PHAROS · HYGEIA active
Earlier safety signals.
No causality cosplay.
JourdanLabs’ clinical-safety flagship — COSMIC applied to pharmacovigilance. PHAROS reads FAERS-scale adverse-event reports, ranks drug–event pairs that deserve review, and refuses to score when confidence isn’t there. Triage for safety teams — not medical advice, not regulatory action, not proof of causality.
pair tofacitinib · thrombosis · FAERS corpus through 2023 Q4
comparators ROR elevated · PRR below threshold · BCPNN elevated
trajectory ECLIPSE — rising quarter-over-quarter since 2021 Q2
stability PULSAR — holds under adversarial drop tests
confidence AURORA calibrated · above signal threshold
PRIORITIZE
alt MONITOR · ABSTAIN when evidence is thin or unstable
Illustrative console — real pipeline runs deterministic engines + AURORA gate on sealed FAERS snapshots.
Why it exists
Post-market drug safety still runs on four disproportionality formulas from 1998–2002. They flag patterns but ship raw scores with no calibrated confidence — so teams drown in false alarms while real Black Box Warnings arrive months or years after the signal was already visible in FDA’s own database. PHAROS matches the best legacy methods on detection and lead time, and adds what none of them have: honest abstention.
Five-engine pipeline
Deterministic Python pipeline — LLM-free in every scoring path, bit-for-bit identical across runs. 947/947 tests passed.
SIGNAL benchmark
SIGNAL is the benchmark receipt — product on the glass is PHAROS. Ground truth sealed before labels were opened.
pairs 36 FDA Black Box drug–event pairs (2018–2025)
corpus 17.76M unique FAERS reports · 44 quarters (2015 Q1 – 2025 Q4)
integrity 0 temporal-leakage violations · predictions SHA-256 sealed pre-label
determinism 5/5 runs identical · frozen-output hash verified
benchmark SIGNAL F1 0.639 benchmark name — product is PHAROS
Headline results
Frozen run against 36 labeled pairs. PHAROS at statistical parity with ROR and BCPNN under bootstrap resampling.
Calibrated abstention
Old methods always return a number — even when they shouldn’t. PHAROS ranks by confidence and stops guessing.
Rank PHAROS predictions by maximum confidence: the top 23 are 100% correct. PRR makes its first ranked error at position 21.
Early-warning examples
Months before FDA’s actual Black Box announcement — from sealed backtests, not retroactive storytelling.
Plain English
PHAROS reads the FDA’s giant database of patient side-effect reports and looks for drugs that might be hurting people earlier than the agency usually catches them. It catches as many real dangers as the best 30-year-old formulas and flags them roughly two years sooner — but unlike those formulas, it tells you when it isn’t sure enough to act. That refusal is the product: fewer false alarms, more time on the signals that matter.
Boundaries
PHAROS lives under HYGEIA — regulated clinical safety and pharmacovigilance. It is not HELIX. Output is safety-review triage from spontaneous report data: prioritize, monitor, or abstain. It does not prove causality, does not replace a qualified drug-safety officer, and does not constitute medical or regulatory advice.
First external-domain validation of COSMIC in medicine · LLM-free scoring · 947/947 tests · sealed SIGNAL receipts on the benchmark wall.