Team (Phase I): Roles, Responsibilities & Execution
Core principle
LLM explains; Bayesian engine determines confidence.
PromptGenix is intentionally designed to be reviewer-traceable: probabilistic confidence (posterior mean + credible interval + evidence coverage) is computed by an evidence-weighted Bayesian decision engine, while the LLM is used only for interpretability— to produce human-readable rationales and cite supporting or conflicting sources.
Guardrail: “No evidence object → no claim.” The LLM cannot invent confidence or override posterior confidence.
Team structure
Key personnel & responsibilities
- PI / Technical Lead (Bioinformatics & Biostatistics): overall architecture; evidence feature design; Bayesian model specification; evaluation design; milestone delivery.
- Pipeline Engineering Lead: deterministic RNA-Seq + flow cytometry evidence generation; QC gates; metadata harmonization; version pinning and reproducible builds.
- ML / NLP Engineer (Interpretability Layer): constrained prompting; citation enforcement; claim filtering; template-driven narratives; safety + hallucination control.
- Product/UX & Reporting: reviewer-ready HTML/PDF reports; figure/table generation; traceability links (evidence IDs, configs, run logs).
- Advisory Support (as-needed): domain review (immunology/infectious disease), methods critique, and usability feedback on pilot public-data studies.
Deterministic pipelines
Evidence objects
Bayesian inference
LLM interpretability only
Reviewer traceability
Execution alignment to Phase I aims
- Aim 1 (Evidence layer): pipeline lead + PI define standardized evidence objects (effect size, uncertainty, context, reproducibility).
- Aim 2 (Decision engine): PI specifies prior construction + likelihood mapping; implements posterior confidence + CrI + coverage.
- Aim 3 (Reports & usability): UX/reporting + NLP engineer produce evidence-linked narratives; pilot KPI collection and feedback.
Reviewer-facing narrative: The “confidence” is not an LLM opinion—confidence is a posterior probability distribution produced by the Bayesian engine.
Responsibilities by work package
Clear ownership for reproducibility, inference, and interpretation.
| Work package | Primary owner | Accountability outputs |
|---|---|---|
| WP1 Evidence layer |
Pipeline Engineering Lead (implementation) + PI (specification of evidence features) | Deterministic pipeline runs; standardized evidence objects; QC logs; version-pinned configs |
| WP2 Bayesian engine |
PI / Technical Lead | Prior construction rules; likelihood mapping; posterior confidence + 95% CrI; calibration tests; coverage metrics |
| WP3 Interpretability & reports |
ML/NLP Engineer (constraints) + Product/UX (report format) + PI (review/approval) | Evidence-linked narrative; citation enforcement; “no evidence → no claim” checks; HTML/PDF exports |
| WP4 Pilot evaluation |
PI (study design) + Advisory Support (domain review) | KPI definitions; pilot runs on public datasets; usability feedback; reviewer-ready summaries |
Quality & governance
Operational controls that protect scientific validity.
Reproducibility & traceability
- Deterministic configs: pinned versions, run manifests, and reproducible parameter sets.
- Evidence IDs: each hypothesis links to evidence objects (effect sizes, uncertainty, context).
- Audit trail: run logs + checksums support reviewer trust and rerun verification.
LLM constraints (interpretability only)
- Claim gating: narratives must reference evidence objects and/or citations.
- No confidence authoring: LLM cannot produce or modify posterior confidence outputs.
- Uncertainty first: reporting requires posterior + CrI + supporting/conflicting coverage.
Bottom line: The Bayesian engine computes “confidence”; the LLM explains “why,” constrained to traceable evidence.
Phase II readiness
Team skills map directly to scalable delivery.
Phase I establishes technical feasibility and reviewer trust through deterministic pipelines, evidence-weighted Bayesian inference, and strict interpretability controls. Phase II will expand dataset coverage, strengthen calibration and benchmarking, and support secure deployments and subscription-based delivery while preserving the same “LLM explains; Bayesian engine computes confidence” principle.