What ProtocolOS does

Every exclusion deserves a rationale. ProtocolOS helps a protocol team see which eligibility restrictions deserve reconsideration before a protocol is finalized, while keeping the safety, enrichment and statistical reasons those restrictions exist. It is built for clinical scientists, protocol authors and cross-functional protocol reviewers, not patients. It does not decide anything: it gathers the rationale, shows where it conflicts or is missing, and records the team's decision.

Eligibility criteria are deliberate trade-offs (safety, predictive or prognostic enrichment, reduced heterogeneity, interpretability). A broader criterion is not automatically better, and a peer trial that used a looser threshold is precedent, not proof.

Where AI is and is not used

The rationale engine (extraction, comparison, classification, verification) is deterministic and auditable; it is not an autonomous AI scientist. A language model, run with your own OpenAI or Anthropic key (the public server has none of its own), is used only to (a) critique a flag, and it may only downgrade it, and (b) turn a reviewer's free-text challenge into checkable claims. Every model output passes through the deterministic verifier. Documents are treated as data: text from PDFs is never executed as instructions, the model has no tools or write access, and every value shown links to its source page.

Pipeline

1. Protocol ingestion. The registry record and the protocol PDF posted to ClinicalTrials.gov are fetched and cached; the eligibility section is located and parsed with deterministic rules. Every value keeps its verbatim source line and page.
2. Regimen. Drugs in every arm come from registry interventions plus protocol pages with dosing context, then a small curated class-level pharmacology table (myelosuppressive, hepatic, renal, immunotherapy).
3. Comparators. Phase 3 trials in the same condition with overlapping drug classes and a posted protocol are discovered from the registry, ranked by drug-class overlap, and parsed with the same extractor. By default only trials that started before the reviewed trial are used.
4. FDA guidance. Passages from the July 2026 FDA eligibility guidances (laboratory values, performance status, washout periods) are matched to criteria. Each quote is verified verbatim against the official PDF at load time, with section, page and FDA line numbers.
5. Classification. A transparent rule over evidence stances: Review recommended, Evidence supports retaining, Insufficient evidence. A toxicity rationale supports having a threshold, never a specific value; a more permissive comparator never proves a restriction unnecessary.
6. Verification. Any proposed recommendation (from a user, an LLM, or the evaluation set) is checked deterministically: source existence, verbatim quotes, threshold values and units, direction, regimen contradictions, evidence sufficiency, and overclaims. An optional LLM critic can only downgrade.
7. Population context (secondary). NHANES 2017–2018 survey-weighted estimates of adults below a laboratory threshold, with design-based confidence intervals.

What it does not do

• It does not determine that any criterion is safe to relax, and does not estimate enrollment, cost or timeline effects.
• NHANES numbers describe the US general population, not people with cancer, and race/ethnicity is not a proxy for Duffy-null phenotype.
• Protocol PDFs on ClinicalTrials.gov are usually the final amendment, not the design-time version.
• The pharmacology table is curated at class level and not re-verified against current labels.