Oncology Failure Atlas

AI Transparency

Where AI is used in the Oncology Failure Atlas — and, just as importantly, where it is not.

Pharma Innovation Research by Sebastian Azar · Current release: 11,525 production records · Last reviewed: June 2026

Deterministic OFA analytics vs AI-assisted research drafting

OFA cleanly separates two layers. The deterministic layer powers every dashboard view, every export, and every score the product surfaces. The AI-assisted layer is downstream of OFA analytics and is used only to help draft human-reviewable research summaries from already-filtered OFA evidence packs.

Deterministic (no LLM at render time): the dashboard, filters, MDIP™ distributions, phase × pattern heatmaps, mechanism harmonization, exports (CSV / XLSX / PDF), and RecurSignal™ are all deterministic code reading from precomputed Supabase tables and views.

AI-assisted (live, server-side): the OFA Research Brief Generator uses Anthropic Claude to draft a human-reviewable core research brief from the OFA-filtered evidence pack the user selects. Claude does not classify records, calculate RecurSignal, alter the database, browse the web, or independently verify claims. Claude output is a research draft, not a scientific conclusion.

Current AI use: Claude-assisted Research Brief Generator

  • Server-side Edge Function: claude-research-hub (Supabase Edge Function). The ANTHROPIC_API_KEY is stored as a server-side secret and is never exposed to the browser. There is no VITE_ANTHROPIC_API_KEY.
  • Model: Anthropic Claude (claude-sonnet-4-5).
  • Input: an OFA-filtered evidence pack — active filters, record counts, top MDIP™ patterns, top harmonized mechanisms, phase distribution, linked-signal availability, a small set of representative records, and limitations.
  • Output: a core research brief — pattern summary, knowledge gaps, hypotheses (each with an ofa_evidence_basis field citing specific evidence-pack facts), research questions, validation plan, and safety notes.
  • Human review required. Every brief is an evidence-bound draft for an expert reviewer. It is not a scientific conclusion.
  • Not advice. Briefs do not provide clinical, regulatory, investment, medical, or treatment advice.

What AI does not do

  • Does not assign MDIP categories. MDIP (Mapped Discontinuation Intelligence Patterns) is deterministic classification from public stop-reason language; Claude only receives MDIP category summaries as part of the filtered evidence pack.
  • Does not classify production records (MDIP for the live dataset is deterministic — see history below).
  • Does not compute RecurSignal™ — RecurSignal is fully deterministic.
  • Does not update, write to, or alter the trial database.
  • Does not make predictions, forecasts, or risk estimates.
  • Does not browse the web or independently verify external sources.
  • Does not replace expert scientific, clinical, regulatory, or investment review.
  • Does not infer hidden sponsor, regulatory, or investor motivation beyond what the evidence pack explicitly contains.

Classification history (for the record)

The current 11,525-record production dataset combines the original 5,000-record baseline (Phase 1) with 6,525 newly added records (Phase 2). The original baseline used a Claude-assisted classification workflow with a five-layer audit. Newly added records were classified using deterministic MDIP™ v4 rules applied primarily to the ClinicalTrials.gov why_stopped field — no LLM in the loop.

The dashboard does not call an LLM at render time. Mechanism filtering uses the harmonized mechanism source (trial_mechanism_harmonization.harmonized_mechanism_class), not the legacy raw public.trials.mechanism_class field.

second_pass = 100.0% — historical caveat

In the original baseline audit, the second_pass = 100.0% metric reflects a deterministic backfill, not a fully independent API re-run. It should not be interpreted as 100% AI agreement on an independent re-classification. The caveat applies only to the original Claude-assisted baseline; newly added records are deterministic.

RecurSignal™ v1 — deterministic

RecurSignal v1 is calculated deterministically across all 11,525 records from discontinuation category, trial phase, and classification confidence. Output is capped 5–95. Full-dataset average: 44.0.

RecurSignal is a historical recurrence signal only. It is not a prediction, probability, forecast, clinical risk score, investment signal, treatment guidance, or asset-quality score. Claude does not compute RecurSignal.

Scope limitations and non-claims

  • OFA does not determine the definitive cause of any trial's discontinuation.
  • OFA does not make clinical, regulatory, investment, medical, or treatment recommendations.
  • RecurSignal™ is a deterministic historical recurrence signal — not a prediction or risk score.
  • A NONE / NO_MATCH / missing CrossRef link means no linked signal currently available in OFA — not proof that no publication, scientific, regulatory, or safety evidence exists for that trial.
  • Strategic / portfolio classifications reflect publicly reported language only; OFA does not infer hidden sponsor rationale, and Claude is instructed not to either.
  • AI-assisted research briefs are drafts that require expert human review before any downstream use.

AI Transparency Notice

This page discloses the AI system used (Anthropic Claude via the claude-research-hub Supabase Edge Function), the stage at which it is applied (downstream research-brief drafting from OFA-filtered evidence packs), and the deterministic boundaries around it (dashboard, exports, RecurSignal, mechanism harmonization, and MDIP classification for newly added records are all deterministic).

This platform is operated by Sebastian Azar, Pharma Innovation Research — Woerden, Netherlands.

Oncology Failure Atlas · Pharma Innovation Research by Sebastian Azar · oncologyfailureatlas.com · MDIP™ and RecurSignal™ are proprietary frameworks. All rights reserved.