4a:I[191445,["/_next/static/immutable/chunks/0v6-3q16afmwn.js","/_next/static/immutable/chunks/2ck6yuttil9xz.js","/_next/static/immutable/chunks/3rsnyfmzi6oij.js","/_next/static/immutable/chunks/0t864vwxbayyr.js","/_next/static/immutable/chunks/3vba9hf7-zrge.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/1w4p5fqe01ywk.js","/_next/static/immutable/chunks/2nv7z5ggl2vxy.js","/_next/static/immutable/chunks/2xoyts_0uq5uj.js","/_next/static/immutable/chunks/2u07s2f6fbivt.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/1eav89ga2ish_.js","/_next/static/immutable/chunks/234j1a4a10s_s.js","/_next/static/immutable/chunks/2gj-fqwe9ql6k.js","/_next/static/immutable/chunks/3b7s527kqlb86.js","/_next/static/immutable/chunks/3qfv9aanu18_5.js","/_next/static/immutable/chunks/1dm59r-7ja6p-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/0q_fqruepll3-.js","/_next/static/immutable/chunks/0-t3xv07arlr0.js","/_next/static/immutable/chunks/2bxez58c_4__7.js","/_next/static/immutable/chunks/3hymg46f2cijm.js","/_next/static/immutable/chunks/3ifxe0l6v3puc.js","/_next/static/immutable/chunks/2ity0kkyaf8hb.js","/_next/static/immutable/chunks/0vrqfaof35wkh.js","/_next/static/immutable/chunks/3_zfwhp9enp9f.js","/_next/static/immutable/chunks/3fohu7x10eje8.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/2azasfm84f4nm.js","/_next/static/immutable/chunks/2wlnf07d6jjbt.js","/_next/static/immutable/chunks/1-au7zm9m8n_w.js","/_next/static/immutable/chunks/0ot5-xo4ttfz-.js","/_next/static/immutable/chunks/1u-ziz4uaewf-.js","/_next/static/immutable/chunks/2h3oa9-6aag1f.js","/_next/static/immutable/chunks/0_l706ym7bmpj.js","/_next/static/immutable/chunks/3vqctoi0mp5pw.js","/_next/static/immutable/chunks/2ddrbcod4yybp.js","/_next/static/immutable/chunks/1q7kned6j2qte.js","/_next/static/immutable/chunks/06f-3e-am-rz0.js","/_next/static/immutable/chunks/04-0klqh106nh.js","/_next/static/immutable/chunks/1ax-wkdfkp0bk.js","/_next/static/immutable/chunks/3_l9h-tkpjgt6.js","/_next/static/immutable/chunks/0yk57mslp5gpo.js","/_next/static/immutable/chunks/2ibdfu2l9iggk.js","/_next/static/immutable/chunks/1vwds7vq-f4ct.js","/_next/static/immutable/chunks/3qeqp7-3q5oxj.js","/_next/static/immutable/chunks/11d1k5hd8gyhp.js","/_next/static/immutable/chunks/1itl0dp78v8hy.js","/_next/static/immutable/chunks/1nv14oq84mc1d.js","/_next/static/immutable/chunks/3k21rgnf_61hs.js"],"WorkTabProvider"] 4b:I[532546,["/_next/static/immutable/chunks/0v6-3q16afmwn.js","/_next/static/immutable/chunks/2ck6yuttil9xz.js","/_next/static/immutable/chunks/3rsnyfmzi6oij.js","/_next/static/immutable/chunks/0t864vwxbayyr.js","/_next/static/immutable/chunks/3vba9hf7-zrge.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/1w4p5fqe01ywk.js","/_next/static/immutable/chunks/2nv7z5ggl2vxy.js","/_next/static/immutable/chunks/2xoyts_0uq5uj.js","/_next/static/immutable/chunks/2u07s2f6fbivt.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/1eav89ga2ish_.js","/_next/static/immutable/chunks/234j1a4a10s_s.js","/_next/static/immutable/chunks/2gj-fqwe9ql6k.js","/_next/static/immutable/chunks/3b7s527kqlb86.js","/_next/static/immutable/chunks/3qfv9aanu18_5.js","/_next/static/immutable/chunks/1dm59r-7ja6p-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/0q_fqruepll3-.js","/_next/static/immutable/chunks/0-t3xv07arlr0.js","/_next/static/immutable/chunks/2bxez58c_4__7.js","/_next/static/immutable/chunks/3hymg46f2cijm.js","/_next/static/immutable/chunks/3ifxe0l6v3puc.js","/_next/static/immutable/chunks/2ity0kkyaf8hb.js","/_next/static/immutable/chunks/0vrqfaof35wkh.js","/_next/static/immutable/chunks/3_zfwhp9enp9f.js","/_next/static/immutable/chunks/3fohu7x10eje8.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/2azasfm84f4nm.js","/_next/static/immutable/chunks/2wlnf07d6jjbt.js","/_next/static/immutable/chunks/1-au7zm9m8n_w.js","/_next/static/immutable/chunks/0ot5-xo4ttfz-.js","/_next/static/immutable/chunks/1u-ziz4uaewf-.js","/_next/static/immutable/chunks/2h3oa9-6aag1f.js","/_next/static/immutable/chunks/0_l706ym7bmpj.js","/_next/static/immutable/chunks/3vqctoi0mp5pw.js","/_next/static/immutable/chunks/2ddrbcod4yybp.js","/_next/static/immutable/chunks/1q7kned6j2qte.js","/_next/static/immutable/chunks/06f-3e-am-rz0.js","/_next/static/immutable/chunks/04-0klqh106nh.js","/_next/static/immutable/chunks/1ax-wkdfkp0bk.js","/_next/static/immutable/chunks/3_l9h-tkpjgt6.js","/_next/static/immutable/chunks/0yk57mslp5gpo.js","/_next/static/immutable/chunks/2ibdfu2l9iggk.js","/_next/static/immutable/chunks/1vwds7vq-f4ct.js","/_next/static/immutable/chunks/3qeqp7-3q5oxj.js","/_next/static/immutable/chunks/11d1k5hd8gyhp.js","/_next/static/immutable/chunks/1itl0dp78v8hy.js","/_next/static/immutable/chunks/1nv14oq84mc1d.js","/_next/static/immutable/chunks/3k21rgnf_61hs.js"],"WorkHeaderProposal"] 57:I[127600,["/_next/static/immutable/chunks/0v6-3q16afmwn.js","/_next/static/immutable/chunks/2ck6yuttil9xz.js","/_next/static/immutable/chunks/3rsnyfmzi6oij.js","/_next/static/immutable/chunks/0t864vwxbayyr.js","/_next/static/immutable/chunks/3vba9hf7-zrge.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/1w4p5fqe01ywk.js","/_next/static/immutable/chunks/2nv7z5ggl2vxy.js","/_next/static/immutable/chunks/2xoyts_0uq5uj.js","/_next/static/immutable/chunks/2u07s2f6fbivt.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/1eav89ga2ish_.js","/_next/static/immutable/chunks/234j1a4a10s_s.js","/_next/static/immutable/chunks/2gj-fqwe9ql6k.js","/_next/static/immutable/chunks/3b7s527kqlb86.js","/_next/static/immutable/chunks/3qfv9aanu18_5.js","/_next/static/immutable/chunks/1dm59r-7ja6p-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/0q_fqruepll3-.js","/_next/static/immutable/chunks/0-t3xv07arlr0.js","/_next/static/immutable/chunks/2bxez58c_4__7.js","/_next/static/immutable/chunks/3hymg46f2cijm.js","/_next/static/immutable/chunks/3ifxe0l6v3puc.js","/_next/static/immutable/chunks/2ity0kkyaf8hb.js","/_next/static/immutable/chunks/0vrqfaof35wkh.js","/_next/static/immutable/chunks/3_zfwhp9enp9f.js","/_next/static/immutable/chunks/3fohu7x10eje8.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/2azasfm84f4nm.js","/_next/static/immutable/chunks/2wlnf07d6jjbt.js","/_next/static/immutable/chunks/1-au7zm9m8n_w.js","/_next/static/immutable/chunks/0ot5-xo4ttfz-.js","/_next/static/immutable/chunks/1u-ziz4uaewf-.js","/_next/static/immutable/chunks/2h3oa9-6aag1f.js","/_next/static/immutable/chunks/0_l706ym7bmpj.js","/_next/static/immutable/chunks/3vqctoi0mp5pw.js","/_next/static/immutable/chunks/2ddrbcod4yybp.js","/_next/static/immutable/chunks/1q7kned6j2qte.js","/_next/static/immutable/chunks/06f-3e-am-rz0.js","/_next/static/immutable/chunks/04-0klqh106nh.js","/_next/static/immutable/chunks/1ax-wkdfkp0bk.js","/_next/static/immutable/chunks/3_l9h-tkpjgt6.js","/_next/static/immutable/chunks/0yk57mslp5gpo.js","/_next/static/immutable/chunks/2ibdfu2l9iggk.js","/_next/static/immutable/chunks/1vwds7vq-f4ct.js","/_next/static/immutable/chunks/3qeqp7-3q5oxj.js","/_next/static/immutable/chunks/11d1k5hd8gyhp.js","/_next/static/immutable/chunks/1itl0dp78v8hy.js","/_next/static/immutable/chunks/1nv14oq84mc1d.js","/_next/static/immutable/chunks/3k21rgnf_61hs.js","/_next/static/immutable/chunks/2h79i-bk3ov-9.js","/_next/static/immutable/chunks/022pez2xgib6m.js","/_next/static/immutable/chunks/0s2fg0442qlk3.js","/_next/static/immutable/chunks/3jyk7ph5wdhcn.js"],"FundDocument"] 4c:T4f82,Learning and Retention from Verified Experience Project host: Dilate Technologies · Programme: Loopseed 1. Abstract A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results. 2. Introduction When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6] Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, Dynamical Synthesis: Learning through Interaction, reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance. We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive. Hypotheses The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types. H1. Improvement beyond memory Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain. H2. Retention of earlier learning After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable. H3. Contribution of stored memory Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples. H4. Contribution of correct training targets The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs. H5. Contribution of replay Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing. H6. Transfer to different rule types The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered. 3. Methods Participants/Sample The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample. Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample. Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection. Materials and Procedures Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions. All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are: A. Frozen reference: no further training or retrieved examples. B. Memory alone: retrieved examples, without further training. C. Training with replay: checked updates and replay, without retrieval. D. Combined procedure: the trained checkpoints from C, evaluated with retrieval. E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval. F. Training without replay: checked new targets and retrieval, without earlier training examples. For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout. The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query. Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule. Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample. After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements. Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed. During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget. Planned Analyses The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses. H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty. These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects. Ethics and Data Management The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores. The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development. 4. Pilot Data The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81. Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate. A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions. Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss. The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning. 5. Budget We propose a six-month budget of USD 25,000, including platform fees. Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure: Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation. Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution. External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements. Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour. Storage and backups: $300: six months at $50 for model files, outputs and backup copies. Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation. Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5] Contingency: $1,650: for additional runtime, price changes and other confirmed project costs. The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity. The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation. After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase. 6. References Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials. Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript. Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article. Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article. ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026). Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper. Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026). Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper. Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book.4d:T500c,Learning and Retention from Verified Experience Project host: Dilate Technologies · Programme: Loopseed 1. Abstract A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results. 2. Introduction When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6] Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, Dynamical Synthesis: Learning through Interaction, reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance. We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive. Hypotheses The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types. H1. Improvement beyond memory Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain. H2. Retention of earlier learning After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable. H3. Contribution of stored memory Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples. H4. Contribution of correct training targets The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs. H5. Contribution of replay Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing. H6. Transfer to different rule types The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered. 3. Methods Participants/Sample The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample. Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample. Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection. Materials and Procedures Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions. All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are: A. Frozen reference: no further training or retrieved examples. B. Memory alone: retrieved examples, without further training. C. Training with replay: checked updates and replay, without retrieval. D. Combined procedure: the trained checkpoints from C, evaluated with retrieval. E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval. F. Training without replay: checked new targets and retrieval, without earlier training examples. For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout. The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query. Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule. Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample. After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements. Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed. During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget. Planned Analyses The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses. H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty. These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects. Ethics and Data Management The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores. The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development. 4. Pilot Data The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81. Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate. A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions. Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss. The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning. 5. Budget We propose a six-month budget of USD 25,000, including platform fees. Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure: Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation. Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution. External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements. Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour. Storage and backups: $300: six months at $50 for model files, outputs and backup copies. Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation. Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5] Contingency: $1,650: for additional runtime, price changes and other confirmed project costs. The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity. The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation. After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase. 6. References Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials. Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript. Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article. Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article. ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026). Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper. Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026). Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper. Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book. 4e:T95a6,{"type":"doc","content":[{"type":"heading","attrs":{"id":"59c2432e-ee81-4730-94dc-99b1ab2aa3d6","data-toc-id":"59c2432e-ee81-4730-94dc-99b1ab2aa3d6","textAlign":null,"level":1},"content":[{"type":"text","text":"Learning and Retention from Verified Experience"}]},{"type":"paragraph","attrs":{"id":"9e864374-dbd3-4016-867e-b31df266ef70","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Project host:"},{"type":"text","text":" Dilate Technologies · "},{"type":"text","marks":[{"type":"bold"}],"text":"Programme:"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://loopseed.io/","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Loopseed"}]},{"type":"heading","attrs":{"id":"e71c737f-7c4f-4039-9c71-ec452d1d50da","data-toc-id":"e71c737f-7c4f-4039-9c71-ec452d1d50da","textAlign":null,"level":2},"content":[{"type":"text","text":"1. Abstract"}]},{"type":"paragraph","attrs":{"id":"f0787fc3-414a-496e-9d55-8cc301aae393","class":null,"textAlign":null},"content":[{"type":"text","text":"A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results."}]},{"type":"heading","attrs":{"id":"d84945ae-5538-4c45-acb1-ee91e8d30164","data-toc-id":"d84945ae-5538-4c45-acb1-ee91e8d30164","textAlign":null,"level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"5d5ab5af-7b55-4d29-9580-ae5bf451779e","class":null,"textAlign":null},"content":[{"type":"text","text":"When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6]"}]},{"type":"paragraph","attrs":{"id":"b4a36e9f-d69f-45b1-9b66-d6191578fadd","class":null,"textAlign":null},"content":[{"type":"text","text":"Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, "},{"type":"text","marks":[{"type":"italic"}],"text":"Dynamical Synthesis: Learning through Interaction"},{"type":"text","text":", reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance."}]},{"type":"paragraph","attrs":{"id":"8ab6bb9f-c94b-4e49-8408-76ce6e52e6b7","class":null,"textAlign":null},"content":[{"type":"text","text":"We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive."}]},{"type":"heading","attrs":{"id":"e1e40413-f49a-4474-8ab1-add1db5fb256","data-toc-id":"e1e40413-f49a-4474-8ab1-add1db5fb256","textAlign":null,"level":3},"content":[{"type":"text","text":"Hypotheses"}]},{"type":"paragraph","attrs":{"id":"4ced654d-61df-4bac-9b64-d7a5fc30bd2b","class":null,"textAlign":null},"content":[{"type":"text","text":"The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types."}]},{"type":"heading","attrs":{"id":"db5e9148-c994-452d-9a15-ae13f1d24066","data-toc-id":"db5e9148-c994-452d-9a15-ae13f1d24066","textAlign":null,"level":4},"content":[{"type":"text","text":"H1. Improvement beyond memory"}]},{"type":"paragraph","attrs":{"id":"12b61285-3068-4636-947d-f69eb1bfd17d","class":null,"textAlign":null},"content":[{"type":"text","text":"Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain."}]},{"type":"heading","attrs":{"id":"a300a23d-1781-4a74-8f66-199ff19245ac","data-toc-id":"a300a23d-1781-4a74-8f66-199ff19245ac","textAlign":null,"level":4},"content":[{"type":"text","text":"H2. Retention of earlier learning"}]},{"type":"paragraph","attrs":{"id":"518805db-fe89-428f-9c75-cd09569a8e7f","class":null,"textAlign":null},"content":[{"type":"text","text":"After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable."}]},{"type":"heading","attrs":{"id":"51e6614c-330c-407d-b870-6ed9eb99f8a0","data-toc-id":"51e6614c-330c-407d-b870-6ed9eb99f8a0","textAlign":null,"level":4},"content":[{"type":"text","text":"H3. Contribution of stored memory"}]},{"type":"paragraph","attrs":{"id":"b98a4928-70fd-4095-b50c-860fa6e0cdf7","class":null,"textAlign":null},"content":[{"type":"text","text":"Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples."}]},{"type":"heading","attrs":{"id":"1efc5946-c9b8-47e6-824c-400a66fc0457","data-toc-id":"1efc5946-c9b8-47e6-824c-400a66fc0457","textAlign":null,"level":4},"content":[{"type":"text","text":"H4. Contribution of correct training targets"}]},{"type":"paragraph","attrs":{"id":"dce59768-157b-4d6e-b9c1-5c11d8763790","class":null,"textAlign":null},"content":[{"type":"text","text":"The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs."}]},{"type":"heading","attrs":{"id":"32ce2688-e7ff-40ce-b4e4-4703569c733a","data-toc-id":"32ce2688-e7ff-40ce-b4e4-4703569c733a","textAlign":null,"level":4},"content":[{"type":"text","text":"H5. Contribution of replay"}]},{"type":"paragraph","attrs":{"id":"da9a2528-82c4-4424-8580-a8d247ab0e0c","class":null,"textAlign":null},"content":[{"type":"text","text":"Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing."}]},{"type":"heading","attrs":{"id":"6873fbef-c5b9-4d8e-b016-aee17479695d","data-toc-id":"6873fbef-c5b9-4d8e-b016-aee17479695d","textAlign":null,"level":4},"content":[{"type":"text","text":"H6. Transfer to different rule types"}]},{"type":"paragraph","attrs":{"id":"02a32b8d-8e9b-4764-82af-85c5d14b0682","class":null,"textAlign":null},"content":[{"type":"text","text":"The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered."}]},{"type":"heading","attrs":{"id":"3face36d-7679-44e1-ac8a-3648bd2f6ae5","data-toc-id":"3face36d-7679-44e1-ac8a-3648bd2f6ae5","textAlign":null,"level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"43cf84eb-d721-4098-be23-10a6b3e3539b","data-toc-id":"43cf84eb-d721-4098-be23-10a6b3e3539b","textAlign":null,"level":3},"content":[{"type":"text","text":"Participants/Sample"}]},{"type":"paragraph","attrs":{"id":"7e38a7b1-42bf-4973-810f-829c565ee20c","class":null,"textAlign":null},"content":[{"type":"text","text":"The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample."}]},{"type":"paragraph","attrs":{"id":"22395820-1656-4f9f-ba4f-30e73cfbbbe4","class":null,"textAlign":null},"content":[{"type":"text","text":"Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample."}]},{"type":"paragraph","attrs":{"id":"f9fbbc2c-3c1e-401e-a762-325b71ac7c47","class":null,"textAlign":null},"content":[{"type":"text","text":"Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection."}]},{"type":"heading","attrs":{"id":"998d7367-d7e0-4d04-9321-91b9fa52c23e","data-toc-id":"998d7367-d7e0-4d04-9321-91b9fa52c23e","textAlign":null,"level":3},"content":[{"type":"text","text":"Materials and Procedures"}]},{"type":"paragraph","attrs":{"id":"21d2f61e-6ff2-408e-bee6-d69918ab9514","class":null,"textAlign":null},"content":[{"type":"text","text":"Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions."}]},{"type":"paragraph","attrs":{"id":"19a83e19-3bf2-4cd5-a7ac-e98eec27a918","class":null,"textAlign":null},"content":[{"type":"text","text":"All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c004852a-3ac2-42c8-9b34-f04c43e950d6","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"A. Frozen reference:"},{"type":"text","text":" no further training or retrieved examples."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"84d8cd76-5066-4e09-b73f-549ae8a0f76b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"B. Memory alone:"},{"type":"text","text":" retrieved examples, without further training."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"00e54fb8-0faa-4c74-ae35-96206b6af0a5","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"C. Training with replay:"},{"type":"text","text":" checked updates and replay, without retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"193f23c6-93a0-4f07-9f6a-f616e9b0f3df","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"D. Combined procedure:"},{"type":"text","text":" the trained checkpoints from C, evaluated with retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f5d59d32-036f-4479-b40c-c52de11c7d0d","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"E. Shuffled-target control:"},{"type":"text","text":" shuffled new training targets, with correct replay and retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b1a8f30c-a796-48ba-baf0-ec2c70dcccc1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"F. Training without replay:"},{"type":"text","text":" checked new targets and retrieval, without earlier training examples."}]}]}]},{"type":"paragraph","attrs":{"id":"615de6d9-c43a-4f1b-94f2-9c324e745777","class":null,"textAlign":null},"content":[{"type":"text","text":"For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout."}]},{"type":"paragraph","attrs":{"id":"c63ba6e8-0de8-498a-9f85-ff0ae290cf84","class":null,"textAlign":null},"content":[{"type":"text","text":"The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query."}]},{"type":"paragraph","attrs":{"id":"296181c4-5418-4919-a58c-07ff13e23c53","class":null,"textAlign":null},"content":[{"type":"text","text":"Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/c77722fc-10f2-4830-bf9f-097c408babf7/study-design.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"269bea55-1e80-4150-b5cd-90a9521d439d","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample."}]},{"type":"paragraph","attrs":{"id":"9c73469f-0395-4fac-9961-a17afc8525e9","class":null,"textAlign":null},"content":[{"type":"text","text":"After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements."}]},{"type":"paragraph","attrs":{"id":"8f84492e-d615-4c0e-88e2-e0b93dc6764e","class":null,"textAlign":null},"content":[{"type":"text","text":"Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed."}]},{"type":"paragraph","attrs":{"id":"ccecaa06-7da4-4f21-a89b-ef6bf231932b","class":null,"textAlign":null},"content":[{"type":"text","text":"During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget."}]},{"type":"heading","attrs":{"id":"564acebb-57da-433a-8ecc-6110db5745d0","data-toc-id":"564acebb-57da-433a-8ecc-6110db5745d0","textAlign":null,"level":3},"content":[{"type":"text","text":"Planned Analyses"}]},{"type":"paragraph","attrs":{"id":"a72835aa-62e4-466c-b064-e742c3574f33","class":null,"textAlign":null},"content":[{"type":"text","text":"The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses."}]},{"type":"paragraph","attrs":{"id":"563cdf66-43d7-4d83-aadb-a6bfbcbc5d56","class":null,"textAlign":null},"content":[{"type":"text","text":"H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty."}]},{"type":"paragraph","attrs":{"id":"20475d66-e4b7-4d9d-a420-68bc7b5e17b6","class":null,"textAlign":null},"content":[{"type":"text","text":"These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects."}]},{"type":"heading","attrs":{"id":"a6c60e17-1db5-43bd-aec0-f894376cc082","data-toc-id":"a6c60e17-1db5-43bd-aec0-f894376cc082","textAlign":null,"level":3},"content":[{"type":"text","text":"Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"261c63d5-9f83-420f-83e0-50a6062291ad","class":null,"textAlign":null},"content":[{"type":"text","text":"The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores."}]},{"type":"paragraph","attrs":{"id":"8b1dea88-e09d-49e7-84f8-e9afd5b5d32e","class":null,"textAlign":null},"content":[{"type":"text","text":"The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development."}]},{"type":"heading","attrs":{"id":"8e9de865-a5c5-4aa9-b250-7b9c0ff1e3e4","data-toc-id":"8e9de865-a5c5-4aa9-b250-7b9c0ff1e3e4","textAlign":null,"level":2},"content":[{"type":"text","text":"4. Pilot Data"}]},{"type":"paragraph","attrs":{"id":"7b482172-b13c-4d69-8f12-52bbe33bc6fc","class":null,"textAlign":null},"content":[{"type":"text","text":"The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/a97eddd2-8196-41ae-96f5-019091b8969a/adapter-transfer.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"359d9736-e38b-4714-80e4-38b7ebbf6f55","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate."}]},{"type":"paragraph","attrs":{"id":"eb5e4eee-0b83-4c12-a68a-2bbd4ef7cf4e","class":null,"textAlign":null},"content":[{"type":"text","text":"A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/bdaabc9b-4ee0-4e43-84a4-72616c168769/continuation.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"5a312d7f-3d8a-4173-9c5d-d7aeed7b6e7e","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss."}]},{"type":"paragraph","attrs":{"id":"7c2444ad-2d85-435e-a140-438d2b1412f2","class":null,"textAlign":null},"content":[{"type":"text","text":"The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning."}]},{"type":"heading","attrs":{"id":"3d9246ca-75e9-448d-8fae-22dcc8517dcf","data-toc-id":"3d9246ca-75e9-448d-8fae-22dcc8517dcf","textAlign":null,"level":2},"content":[{"type":"text","text":"5. Budget"}]},{"type":"paragraph","attrs":{"id":"c38ef251-75cf-4c4f-a74c-e993cbc24bd1","class":null,"textAlign":null},"content":[{"type":"text","text":"We propose a six-month budget of "},{"type":"text","marks":[{"type":"bold"}],"text":"USD 25,000, including platform fees"},{"type":"text","text":". Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f2ba64df-0487-4890-84a0-b358380ce949","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research lead: $7,200:"},{"type":"text","text":" 240 hours at $30 for design, supervision, analysis and paper preparation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"94c62709-b905-4fa9-8114-d966640be559","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research engineering: $8,000:"},{"type":"text","text":" 320 hours at $25 for generators, checkers, the training workflow and reproducible execution."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0eab8143-ae90-4112-9ec2-774da98ca18f","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"External review: $2,400:"},{"type":"text","text":" 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"434c4094-9d57-43fb-a085-c631668e5b87","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Cloud compute: $2,400:"},{"type":"text","text":" 1,200 GPU-hours at a planning rate of $2 per hour."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fdb9481f-937a-4929-af29-2ae2816f7be0","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Storage and backups: $300:"},{"type":"text","text":" six months at $50 for model files, outputs and backup copies."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ee1036e8-88f7-4bc1-afcc-ad1cf509017a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Release documentation: $800:"},{"type":"text","text":" 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7ca6c020-eb84-449a-86e1-29230d998c30","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Platform fee allowance: $2,250:"},{"type":"text","text":" 9% of the gross request, using ResearchHub's published rate. [5]"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"d2831113-e71e-4d22-97dc-eb3157af7a26","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Contingency: $1,650:"},{"type":"text","text":" for additional runtime, price changes and other confirmed project costs."}]}]}]},{"type":"paragraph","attrs":{"id":"86d52b9d-1155-4e9e-ba77-19d7554d1fa6","class":null,"textAlign":null},"content":[{"type":"text","text":"The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity."}]},{"type":"paragraph","attrs":{"id":"ba832805-1168-4a0b-90bc-601cc2b82d59","class":null,"textAlign":null},"content":[{"type":"text","text":"The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation."}]},{"type":"paragraph","attrs":{"id":"320e3f4e-4e35-4cb0-ad06-fe04088b9626","class":null,"textAlign":null},"content":[{"type":"text","text":"After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase."}]},{"type":"heading","attrs":{"id":"9502b683-f126-450c-8295-b46906854ed0","data-toc-id":"9502b683-f126-450c-8295-b46906854ed0","textAlign":null,"level":2},"content":[{"type":"text","text":"6. References"}]},{"type":"orderedList","attrs":{"start":1,"type":"1"},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"87e6b8b9-7a95-41d6-b83b-9d04d245d917","class":null,"textAlign":null},"content":[{"type":"text","text":"Ashiundu, B. "},{"type":"text","marks":[{"type":"italic"}],"text":"Dynamical Synthesis: Learning through Interaction"},{"type":"text","text":". Dilate Technologies (2026). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://loopseed.io/papers/dynamical-synthesis.html","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Paper and supporting materials"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"bfd2308a-0b8b-44c4-aa28-582b455e9b1a","class":null,"textAlign":null},"content":[{"type":"text","text":"Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In "},{"type":"text","marks":[{"type":"italic"}],"text":"International Conference on Learning Representations"},{"type":"text","text":" (2022). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://arxiv.org/abs/2106.09685","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"LoRA manuscript"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2dbd2687-9844-45f3-bb9f-10446f86c15c","class":null,"textAlign":null},"content":[{"type":"text","text":"Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. "},{"type":"text","marks":[{"type":"italic"}],"text":"Proc. Natl Acad. Sci. USA"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"114"},{"type":"text","text":", 3521–3526 (2017). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1073/pnas.1611835114","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Article"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"eecc9056-45ab-4c02-a4f9-e95e44a66d06","class":null,"textAlign":null},"content":[{"type":"text","text":"Lakens, D. Sample size justification. "},{"type":"text","marks":[{"type":"italic"}],"text":"Collabra: Psychology"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"8"},{"type":"text","text":", 33267 (2022). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1525/collabra.33267","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Article"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"66933723-8a9b-40ca-956b-2613a675cf5e","class":null,"textAlign":null},"content":[{"type":"text","text":"ResearchHub. "},{"type":"text","marks":[{"type":"italic"}],"text":"Funders & Donors"},{"type":"text","text":". ResearchHub documentation. "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://docs.researchhub.com/researchhub/product-features/fund/funders-and-donors","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Platform funding fees"},{"type":"text","text":" (accessed 13 September 2026)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"faf1ccda-cb19-4652-9d2b-91d6723d2180","class":null,"textAlign":null},"content":[{"type":"text","text":"Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In "},{"type":"text","marks":[{"type":"italic"}],"text":"Advances in Neural Information Processing Systems"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"32"},{"type":"text","text":" (2019). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://proceedings.neurips.cc/paper_files/paper/2019/hash/fa7cdfad1a5aaf8370ebeda47a1ff1c3-Abstract.html","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Conference paper"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"e4697902-4430-4bad-8f5b-57dc8ee7626f","class":null,"textAlign":null},"content":[{"type":"text","text":"Runpod. "},{"type":"text","marks":[{"type":"italic"}],"text":"GPU Cloud Pricing"},{"type":"text","text":". "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://www.runpod.io/pricing","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Published GPU and storage rates"},{"type":"text","text":" (accessed 13 September 2026)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"d1a2d68c-8535-43b6-87a6-a990b91b6e12","class":null,"textAlign":null},"content":[{"type":"text","text":"Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In "},{"type":"text","marks":[{"type":"italic"}],"text":"Proceedings of the Seventh International Conference on Machine Learning"},{"type":"text","text":", 216–224 (1990). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1016/B978-1-55860-141-3.50030-4","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Conference paper"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9f5b49c2-81eb-4ac1-be0b-216ef0e5a340","class":null,"textAlign":null},"content":[{"type":"text","text":"Wolfram, S. "},{"type":"text","marks":[{"type":"italic"}],"text":"A New Kind of Science"},{"type":"text","text":". Wolfram Media (2002). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://www.wolframscience.com/nks/","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Online book"},{"type":"text","text":"."}]}]}]},{"type":"paragraph","attrs":{"id":"5d43023a-db0e-4af0-9440-3fb6adf77600","class":null,"textAlign":null}}]}4f:T500c,Learning and Retention from Verified Experience Project host: Dilate Technologies · Programme: Loopseed 1. Abstract A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results. 2. Introduction When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6] Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, Dynamical Synthesis: Learning through Interaction, reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance. We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive. Hypotheses The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types. H1. Improvement beyond memory Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain. H2. Retention of earlier learning After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable. H3. Contribution of stored memory Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples. H4. Contribution of correct training targets The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs. H5. Contribution of replay Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing. H6. Transfer to different rule types The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered. 3. Methods Participants/Sample The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample. Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample. Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection. Materials and Procedures Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions. All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are: A. Frozen reference: no further training or retrieved examples. B. Memory alone: retrieved examples, without further training. C. Training with replay: checked updates and replay, without retrieval. D. Combined procedure: the trained checkpoints from C, evaluated with retrieval. E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval. F. Training without replay: checked new targets and retrieval, without earlier training examples. For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout. The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query. Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule. Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample. After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements. Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed. During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget. Planned Analyses The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses. H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty. These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects. Ethics and Data Management The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores. The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development. 4. Pilot Data The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81. Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate. A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions. Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss. The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning. 5. Budget We propose a six-month budget of USD 25,000, including platform fees. Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure: Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation. Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution. External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements. Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour. Storage and backups: $300: six months at $50 for model files, outputs and backup copies. Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation. Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5] Contingency: $1,650: for additional runtime, price changes and other confirmed project costs. The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity. The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation. After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase. 6. References Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials. Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript. Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article. Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article. ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026). Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper. Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026). Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper. Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book. 50:T95a6,{"type":"doc","content":[{"type":"heading","attrs":{"id":"59c2432e-ee81-4730-94dc-99b1ab2aa3d6","data-toc-id":"59c2432e-ee81-4730-94dc-99b1ab2aa3d6","textAlign":null,"level":1},"content":[{"type":"text","text":"Learning and Retention from Verified Experience"}]},{"type":"paragraph","attrs":{"id":"9e864374-dbd3-4016-867e-b31df266ef70","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Project host:"},{"type":"text","text":" Dilate Technologies · "},{"type":"text","marks":[{"type":"bold"}],"text":"Programme:"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://loopseed.io/","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Loopseed"}]},{"type":"heading","attrs":{"id":"e71c737f-7c4f-4039-9c71-ec452d1d50da","data-toc-id":"e71c737f-7c4f-4039-9c71-ec452d1d50da","textAlign":null,"level":2},"content":[{"type":"text","text":"1. Abstract"}]},{"type":"paragraph","attrs":{"id":"f0787fc3-414a-496e-9d55-8cc301aae393","class":null,"textAlign":null},"content":[{"type":"text","text":"A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results."}]},{"type":"heading","attrs":{"id":"d84945ae-5538-4c45-acb1-ee91e8d30164","data-toc-id":"d84945ae-5538-4c45-acb1-ee91e8d30164","textAlign":null,"level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"5d5ab5af-7b55-4d29-9580-ae5bf451779e","class":null,"textAlign":null},"content":[{"type":"text","text":"When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6]"}]},{"type":"paragraph","attrs":{"id":"b4a36e9f-d69f-45b1-9b66-d6191578fadd","class":null,"textAlign":null},"content":[{"type":"text","text":"Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, "},{"type":"text","marks":[{"type":"italic"}],"text":"Dynamical Synthesis: Learning through Interaction"},{"type":"text","text":", reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance."}]},{"type":"paragraph","attrs":{"id":"8ab6bb9f-c94b-4e49-8408-76ce6e52e6b7","class":null,"textAlign":null},"content":[{"type":"text","text":"We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive."}]},{"type":"heading","attrs":{"id":"e1e40413-f49a-4474-8ab1-add1db5fb256","data-toc-id":"e1e40413-f49a-4474-8ab1-add1db5fb256","textAlign":null,"level":3},"content":[{"type":"text","text":"Hypotheses"}]},{"type":"paragraph","attrs":{"id":"4ced654d-61df-4bac-9b64-d7a5fc30bd2b","class":null,"textAlign":null},"content":[{"type":"text","text":"The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types."}]},{"type":"heading","attrs":{"id":"db5e9148-c994-452d-9a15-ae13f1d24066","data-toc-id":"db5e9148-c994-452d-9a15-ae13f1d24066","textAlign":null,"level":4},"content":[{"type":"text","text":"H1. Improvement beyond memory"}]},{"type":"paragraph","attrs":{"id":"12b61285-3068-4636-947d-f69eb1bfd17d","class":null,"textAlign":null},"content":[{"type":"text","text":"Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain."}]},{"type":"heading","attrs":{"id":"a300a23d-1781-4a74-8f66-199ff19245ac","data-toc-id":"a300a23d-1781-4a74-8f66-199ff19245ac","textAlign":null,"level":4},"content":[{"type":"text","text":"H2. Retention of earlier learning"}]},{"type":"paragraph","attrs":{"id":"518805db-fe89-428f-9c75-cd09569a8e7f","class":null,"textAlign":null},"content":[{"type":"text","text":"After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable."}]},{"type":"heading","attrs":{"id":"51e6614c-330c-407d-b870-6ed9eb99f8a0","data-toc-id":"51e6614c-330c-407d-b870-6ed9eb99f8a0","textAlign":null,"level":4},"content":[{"type":"text","text":"H3. Contribution of stored memory"}]},{"type":"paragraph","attrs":{"id":"b98a4928-70fd-4095-b50c-860fa6e0cdf7","class":null,"textAlign":null},"content":[{"type":"text","text":"Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples."}]},{"type":"heading","attrs":{"id":"1efc5946-c9b8-47e6-824c-400a66fc0457","data-toc-id":"1efc5946-c9b8-47e6-824c-400a66fc0457","textAlign":null,"level":4},"content":[{"type":"text","text":"H4. Contribution of correct training targets"}]},{"type":"paragraph","attrs":{"id":"dce59768-157b-4d6e-b9c1-5c11d8763790","class":null,"textAlign":null},"content":[{"type":"text","text":"The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs."}]},{"type":"heading","attrs":{"id":"32ce2688-e7ff-40ce-b4e4-4703569c733a","data-toc-id":"32ce2688-e7ff-40ce-b4e4-4703569c733a","textAlign":null,"level":4},"content":[{"type":"text","text":"H5. Contribution of replay"}]},{"type":"paragraph","attrs":{"id":"da9a2528-82c4-4424-8580-a8d247ab0e0c","class":null,"textAlign":null},"content":[{"type":"text","text":"Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing."}]},{"type":"heading","attrs":{"id":"6873fbef-c5b9-4d8e-b016-aee17479695d","data-toc-id":"6873fbef-c5b9-4d8e-b016-aee17479695d","textAlign":null,"level":4},"content":[{"type":"text","text":"H6. Transfer to different rule types"}]},{"type":"paragraph","attrs":{"id":"02a32b8d-8e9b-4764-82af-85c5d14b0682","class":null,"textAlign":null},"content":[{"type":"text","text":"The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered."}]},{"type":"heading","attrs":{"id":"3face36d-7679-44e1-ac8a-3648bd2f6ae5","data-toc-id":"3face36d-7679-44e1-ac8a-3648bd2f6ae5","textAlign":null,"level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"43cf84eb-d721-4098-be23-10a6b3e3539b","data-toc-id":"43cf84eb-d721-4098-be23-10a6b3e3539b","textAlign":null,"level":3},"content":[{"type":"text","text":"Participants/Sample"}]},{"type":"paragraph","attrs":{"id":"7e38a7b1-42bf-4973-810f-829c565ee20c","class":null,"textAlign":null},"content":[{"type":"text","text":"The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample."}]},{"type":"paragraph","attrs":{"id":"22395820-1656-4f9f-ba4f-30e73cfbbbe4","class":null,"textAlign":null},"content":[{"type":"text","text":"Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample."}]},{"type":"paragraph","attrs":{"id":"f9fbbc2c-3c1e-401e-a762-325b71ac7c47","class":null,"textAlign":null},"content":[{"type":"text","text":"Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection."}]},{"type":"heading","attrs":{"id":"998d7367-d7e0-4d04-9321-91b9fa52c23e","data-toc-id":"998d7367-d7e0-4d04-9321-91b9fa52c23e","textAlign":null,"level":3},"content":[{"type":"text","text":"Materials and Procedures"}]},{"type":"paragraph","attrs":{"id":"21d2f61e-6ff2-408e-bee6-d69918ab9514","class":null,"textAlign":null},"content":[{"type":"text","text":"Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions."}]},{"type":"paragraph","attrs":{"id":"19a83e19-3bf2-4cd5-a7ac-e98eec27a918","class":null,"textAlign":null},"content":[{"type":"text","text":"All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c004852a-3ac2-42c8-9b34-f04c43e950d6","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"A. Frozen reference:"},{"type":"text","text":" no further training or retrieved examples."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"84d8cd76-5066-4e09-b73f-549ae8a0f76b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"B. Memory alone:"},{"type":"text","text":" retrieved examples, without further training."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"00e54fb8-0faa-4c74-ae35-96206b6af0a5","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"C. Training with replay:"},{"type":"text","text":" checked updates and replay, without retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"193f23c6-93a0-4f07-9f6a-f616e9b0f3df","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"D. Combined procedure:"},{"type":"text","text":" the trained checkpoints from C, evaluated with retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f5d59d32-036f-4479-b40c-c52de11c7d0d","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"E. Shuffled-target control:"},{"type":"text","text":" shuffled new training targets, with correct replay and retrieval."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b1a8f30c-a796-48ba-baf0-ec2c70dcccc1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"F. Training without replay:"},{"type":"text","text":" checked new targets and retrieval, without earlier training examples."}]}]}]},{"type":"paragraph","attrs":{"id":"615de6d9-c43a-4f1b-94f2-9c324e745777","class":null,"textAlign":null},"content":[{"type":"text","text":"For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout."}]},{"type":"paragraph","attrs":{"id":"c63ba6e8-0de8-498a-9f85-ff0ae290cf84","class":null,"textAlign":null},"content":[{"type":"text","text":"The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query."}]},{"type":"paragraph","attrs":{"id":"296181c4-5418-4919-a58c-07ff13e23c53","class":null,"textAlign":null},"content":[{"type":"text","text":"Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/c77722fc-10f2-4830-bf9f-097c408babf7/study-design.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"269bea55-1e80-4150-b5cd-90a9521d439d","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample."}]},{"type":"paragraph","attrs":{"id":"9c73469f-0395-4fac-9961-a17afc8525e9","class":null,"textAlign":null},"content":[{"type":"text","text":"After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements."}]},{"type":"paragraph","attrs":{"id":"8f84492e-d615-4c0e-88e2-e0b93dc6764e","class":null,"textAlign":null},"content":[{"type":"text","text":"Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed."}]},{"type":"paragraph","attrs":{"id":"ccecaa06-7da4-4f21-a89b-ef6bf231932b","class":null,"textAlign":null},"content":[{"type":"text","text":"During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget."}]},{"type":"heading","attrs":{"id":"564acebb-57da-433a-8ecc-6110db5745d0","data-toc-id":"564acebb-57da-433a-8ecc-6110db5745d0","textAlign":null,"level":3},"content":[{"type":"text","text":"Planned Analyses"}]},{"type":"paragraph","attrs":{"id":"a72835aa-62e4-466c-b064-e742c3574f33","class":null,"textAlign":null},"content":[{"type":"text","text":"The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses."}]},{"type":"paragraph","attrs":{"id":"563cdf66-43d7-4d83-aadb-a6bfbcbc5d56","class":null,"textAlign":null},"content":[{"type":"text","text":"H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty."}]},{"type":"paragraph","attrs":{"id":"20475d66-e4b7-4d9d-a420-68bc7b5e17b6","class":null,"textAlign":null},"content":[{"type":"text","text":"These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects."}]},{"type":"heading","attrs":{"id":"a6c60e17-1db5-43bd-aec0-f894376cc082","data-toc-id":"a6c60e17-1db5-43bd-aec0-f894376cc082","textAlign":null,"level":3},"content":[{"type":"text","text":"Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"261c63d5-9f83-420f-83e0-50a6062291ad","class":null,"textAlign":null},"content":[{"type":"text","text":"The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores."}]},{"type":"paragraph","attrs":{"id":"8b1dea88-e09d-49e7-84f8-e9afd5b5d32e","class":null,"textAlign":null},"content":[{"type":"text","text":"The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development."}]},{"type":"heading","attrs":{"id":"8e9de865-a5c5-4aa9-b250-7b9c0ff1e3e4","data-toc-id":"8e9de865-a5c5-4aa9-b250-7b9c0ff1e3e4","textAlign":null,"level":2},"content":[{"type":"text","text":"4. Pilot Data"}]},{"type":"paragraph","attrs":{"id":"7b482172-b13c-4d69-8f12-52bbe33bc6fc","class":null,"textAlign":null},"content":[{"type":"text","text":"The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/a97eddd2-8196-41ae-96f5-019091b8969a/adapter-transfer.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"359d9736-e38b-4714-80e4-38b7ebbf6f55","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate."}]},{"type":"paragraph","attrs":{"id":"eb5e4eee-0b83-4c12-a68a-2bbd4ef7cf4e","class":null,"textAlign":null},"content":[{"type":"text","text":"A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions."}]},{"type":"imageBlock","attrs":{"src":"https://storage.prod.researchhub.com/uploads/notes/users/201779/bdaabc9b-4ee0-4e43-84a4-72616c168769/continuation.png","width":"100%","align":"center"}},{"type":"paragraph","attrs":{"id":"5a312d7f-3d8a-4173-9c5d-d7aeed7b6e7e","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss."}]},{"type":"paragraph","attrs":{"id":"7c2444ad-2d85-435e-a140-438d2b1412f2","class":null,"textAlign":null},"content":[{"type":"text","text":"The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning."}]},{"type":"heading","attrs":{"id":"3d9246ca-75e9-448d-8fae-22dcc8517dcf","data-toc-id":"3d9246ca-75e9-448d-8fae-22dcc8517dcf","textAlign":null,"level":2},"content":[{"type":"text","text":"5. Budget"}]},{"type":"paragraph","attrs":{"id":"c38ef251-75cf-4c4f-a74c-e993cbc24bd1","class":null,"textAlign":null},"content":[{"type":"text","text":"We propose a six-month budget of "},{"type":"text","marks":[{"type":"bold"}],"text":"USD 25,000, including platform fees"},{"type":"text","text":". Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f2ba64df-0487-4890-84a0-b358380ce949","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research lead: $7,200:"},{"type":"text","text":" 240 hours at $30 for design, supervision, analysis and paper preparation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"94c62709-b905-4fa9-8114-d966640be559","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research engineering: $8,000:"},{"type":"text","text":" 320 hours at $25 for generators, checkers, the training workflow and reproducible execution."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0eab8143-ae90-4112-9ec2-774da98ca18f","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"External review: $2,400:"},{"type":"text","text":" 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"434c4094-9d57-43fb-a085-c631668e5b87","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Cloud compute: $2,400:"},{"type":"text","text":" 1,200 GPU-hours at a planning rate of $2 per hour."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fdb9481f-937a-4929-af29-2ae2816f7be0","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Storage and backups: $300:"},{"type":"text","text":" six months at $50 for model files, outputs and backup copies."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ee1036e8-88f7-4bc1-afcc-ad1cf509017a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Release documentation: $800:"},{"type":"text","text":" 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7ca6c020-eb84-449a-86e1-29230d998c30","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Platform fee allowance: $2,250:"},{"type":"text","text":" 9% of the gross request, using ResearchHub's published rate. [5]"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"d2831113-e71e-4d22-97dc-eb3157af7a26","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Contingency: $1,650:"},{"type":"text","text":" for additional runtime, price changes and other confirmed project costs."}]}]}]},{"type":"paragraph","attrs":{"id":"86d52b9d-1155-4e9e-ba77-19d7554d1fa6","class":null,"textAlign":null},"content":[{"type":"text","text":"The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity."}]},{"type":"paragraph","attrs":{"id":"ba832805-1168-4a0b-90bc-601cc2b82d59","class":null,"textAlign":null},"content":[{"type":"text","text":"The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation."}]},{"type":"paragraph","attrs":{"id":"320e3f4e-4e35-4cb0-ad06-fe04088b9626","class":null,"textAlign":null},"content":[{"type":"text","text":"After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase."}]},{"type":"heading","attrs":{"id":"9502b683-f126-450c-8295-b46906854ed0","data-toc-id":"9502b683-f126-450c-8295-b46906854ed0","textAlign":null,"level":2},"content":[{"type":"text","text":"6. References"}]},{"type":"orderedList","attrs":{"start":1,"type":"1"},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"87e6b8b9-7a95-41d6-b83b-9d04d245d917","class":null,"textAlign":null},"content":[{"type":"text","text":"Ashiundu, B. "},{"type":"text","marks":[{"type":"italic"}],"text":"Dynamical Synthesis: Learning through Interaction"},{"type":"text","text":". Dilate Technologies (2026). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://loopseed.io/papers/dynamical-synthesis.html","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Paper and supporting materials"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"bfd2308a-0b8b-44c4-aa28-582b455e9b1a","class":null,"textAlign":null},"content":[{"type":"text","text":"Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In "},{"type":"text","marks":[{"type":"italic"}],"text":"International Conference on Learning Representations"},{"type":"text","text":" (2022). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://arxiv.org/abs/2106.09685","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"LoRA manuscript"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2dbd2687-9844-45f3-bb9f-10446f86c15c","class":null,"textAlign":null},"content":[{"type":"text","text":"Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. "},{"type":"text","marks":[{"type":"italic"}],"text":"Proc. Natl Acad. Sci. USA"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"114"},{"type":"text","text":", 3521–3526 (2017). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1073/pnas.1611835114","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Article"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"eecc9056-45ab-4c02-a4f9-e95e44a66d06","class":null,"textAlign":null},"content":[{"type":"text","text":"Lakens, D. Sample size justification. "},{"type":"text","marks":[{"type":"italic"}],"text":"Collabra: Psychology"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"8"},{"type":"text","text":", 33267 (2022). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1525/collabra.33267","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Article"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"66933723-8a9b-40ca-956b-2613a675cf5e","class":null,"textAlign":null},"content":[{"type":"text","text":"ResearchHub. "},{"type":"text","marks":[{"type":"italic"}],"text":"Funders & Donors"},{"type":"text","text":". ResearchHub documentation. "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://docs.researchhub.com/researchhub/product-features/fund/funders-and-donors","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Platform funding fees"},{"type":"text","text":" (accessed 13 September 2026)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"faf1ccda-cb19-4652-9d2b-91d6723d2180","class":null,"textAlign":null},"content":[{"type":"text","text":"Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In "},{"type":"text","marks":[{"type":"italic"}],"text":"Advances in Neural Information Processing Systems"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"32"},{"type":"text","text":" (2019). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://proceedings.neurips.cc/paper_files/paper/2019/hash/fa7cdfad1a5aaf8370ebeda47a1ff1c3-Abstract.html","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Conference paper"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"e4697902-4430-4bad-8f5b-57dc8ee7626f","class":null,"textAlign":null},"content":[{"type":"text","text":"Runpod. "},{"type":"text","marks":[{"type":"italic"}],"text":"GPU Cloud Pricing"},{"type":"text","text":". "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://www.runpod.io/pricing","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Published GPU and storage rates"},{"type":"text","text":" (accessed 13 September 2026)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"d1a2d68c-8535-43b6-87a6-a990b91b6e12","class":null,"textAlign":null},"content":[{"type":"text","text":"Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In "},{"type":"text","marks":[{"type":"italic"}],"text":"Proceedings of the Seventh International Conference on Machine Learning"},{"type":"text","text":", 216–224 (1990). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://doi.org/10.1016/B978-1-55860-141-3.50030-4","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Conference paper"},{"type":"text","text":"."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9f5b49c2-81eb-4ac1-be0b-216ef0e5a340","class":null,"textAlign":null},"content":[{"type":"text","text":"Wolfram, S. "},{"type":"text","marks":[{"type":"italic"}],"text":"A New Kind of Science"},{"type":"text","text":". Wolfram Media (2002). "},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://www.wolframscience.com/nks/","target":"_blank","rel":"noopener noreferrer nofollow","class":null,"title":null}}],"text":"Online book"},{"type":"text","text":"."}]}]}]},{"type":"paragraph","attrs":{"id":"5d43023a-db0e-4af0-9440-3fb6adf77600","class":null,"textAlign":null}}]}51:T733c,

Learning and Retention from Verified Experience

Project host: Dilate Technologies · Programme: Loopseed

1. Abstract

A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results.

2. Introduction

When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6]

Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, Dynamical Synthesis: Learning through Interaction, reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance.

We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive.

Hypotheses

The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types.

H1. Improvement beyond memory

Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain.

H2. Retention of earlier learning

After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable.

H3. Contribution of stored memory

Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples.

H4. Contribution of correct training targets

The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs.

H5. Contribution of replay

Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing.

H6. Transfer to different rule types

The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered.

3. Methods

Participants/Sample

The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample.

Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample.

Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection.

Materials and Procedures

Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions.

All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are:

A. Frozen reference: no further training or retrieved examples.
B. Memory alone: retrieved examples, without further training.
C. Training with replay: checked updates and replay, without retrieval.
D. Combined procedure: the trained checkpoints from C, evaluated with retrieval.
E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval.
F. Training without replay: checked new targets and retrieval, without earlier training examples.

For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout.

The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query.

Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule.

Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample.

After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements.

Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed.

During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget.

Planned Analyses

The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses.

H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty.

These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects.

Ethics and Data Management

The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores.

The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development.

4. Pilot Data

The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81.

Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate.

A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions.

Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss.

The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning.

5. Budget

We propose a six-month budget of USD 25,000, including platform fees. Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure:

Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation.
Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution.
External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements.
Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour.
Storage and backups: $300: six months at $50 for model files, outputs and backup copies.
Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation.
Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5]
Contingency: $1,650: for additional runtime, price changes and other confirmed project costs.

The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity.

The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation.

After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase.

6. References

Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials.
Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript.
Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article.
Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article.
ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026).
Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper.
Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026).
Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper.
Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book.

52:T733c,

Learning and Retention from Verified Experience

Project host: Dilate Technologies · Programme: Loopseed

1. Abstract

2. Introduction

Hypotheses

H1. Improvement beyond memory

H2. Retention of earlier learning

H3. Contribution of stored memory

Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples.

H4. Contribution of correct training targets

The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs.

H5. Contribution of replay

Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing.

H6. Transfer to different rule types

The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered.

3. Methods

Participants/Sample

Materials and Procedures

A. Frozen reference: no further training or retrieved examples.
B. Memory alone: retrieved examples, without further training.
C. Training with replay: checked updates and replay, without retrieval.
D. Combined procedure: the trained checkpoints from C, evaluated with retrieval.
E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval.
F. Training without replay: checked new targets and retrieval, without earlier training examples.

Planned Analyses

Ethics and Data Management

4. Pilot Data

5. Budget

Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation.
Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution.
External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements.
Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour.
Storage and backups: $300: six months at $50 for model files, outputs and backup copies.
Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation.
Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5]
Contingency: $1,650: for additional runtime, price changes and other confirmed project costs.

6. References

Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials.
Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript.
Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article.
Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article.
ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026).
Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper.
Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026).
Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper.
Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book.

53:T500a,Learning and Retention from Verified Experience Project host: Dilate Technologies · Programme: Loopseed 1. Abstract A central research challenge is developing computational systems that reliably learn from new experience without losing earlier abilities. Loopseed's first paper reported improved answers on selected mathematics problems, followed by limited transfer and losses after further training. We will investigate these limitations by comparing access to stored examples, further model training and reuse of earlier training examples. For comparison, we will train another model on the same new questions paired with answers from other questions. Tasks will follow fixed computational rules so that correct answers can be calculated exactly and the tests repeated. A separate pilot of 12 task sequences will prepare a main study of 100 generated sequences, each tested under six conditions over three learning rounds. The primary outcomes are improvement on new inputs and retention of initial-task performance after restarting the system. Brian Ashiundu will lead the six-month project at Dilate Technologies, using existing research software, local development facilities and budgeted cloud compute, engineering and external review. The study uses synthetic data; applicable ethics review requirements will be confirmed before collection. We will publish the protocol, task generators, model outputs and analysis code, together with a follow-up paper reporting positive, negative and uncertain results. 2. Introduction When an adaptive system answers a problem correctly after seeing new examples, the improvement may depend on information still available in its context or memory, on changes to its parameters, or on both. Distinguishing these effects is necessary to establish what the system has learned and whether that learning persists. Further training introduces another difficulty: improvements on recent examples can accompany losses on earlier tasks. Research on continual learning has examined ways to limit these losses, including constraints on parameter changes and replay of earlier examples. [3,6] Loopseed is a research programme investigating learning through interaction in computational systems. Its first paper, Dynamical Synthesis: Learning through Interaction, reported improved answers in a controlled mathematics comparison, but subsequent studies found limited transfer and losses after further training. [1] These results motivate a study of repeated updates in which memory, training and replay can be compared on the same tasks. The question is whether training adds a measurable benefit beyond access to stored examples, and whether that benefit is compatible with retaining earlier performance. We will address this question using low-rank parameter updates [2] and tasks generated by finite computational rules. This design draws on work integrating experience and learning [8] and on the use of simple programs to study complex behaviour. [9] Fixed exposure sequences will make the training conditions comparable, while exact reference answers will allow gains and losses to be checked directly. The resulting data and software will provide a basis for evaluating updates before their adoption. Later work can extend the comparison to systems whose actions determine the experience they receive. Hypotheses The primary claim requires both improved performance on new inputs and acceptable retention of earlier performance. Three secondary comparisons examine which parts of the procedure contribute to these outcomes; a final exploratory comparison tests transfer to different rule types. H1. Improvement beyond memory Combining stored memory with checked training and replay will produce higher accuracy on new inputs than memory alone. Accuracy will be averaged across three learning rounds. The planning effect is a 10 percentage-point gain, with the statistical test against zero gain. H2. Retention of earlier learning After the third round and a process restart, the combined procedure will retain initial-task accuracy within five percentage points of the reference frozen after initial training. Both will be tested with memory disabled. This margin is the proposed maximum acceptable average loss for the study. Mean initial-task accuracy must also reach 60% for retention to be interpretable. H3. Contribution of stored memory Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples. H4. Contribution of correct training targets The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs. H5. Contribution of replay Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing. H6. Transfer to different rule types The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered. 3. Methods Participants/Sample The sample will consist of 100 independently generated task sequences, or curricula, following a separate 12-curriculum pilot. Each curriculum will be tested under all six conditions over three learning rounds. Environment parameters and adaptation seeds will be drawn independently between curricula, while the base-model family and task generators remain fixed. The analysis therefore uses 100 matched units; repeated answers within a curriculum do not increase that independent sample. Before collecting model responses, we will check that each generated environment terminates within its specified limits and agrees with its reference implementation. Invalid generator outputs will be recorded and replaced using a fixed seed rule. Valid tasks will remain in the analysis regardless of difficulty, initial performance or training failure. Training examples, update-selection queries and final-test inputs will be separate, and pilot examples will not enter the main sample. Power calculations use differences between curriculum-level scores. [4] With 100 curricula and one-sided alpha 0.025, H1 has 99.9% power under an assumed 10-point gain and paired standard deviation of 20 points. H2 has 91.0% power under an assumed zero loss, paired standard deviation of 15 points and a five-point margin. The probability of passing both tests is at least 90.9% under these assumptions, excluding the initial-accuracy requirement. These values are sensitive to the assumptions: H1 power falls to 69.7% for a five-point gain, and H2 power falls to 37.8% if the true loss is 2.5 points. The historical studies cannot supply reliable estimates of variation between independent training runs. The companion protocol provides the full calculation and sensitivity analysis; any revision following the pilot will be registered before main collection. Materials and Procedures Each curriculum will begin with four environments and introduce four more in each round, covering modular affine updates, eight-state transition systems, symbol substitutions and binary cellular automata. Hidden parameters and test inputs will be drawn from the generators specified in the companion protocol. Two independently written reference implementations must agree before collection. Their known rules allow exact answers to be calculated while the learner receives only exposure examples and task instructions. All conditions will start from a shared checkpoint trained on 48 checked examples. Each subsequent round will supply 48 new examples. Conditions using replay will add 24 earlier examples; training without replay will reuse new examples to match the number of training records and optimisation steps. The six conditions are: A. Frozen reference: no further training or retrieved examples. B. Memory alone: retrieved examples, without further training. C. Training with replay: checked updates and replay, without retrieval. D. Combined procedure: the trained checkpoints from C, evaluated with retrieval. E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval. F. Training without replay: checked new targets and retrieval, without earlier training examples. For conditions using memory, retrieval will return the relevant environment's 12 exposure examples in a fixed order. Test answers will remain outside both memory and training. Sharing checkpoints between C and D isolates the effect of retrieval, while the separate comparisons with E and F test target correctness and replay. Prediction adapters will remain fixed throughout. The proposed starting model is Qwen3-8B. Pilot settings are rank-8 LoRA on 16 layers, scale 20, Adam at learning rate 0.00001, batch size one and 200 steps per update. Training sequences will be capped at 1,024 tokens. Evaluation will use greedy decoding, an 8,192-token context limit and a 256-token output cap, requesting short JSON answers. The model will have no calculator, internet or reference-simulator access during evaluation. The pilot will establish the configuration; model revisions, software, hardware, seeds and settings will then be fixed for the main study. Evaluation will load checkpoints in a separate process and use a fresh context for each query. Before an update is adopted, a separate set of 24 new-task and 24 initial-task queries will compare it with the current checkpoint, with memory disabled. Adoption requires at least one additional correct new-task answer, no more than one lost initial-task answer, and successful loading and isolation checks. A rejected or failed update will leave the preceding checkpoint in place. Because these decisions affect what is tested later, the primary analysis will evaluate the procedure including its update-selection rule. Figure 1. Proposed design. Each of the 100 main curricula receives all six conditions. Training with replay includes 24 earlier examples per round; training without replay reuses new examples to match the training budget. Update selection and final evaluation use separate inputs. The 12-curriculum pilot is a separate sample. After each round, we will evaluate 24 new queries and the same 24 initial-task queries in every condition. Retention tests will always disable memory. A final 24-query test will assess different rule types using identical supplied examples. This produces 103,200 main-study answers within the 100 curricula. Execution order will be randomised, and condition labels will be coded during scoring. Correctness will require valid JSON and exact agreement with the reference answer. An external reviewer will examine the protocol, a balanced 5% sample of main evaluation outputs and all checker disagreements. Brian Ashiundu will lead design, analysis and reporting under Dilate Technologies. The project already has a local development environment, research software and published study records. Grant funding will cover engineering, cloud execution and external review; the new generators and execution configuration will be validated in the pilot. Engineering and review appointments remain to be confirmed. During Months 1–2, we will develop the generators and independent checkers, complete the pilot, measure resource use and register the main protocol. Collection in Months 3–4 will depend on pilot readiness and a cost estimate that fits the remaining funds. Month 5 will cover the scoring audit and analysis, followed by the paper and data release in Month 6. Monthly public updates will report progress, spending and deviations. If the pilot shows that the planned study is not feasible, we will publish its results and propose a revised scope before committing the remaining main-study budget. Planned Analyses The H1 outcome will be each curriculum's mean accuracy across the three new-input rounds; H2 will use final initial-task accuracy. We will estimate the paired differences D−B and D−A in percentage points, with two-sided 95% t confidence intervals. H1 will pass if the lower bound exceeds zero and H2 if it exceeds −5 points. Both criteria, together with the initial-accuracy requirement, must be met for the primary claim. A sensitivity analysis will resample complete curricula 10,000 times, and results by task family will show whether average gains conceal particular losses. H3–H5 will use paired two-sided tests with Holm correction across the three comparisons at 0.05. Transfer, update acceptance and resource costs will be exploratory. Wrong answers, invalid JSON, refusals and output-limit failures will score zero. Infrastructure failures may receive at most two retries with the same input, seed and checkpoint; unresolved answers will score zero, with additional bounds showing the effect of more favourable assignments. Collection will use the registered sample size rather than stop according to significance, and deviations will be reported with the achieved sample and uncertainty. These analyses distinguish improvement with acceptable retention from improvement accompanied by excessive loss. A null or uncertain difference would leave the proposed advantage unestablished, while low initial accuracy would limit conclusions about retention. The paper will report these outcomes alongside any supported effects. Ethics and Data Management The study will generate computational tasks and model outputs, without recruiting people or using private conversation records. Dilate Technologies will confirm applicable review requirements before collection; no approval or exemption is currently claimed. Experimental copies will be isolated from the live system, and test answers will be kept outside training stores. The protocol, generator code and commitments to the test seeds will be public before main collection. Within 30 days of the final report, we will release the seeds, synthetic examples, raw outputs, scores, failures, update decisions and analysis code through Loopseed and a persistent archive. Reuse licences will follow ownership and model terms. Credentials and administrative records will remain separate, and live research sessions will use demonstration tasks without exposing main-test outcomes during development. 4. Pilot Data The first Loopseed paper provides evidence that the project can run and document controlled comparisons, as well as identifying the limitations this study addresses. [1] On new numerical inputs within trained mathematics types, additional response training produced 43 accepted answers out of 64, compared with zero for both the base model and shuffled control. On harder mathematics, however, the candidate and shuffled control each produced only 2 accepted answers out of 81. Figure 2. Recorded mathematics results, reproduced unchanged from the first paper. The two panels use different tasks and output conditions. Their counts are separate comparisons and cannot be combined into an overall success rate. A later update lowered validation loss but produced 12 gains, 17 losses and 55 ties across 84 paired comparisons. Hidden-rule work also showed limited acquisition and reuse: 2 of 6 rules were acquired correctly, and 3 of 12 later predictions were correct using the system's own records. Because these studies used different tasks and scoring rules, their results cannot be pooled into an overall success rate or treated as independent training replications. The public paper page links supporting records, including 1,024 mathematics replies and 155 whole-reply review decisions. Figure 3. Recorded continuation results, reproduced unchanged from the first paper. Panel A evaluates a prediction update. Panel B shows a separate response update, including seven losses on earlier skills. The 55 tied comparisons are not plotted as bars. These results motivate testing answers and retention directly, alongside training loss. The proposed pilot will assess whether the new workflow can support the main comparison. It will check reference agreement, separation of training and test data, behaviour after restart, runtime, and floor or ceiling effects. Readiness requires initial accuracy of at least 60% overall with no family below 40%, fewer than 10% unresolved infrastructure or output-limit failures, and projected costs within the available funds. All pilot results will be retained when the main design is fixed; the pilot will not be used as a powered test of learning. 5. Budget We propose a six-month budget of USD 25,000, including platform fees. Existing local facilities will support development, while funding will cover the research, implementation, cloud execution and review required for this study. The following allocations assume paid project time and are estimates to be confirmed before expenditure: Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation. Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution. External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements. Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour. Storage and backups: $300: six months at $50 for model files, outputs and backup copies. Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation. Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5] Contingency: $1,650: for additional runtime, price changes and other confirmed project costs. The compute allowance is based on a workload scenario: 1,120 training jobs averaging 30 minutes would use 560 GPU-hours, while approximately 212,352 pilot and main evaluation queries, including update-selection queries before caching, would use 590 hours at an average of 10 seconds each. A further 50 hours allows for setup and retries. The $2 hourly allowance is informed by published GPU prices, including an A100 80 GB listing at $1.59 per hour when checked on 13 September 2026. [7] Runtime and hardware suitability still need to be measured together in the pilot; the allowance is not a performance forecast or reserved capacity. The review allocation allows 12 hours for the protocol, about 43 hours for the 5,160-output sample at 30 seconds per output, and five hours for disagreements and deviations. The pilot will also check this workload. Personnel hours are part-time allocations across six months, with separate records for research, engineering, review and documentation. After the platform-fee allowance, $22,750 remains for $21,100 in direct costs and $1,650 contingency. The detailed budget assigns $7,460 in direct costs to the first two months and $13,640 to the main study, audit and release. These stages let the measured pilot costs inform the decision to proceed. If quotes or runtime exceed the available funds, the scope and sample plan must be reconsidered before main registration. Independent replication and a second model family would require a later funding phase. 6. References Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials. Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript. Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article. Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article. ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026). Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper. Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026). Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper. Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book.21:["$","$L4a",null,{"children":["$","$L4",null,{"topBanner":["$","$L4b",null,{"work":{"id":32899,"type":"$undefined","contentType":"preregistration","title":"Learning and Retention from Verified Experience","slug":"dynamical-synthesis-learning-and-retention-from-verified-experience","createdDate":"2026-09-12T22:56:09.972699Z","createdByUserId":201779,"updatedDate":"2026-09-22T13:41:15.948688Z","publishedDate":"2026-09-12T22:56:09.972699Z","authors":[{"authorProfile":{"id":8658950,"fullName":"Brian Ashiundu","firstName":"Brian","lastName":"Ashiundu","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2026/06/05/blob_Yi1vStJ","headline":"Research Fellow at Wolfram institute","profileUrl":"/author/8658950","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":201779},"description":"I'm currently doing research in computational biology; understanding the intersection of artificial and biological systems. Other interests include cryptography, number theory, game theory, and computational neuroscience.","createdDate":"2026-06-05T18:40:08.174919Z","education":[{"id":1534,"name":"Dedan Kimathi University of Technology","year":{"label":"2018","value":"2018"},"major":"Computer Science","degree":{"label":"Bachelor's Degree","value":"BA"},"summary":"Computer Science BA '18, Dedan Kimathi University of Technology","is_public":true,"university":{"id":1534,"city":"","name":"Dedan Kimathi University of Technology","state":"","country":"Kenya","updated_date":"2020-11-24T21:44:27.594405Z"}}],"twitter":"https://x.com/bm_asheuh","facebook":"$undefined","linkedin":"https://www.linkedin.com/in/brian-a-007241135/","googleScholar":"https://scholar.google.com/citations?hl=en&view_op=list_works&authuser=1&gmla=AKCpqFxSdqmIVCGk7ItA1ASB9OU9A1GZWYVhISElRZRWm6yIk7jsTAmcgpblu6yXObt3uLutv39hJOG1_ppGWnZ7WzP4&user=hsg80rMAAAAJ","orcidId":"https://orcid.org/0009-0008-5238-660X","isOrcidConnected":true,"isClaimed":true,"userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":8658950,"is_public":true,"is_removed":false,"is_removed_date":null,"user":201779,"first_name":"Brian","last_name":"Ashiundu","created_date":"2026-06-05T18:40:08.174919Z","updated_date":"2026-09-21T16:45:32.982230Z","description":"I'm currently doing research in computational biology; understanding the intersection of artificial and biological systems. Other interests include cryptography, number theory, game theory, and computational neuroscience.","h_index":1,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2026/06/05/blob_Yi1vStJ","author_score":0,"university":null,"orcid_id":"https://orcid.org/0009-0008-5238-660X","openalex_ids":[],"education":"$21:props:children:props:topBanner:props:work:authors:0:authorProfile:education","headline":"Research Fellow at Wolfram institute","facebook":null,"twitter":"https://x.com/bm_asheuh","linkedin":"https://www.linkedin.com/in/brian-a-007241135/","google_scholar":"https://scholar.google.com/citations?hl=en&view_op=list_works&authuser=1&gmla=AKCpqFxSdqmIVCGk7ItA1ASB9OU9A1GZWYVhISElRZRWm6yIk7jsTAmcgpblu6yXObt3uLutv39hJOG1_ppGWnZ7WzP4&user=hsg80rMAAAAJ","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":9,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":2,"is_orcid_connected":true,"reputation":100,"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true}},"isCorresponding":false,"position":"middle","raw":"$21:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"}],"abstract":"$4c","doi":null,"journal":"$undefined","formats":[],"license":"$undefined","pdfCopyrightAllowsDisplay":true,"figures":[],"versions":[],"metrics":{"votes":6,"adjustedScore":"$undefined","comments":10,"saves":0,"reviewScore":0,"reviews":0,"earned":0,"views":0},"unifiedDocumentId":9366982,"registeredReportId":null,"postType":"PREREGISTRATION","fundraise":"$undefined","note":{"id":58070,"access":"WORKSPACE","organization":{"id":196976,"memberCount":1,"userPermission":"$undefined","createdDate":"2026-06-05T18:40:08.249522Z","coverImage":null,"description":"","name":"Brian Mboya's Org","slug":"brian-mboyas-org"},"createdDate":"2026-09-12T21:26:44.387462Z","updatedDate":"2026-09-21T16:22:57.343405Z","title":"Learning and Retention from Verified Experience","isRemoved":false,"post":{"id":32899,"slug":"dynamical-synthesis-learning-and-retention-from-verified-experience","contentType":"preregistration","documentType":"PREREGISTRATION","moderationStatus":"APPROVED","fundraise":{"id":1143,"amountRaised":{"usd":0,"rsc":0},"goalAmount":{"usd":25000,"rsc":428446.41745504964},"contributors":{"numContributors":0,"topContributors":[]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2026-09-12T22:56:10.023440Z","endDate":"2026-12-11T22:56:08.515000Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":1143,"amount_raised":{"usd":0,"rsc":0},"goal_amount":{"usd":25000,"rsc":428446.41745504964},"contributors":{"total":0,"top":[]},"status":"OPEN","goal_currency":"USD","start_date":"2026-09-12T22:56:10.023440Z","end_date":"2026-12-11T22:56:08.515000Z"}},"grant":"$undefined","doi":null,"authors":[{"authorId":8658950,"userId":201779,"name":"Brian Ashiundu","raw":{"id":8658950,"first_name":"Brian","last_name":"Ashiundu","user":201779}}],"contacts":"$undefined","image":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","raw":{"id":32899,"status":"APPROVED","authors":["$21:props:children:props:topBanner:props:work:note:post:authors:0:raw"],"unified_document":{"fundraise":"$21:props:children:props:topBanner:props:work:note:post:fundraise:raw","grant":null},"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","document_type":"PREREGISTRATION","slug":"dynamical-synthesis-learning-and-retention-from-verified-experience","doi":null}},"documentType":"PREREGISTRATION","proposalId":null,"registeredReportPrefill":null,"image":"uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","previewImage":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","authors":[{"authorId":8658950,"userId":201779,"name":"Brian Ashiundu","raw":{"id":8658950,"first_name":"Brian","last_name":"Ashiundu","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2026/06/05/blob_Yi1vStJ","user":201779}}],"grantSettings":null,"preregistrationSettings":{"goalAmount":"25000.00","durationDays":90,"isPublic":true,"nonprofit":null,"raw":{"duration_days":90,"goal_amount":"25000.00","goal_currency":"USD","is_public":true,"nonprofit_details":null,"nonprofit_id":null}},"selectedGrant":null,"raw":{"id":58070,"access":"WORKSPACE","authors":["$21:props:children:props:topBanner:props:work:note:authors:0:raw"],"grant_settings":null,"latest_version":{"id":585841,"src":null,"created_date":"2026-09-21T16:22:56.715785Z","created_via":"editor","plain_text":"$4d","json":"$4e","created_by":201779,"note":58070,"parent_version":null},"organization":{"id":196976,"member_count":1,"user_permission":null,"created_date":"2026-06-05T18:40:08.249522Z","updated_date":"2026-06-05T18:40:08.249534Z","cover_image":null,"description":"","name":"Brian Mboya's Org","note_created":false,"org_type":"PERSONAL","slug":"brian-mboyas-org","user":201779},"post":"$21:props:children:props:topBanner:props:work:note:post:raw","preregistration_settings":"$21:props:children:props:topBanner:props:work:note:preregistrationSettings:raw","registered_report_prefill":null,"selected_grant":null,"selected_grant_details":null,"unified_document":{"id":9366980,"is_removed":false},"created_date":"2026-09-12T21:26:44.387462Z","updated_date":"2026-09-21T16:22:57.343405Z","document_type":"PREREGISTRATION","image":"uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","preview_img":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","title":"Learning and Retention from Verified Experience","created_by":201779},"content":null,"versionId":585841,"versionDate":"2026-09-21T16:22:56.715785Z","plainText":"$4f","contentJson":"$50","versionCreatedVia":"editor"},"previewContent":"$51","contentUrl":"https://storage.prod.researchhub.com/uploads/post_discussion/2026/09/12/RH-POST-PREREGISTRATION-USER-201779_LGxqGS0.txt","tips":[],"peerReviews":[],"image":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","moderationStatus":"APPROVED","raw":{"score":6,"user_flag":null,"authors":["$21:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"],"id":32899,"created_by":{"id":201779,"author_profile":{"id":8658950,"is_public":true,"is_removed":false,"is_removed_date":null,"user":201779,"first_name":"Brian","last_name":"Ashiundu","created_date":"2026-06-05T18:40:08.174919Z","updated_date":"2026-09-21T16:45:32.982230Z","description":"I'm currently doing research in computational biology; understanding the intersection of artificial and biological systems. Other interests include cryptography, number theory, game theory, and computational neuroscience.","h_index":1,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2026/06/05/blob_Yi1vStJ","author_score":0,"university":null,"orcid_id":"https://orcid.org/0009-0008-5238-660X","openalex_ids":[],"education":[{"id":1534,"name":"Dedan Kimathi University of Technology","year":{"label":"2018","value":"2018"},"major":"Computer Science","degree":{"label":"Bachelor's Degree","value":"BA"},"summary":"Computer Science BA '18, Dedan Kimathi University of Technology","is_public":true,"university":{"id":1534,"city":"","name":"Dedan Kimathi University of Technology","state":"","country":"Kenya","updated_date":"2020-11-24T21:44:27.594405Z"}}],"headline":"Research Fellow at Wolfram institute","facebook":null,"twitter":"https://x.com/bm_asheuh","linkedin":"https://www.linkedin.com/in/brian-a-007241135/","google_scholar":"https://scholar.google.com/citations?hl=en&view_op=list_works&authuser=1&gmla=AKCpqFxSdqmIVCGk7ItA1ASB9OU9A1GZWYVhISElRZRWm6yIk7jsTAmcgpblu6yXObt3uLutv39hJOG1_ppGWnZ7WzP4&user=hsg80rMAAAAJ","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":9,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":2,"is_orcid_connected":true,"reputation":100,"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true},"balance":null,"balances":null,"created_date":"2026-06-05T18:40:08.144572Z","has_seen_first_coin_modal":false,"has_seen_orcid_connect_modal":false,"is_funder":false,"is_staking_opted_in":true,"is_suspended":false,"probable_spammer":false,"is_verified":true,"moderator":false,"reputation":100,"updated_date":"2026-06-05T18:40:08.144606Z","upload_tutorial_complete":false,"hub_rep":null,"time_rep":null},"created_date":"2026-09-12T22:56:09.972699Z","discussion_count":10,"document_type":"PREREGISTRATION","doi":null,"editor_type":"CK_EDITOR","full_markdown":"$52","grants":[],"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/201779/5988e502-5423-4cf6-a6a9-d1052a438709/loopseed-proposal-cover.png","is_latest_version":null,"is_removed":false,"is_root_version":true,"note":"$21:props:children:props:topBanner:props:work:note:raw","peer_reviews":[],"post_src":"https://storage.prod.researchhub.com/uploads/post_discussion/2026/09/12/RH-POST-PREREGISTRATION-USER-201779_LGxqGS0.txt","preview_img":null,"registered_report_id":null,"renderable_text":"$53","reviewed_by":null,"reviewed_date":null,"slug":"dynamical-synthesis-learning-and-retention-from-verified-experience","status":"APPROVED","title":"Learning and Retention from Verified Experience","unified_document_id":9366982,"unified_document":{"id":9366982,"documents":[{"id":32899,"title":"Learning and Retention from Verified Experience","slug":"dynamical-synthesis-learning-and-retention-from-verified-experience"}],"created_by":{"id":201779,"author_profile":{"id":8658950,"first_name":"Brian","last_name":"Ashiundu","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2026/06/05/blob_Yi1vStJ"}},"reviews":{"avg":0,"count":0},"is_public":true,"is_removed":false,"document_type":"PREREGISTRATION"},"version_number":1,"updated_date":"2026-09-22T13:41:15.948688Z","bounty_type":null},"aiPeerReview":null,"linkedGrant":null,"isPublic":true},"metadata":{"id":9366982,"score":6,"metrics":{"votes":6,"comments":10,"saves":0,"reviewScore":0,"conversationComments":10,"reviewComments":0,"bountyComments":0},"fundraising":{"id":1143,"amountRaised":{"usd":0,"rsc":0},"goalAmount":{"usd":25000,"rsc":428446.41745504964},"contributors":{"numContributors":0,"topContributors":[]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2026-09-12T22:56:10.023440Z","endDate":"2026-12-11T22:56:08.515000Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":1143,"amount_raised":{"usd":0,"rsc":0},"goal_amount":{"usd":25000,"rsc":428446.41745504964},"contributors":{"total":0,"top":[]},"status":"OPEN","goal_currency":"USD","start_date":"2026-09-12T22:56:10.023440Z","end_date":"2026-12-11T22:56:08.515000Z"}},"bounties":[],"activeBounties":0,"closedBounties":0},"updatesCount":0,"preTitle":"$L54"}],"rightSidebar":"$L55","children":"$L56"}]}] 58:T5c00,

Project host: Dilate Technologies · Programme: Loopseed

1. Abstract

2. Introduction

Hypotheses

H1. Improvement beyond memory

H2. Retention of earlier learning

H3. Contribution of stored memory

Providing stored examples will improve new-input accuracy relative to the same trained checkpoints evaluated without those examples.

H4. Contribution of correct training targets

The combined procedure will outperform an otherwise matched procedure trained with shuffled new question–answer pairs.

H5. Contribution of replay

Including earlier examples during training will improve final retention relative to training on new examples alone, with memory disabled during testing.

H6. Transfer to different rule types

The combined procedure will outperform the frozen reference on rule types excluded from training, with identical example context supplied to both. This comparison is not separately powered.

3. Methods

Participants/Sample

Materials and Procedures

A. Frozen reference: no further training or retrieved examples.
B. Memory alone: retrieved examples, without further training.
C. Training with replay: checked updates and replay, without retrieval.
D. Combined procedure: the trained checkpoints from C, evaluated with retrieval.
E. Shuffled-target control: shuffled new training targets, with correct replay and retrieval.
F. Training without replay: checked new targets and retrieval, without earlier training examples.

Planned Analyses

Ethics and Data Management

4. Pilot Data

5. Budget

Research lead: $7,200: 240 hours at $30 for design, supervision, analysis and paper preparation.
Research engineering: $8,000: 320 hours at $25 for generators, checkers, the training workflow and reproducible execution.
External review: $2,400: 60 hours at $40 for protocol review, the scoring sample and investigation of disagreements.
Cloud compute: $2,400: 1,200 GPU-hours at a planning rate of $2 per hour.
Storage and backups: $300: six months at $50 for model files, outputs and backup copies.
Release documentation: $800: 40 hours at $20 for the data dictionary, archive and reproduction instructions, separate from software implementation.
Platform fee allowance: $2,250: 9% of the gross request, using ResearchHub's published rate. [5]
Contingency: $1,650: for additional runtime, price changes and other confirmed project costs.

6. References

Ashiundu, B. Dynamical Synthesis: Learning through Interaction. Dilate Technologies (2026). Paper and supporting materials.
Hu, E. J. et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (2022). LoRA manuscript.
Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). Article.
Lakens, D. Sample size justification. Collabra: Psychology 8, 33267 (2022). Article.
ResearchHub. Funders & Donors. ResearchHub documentation. Platform funding fees (accessed 13 September 2026).
Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T. P. & Wayne, G. Experience replay for continual learning. In Advances in Neural Information Processing Systems 32 (2019). Conference paper.
Runpod. GPU Cloud Pricing. Published GPU and storage rates (accessed 13 September 2026).
Sutton, R. S. Integrated architectures for learning, planning, and reacting based on approximating dynamic programming. In Proceedings of the Seventh International Conference on Machine Learning, 216–224 (1990). Conference paper.
Wolfram, S. A New Kind of Science. Wolfram Media (2002). Online book.

22:[["$","$L57",null,{"work":"$21:props:children:props:topBanner:props:work","metadata":"$21:props:children:props:topBanner:props:metadata","content":"$58","authorPosts":[]}],"$L59","$L5a"]