BioPromptLab: Prompt Protocols as a Validated Intervention for Three Biomedical GPT Agents

Funding call: AI Tools & AI Agents for Biomedical Research
Duration: 6 months (proof-of-concept)
Primary research artifact: Prompt Protocols (versioned, modular, auditable)
Primary scientific contribution: causal evidence (via controlled evaluation) that structured prompt protocols improve reliability/traceability/actionability in biomedical agent tasks

1. Background

1.1 Why this project, and why now

Biomedical researchers increasingly use GPT-like systems for literature synthesis, methods interpretation, and analysis planning. The bottleneck is not only model capability, but reliability: unsupported claims, “citation laundering” (citations that do not actually support a statement), incomplete or non-actionable outputs, and lack of provenance.

Most prompt engineering is not evaluated as science. Prompts are tuned informally, with little transparency about:

which prompt components cause improvements,
whether improvements generalize to new items,
whether prompting introduces systematic biases across subdomains,
and how to measure hallucination and traceability rigorously.

This RFP explicitly encourages AI agents, reuse of benchmarks/datasets, and fast proof-of-concept results within 6 months. Our project fits this by treating prompt protocols as an experimentally testable intervention and producing a public benchmark + evaluation rubric.

1.2 What the project is (and is not)

Is: A rigorous, preregistered study to build and validate prompt protocols for three GPT agents, using controlled closed-context tasks and blinded evaluation.
Is not: A claim that we solved biomedical knowledge; we do not aim to build a full biomedical platform, nor to provide clinical decision support, nor to run wet-lab experiments.

Why closed-context? If the model can use pretraining knowledge, evaluation is confounded by memorization. Closed-context forces the agent to operate only on the provided evidence packet or dataset card, turning the study into a test of prompt protocol discipline (evidence gating, uncertainty handling, auditability), not an “open-book trivia contest.”

1.3 The three GPT agents and why these tasks

We chose three tasks that are (i) central to biomedical research workflows, (ii) feasible to evaluate within 6 months, (iii) naturally suited to closed-context evaluation, and (iv) aligned with the RFP examples (literature synthesis, experimental planning, data analysis).

EvidenceSynthGPT — literature synthesis with explicit claim–evidence mapping
MethodsChecklistGPT — Methods → reproducibility checklist + ambiguity log
DataAnalysisPlanGPT — dataset card + question → preregisterable analysis plan (no code)

Justification: These cover a spectrum of biomedical agentic functions: synthesizing evidence, converting narrative methods into operational steps, and planning analysis with bias/confound checks. Each can be meaningfully benchmarked with structured outputs.

1.4 The key gap we address

We address the lack of validated prompt protocols that: [i] reduce unsupported claims; [ii] enforce evidence traceability; [iii] produce actionable, structured outputs; and [iv] remain robust across biomedical domains.

2. Hypotheses

H1 (Evidence discipline): Prompt protocols with evidence gating + structured schemas reduce Unsupported Claim Rate (UCR) and improve Claim–Evidence Alignment (CEA) versus a minimal baseline prompt.
H2 (Auditability gains): Adding an explicit Verifier Pass (self-audit step that checks each claim against evidence) further improves CEA and lowers UCR.
H3 (Actionability gains): Protocol-driven outputs improve actionability/completeness of methods checklists and analysis plans, measured via blinded rubric scoring.
H4 (Bias transparency and robustness): With preregistered stratification and subgroup evaluation, prompt protocols will not create hidden performance collapse in any subdomain; if subgroup degradation is observed, it will be quantified and reported (no post-hoc prompt changes on the locked test set).

3. Methods

3.1 Study design overview (prompt protocols as the intervention)

We compare four preregistered prompt conditions, all using the same base model/version:

BP (Baseline Prompt) —> short instruction “be accurate, cite evidence.”
P1 (Protocol v1) —> schema + evidence gating + refusal rules.
P2 (Protocol v2) —> P1 + Verifier Pass (self-audit).
P3 (Protocol v3) —> P2 + Critic Pass (separate critique agent reviews completeness/bias, then revision).

Why this ladder? It isolates causal contributions of structure and verification. We can quantify whether “more prompt machinery” yields real gains or just longer answers.

3.2 Prompt Protocol architecture (what exactly is being engineered)

Each Protocol is modular, allowing ablation:

Module A: Scope contract —> forbids external knowledge beyond the provided packet + forbids clinical advice + requires stating limitations)
Module B: Output schema —> required tables/fields + machine-checkable formatting rules (e.g., JSON-like blocks)
Module C: Evidence gating —> “Every claim must reference ≥1 evidence span from the packet; otherwise mark as ‘Insufficient evidence’ or refuse.”
Module D: Uncertainty calibration —> confidence levels anchored to evidence type (clinical trial excerpt vs observational vs mechanistic) and directness (explicit vs implied)
Module E: Verifier Pass (P2+) —> re-check each claim: does the cited span actually support the claim wording? if not, revise or downgrade/withdraw claim.
Module F: Critic Pass (P3) —> checks completeness and bias: missing key items? overclaiming? ignoring confounds? unclear parameters?

Why these modules? They map directly to known failure modes (hallucination, weak citations, missing parameters, overconfidence) and create measurable outputs.

3.3 Agent definitions (inputs, outputs, hard rules)

3.3.1 EvidenceSynthGPT

Input: research question + bounded evidence packet (4–8 excerpts, each with a source ID).
Output (required):

Claim Table (row = claim) with fields:
- Claim text (single sentence, falsifiable)
- Evidence span(s) (short quotes)
- Evidence source IDs
- Claim category (mechanistic / correlational / clinical / methodological)
- Confidence (High/Med/Low) with justification
Synthesis narrative that may only reuse claims already in the table
Limitations & Unknowns with explicit “cannot conclude because…”

Hard rules: [i] No claim without evidence span(s); [ii] No external knowledge; [iii] Must represent conflicting evidence if present.

Why this structure? It converts a vague “summary” task into auditable units that can be graded.

3.3.2 MethodsChecklistGPT

Input: Methods text excerpt + study type label.
Output (required):

Checklist by stages (Data → Preprocessing → Modeling → Statistics → QC → Reporting)
Parameter Registry (all numeric thresholds, versions, seeds, inclusion/exclusion)
Ambiguity Log: missing critical details + exact questions to resolve
Risk Flags: steps likely to prevent reproducibility

Hard rules: [i] Must label extracted vs inferred; inference should be minimized and never presented as fact; [ii] Missing info must be logged, not guessed.

Why this structure? Reproducibility failures often come from omitted parameters; this output forces the model to surface them.

3.3.3 DataAnalysisPlanGPT

Input: dataset card (variables, sample size, missingness, collection design, confounds) + research question.
Output (required):

Primary/secondary hypotheses
Variables & operational definitions
QC + preprocessing steps
Statistical plan (models/tests, assumptions, multiple testing if needed)
Confounding/bias plan (covariates, stratification, sensitivity analyses)
Reporting plan (tables/figures)
Validity “stop conditions” (when not to proceed)

Hard rules: [i] Must not claim results; [ii] Must explicitly list assumptions and failure modes.

Why this structure? It yields a preregistration-like plan and makes missingness/confounding explicit-common sources of bias.

3.4 Benchmark construction (avoiding bias, leakage, and loose ends)

3.4.1 Closed-context items only: Each item includes all necessary context (excerpts or dataset card). The agent is forbidden to use external knowledge. This makes evaluation fair and reproducible.

3.4.2 Stratified sampling (prevents hidden domain bias): For each benchmark we will predefine strata and target counts:

Domain strata: e.g., neurology, immunology, oncology, infectious disease
Evidence-type strata: clinical vs mechanistic vs mixed
Conflict strata: packets with conflicting results vs consistent packets
For Methods: assay types (omics, imaging, in vivo, cohorts)
For AnalysisPlan: dataset size tier, missingness tier, complexity tier

Preregistered rule: strata quotas must be met before finalizing the test set.

3.4.3 Development vs locked test set (no overfitting): Dev set (used to iterate prompts and rubrics) + Test set (locked; never used for prompt edits) + Freeze milestone (prompts are frozen before any test outputs are generated).

3.4.4 Copyright and openness: We will use (i) open-access papers or (ii) short excerpts within permissible limits. Benchmark will store only IDs and short excerpts + metadata, not full copyrighted texts.

3.5 Evaluation plan (blinded + automated)

3.5.1 Human evaluation (primary)

Two independent raters per item, blinded to prompt condition.
Rubrics are preregistered with clear anchors for each score level.
Adjudication by a third rater only when disagreement exceeds a threshold.

Inter-rater reliability will be computed (weighted kappa / Krippendorff’s alpha).

3.5.2 Automated checks (secondary, to reduce subjectivity)

Schema compliance (pass/fail + % fields present)
Evidence span validity (span must match provided packet)
Claim count, refusal count, ambiguity log completeness proxies

3.6 Failure modes and mitigation (pre-specified, no loose ends)

Potential failure: Protocol becomes too conservative and refuses too often.

Metric: Refusal appropriateness (human-rated).
Mitigation: documented in results; no prompt retuning on test set.

Potential failure: Longer prompts inflate verbosity without improving reliability.

Metric: delta in primary endpoints vs output length.
Mitigation: report efficiency trade-off; propose future compression.

Potential failure: Subgroup degradation (e.g., worse on one domain).

Mitigation: preregistered subgroup reporting + interaction analysis.

4. Plan for Statistical Analysis

Primary endpoints (per agent):

EvidenceSynthGPT: Unsupported Claim Rate (UCR: lower is better) + Claim–Evidence Alignment Score (CEA: higher is better)
MethodsChecklistGPT: Checklist Completeness (CC) + Ambiguity Detection Accuracy (ADA)
DataAnalysisPlanGPT: Actionability Score (AS) + Statistical Validity Score (SVS)

Models:

Mixed-effects regression with fixed effect = prompt condition (BP/P1/P2/P3)
Random intercepts for item and rater
Where appropriate, ordinal models for 0–5 scores; logistic models for binary outcomes

Multiple comparisons: Benjamini–Hochberg FDR at q=0.05 across primary endpoints.

Subgroup bias analysis: Report per-stratum performance + condition×stratum interaction tests.

Pre-registered reporting: Full score distributions, not only means + Effect sizes and confidence intervals + A failure taxonomy with representative anonymized examples

5. Budget

Lead Investigator / Junior Researcher: $4,000 … Tasks: preregistration package; prompt protocol design and versioning; schema design; ablation experiments; evaluation orchestration; analysis and manuscript leadership; open-data packaging.

Human evaluation / annotation: $1,000 … Tasks: blinded rubric scoring; claim–evidence verification; checklist completeness scoring; analysis plan actionability scoring; adjudication buffer.

Compute, storage, and tooling (In-kind Contributions from UFRN) … Tasks: batch inference runs; storage for benchmark artifacts and logs; reproducibility bundles.

6. References

Farquhar S, Kossen J, Kuhn L, et al. Detecting hallucinations in large language models using semantic entropy. Nature. 2024;630:625–630. doi:10.1038/s41586-024-07421-0.
Singhal K, et al. Large Language Models Encode Clinical Knowledge (MultiMedQA / Med-PaLM). arXiv. 2022. arXiv:2212.13138.
Lee J, Yoon W, Kim S, et al. BioBERT: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics. 2020;36(4):1234–1240. doi:10.1093/bioinformatics/btz682.
ruz Rivera S, Liu X, Chan A-W, et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ. 2020;370:m3210. doi:10.1136/bmj.m3210.
Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Series B. 1995;57(1):289–300. doi:10.1111/j.2517-6161.1995.tb02031.x.
Krippendorff K. Measuring the Reliability of Qualitative Text Analysis Data. Qual Quant. 2004;38:787–800. doi:10.1007/s11135-004-8107-7.
Fleiss JL. Measuring nominal scale agreement among many raters. Psychol Bull. 1971;76(5):378–382. doi:10.1037/h0031619.
Wei J, Wang X, Schuurmans D, et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv. 2022. doi:10.48550/arXiv.2201.11903.
Madaan A, Tandon N, Gupta P, et al. Self-Refine: Iterative Refinement with Self-Feedback. arXiv. 2023. arXiv:2303.17651.
Yao S, Zhao J, Yu D, et al. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv. 2022. doi:10.48550/arXiv.2210.03629.
Shinn N, Cassano F, Berman E, et al. Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv. 2023. arXiv:2303.11366.
Lewis P, Perez E, Piktus A, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv. 2020. doi:10.48550/arXiv.2005.11401.
Liu Y, Iter D, Xu Y, et al. G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment. arXiv. 2023. arXiv:2303.16634.

BioPromptLab: Prompt Protocols as a Validated Intervention for Three Biomedical GPT Agents

BioPromptLab: Prompt Protocols as a Validated Intervention for Three Biomedical GPT Agents

Funding Progress

Related Funding

Request for Proposals - AI Tools & AI Agents for Biomedical Research

Nonprofit Recipient

Topics

DOI