1. Extractor
The v4 pipeline uses an LLM extractor to convert a Phase 2 abstract into structured JSON. It captures endpoint outcomes, p-values, hazard-ratio context, trial design cues, and 1 to 3 verbatim evidence quotes.
Dive Deeper
How the BioLensTracker System Works
The current v4 design is intentionally split between explicit extraction and a calibrated forecaster so the process stays measurable, auditable, and reusable.
Technical Workflow
2. Feature Table
The extracted record is transformed into typed, model-ready features such as log-scaled p-values, confidence-interval width, hazard-ratio distance from null, and per-arm enrollment summaries.
3. Forecaster + Calibration
A small supervised model consumes those features and outputs a calibrated probability of Phase 3 success. Calibration matters because the product is built around usable probability estimates, not just ordering names from best to worst.
4. Abstain + Governance
Predictions near the middle can be marked abstain instead of being forced into a weak call. Runs are also evaluated on time-aware holdouts and measured with gates such as AUC, Brier score, and calibration error before they are trusted.
Operational Workflow
Sourcing and Ranking
On a recurring cadence, upcoming Phase 3 catalysts are sourced, normalized, and scored. High-confidence, non-abstained names are pushed forward into the real shortlist.
Investability Audit
Model output narrows the field, but final candidates still go through a human audit for tradability, sponsor quality, price discipline, and broader pipeline context.
Monitoring and Messaging
Once a round is active, the system shifts into monitoring mode. Subscribers then get new-round messages, followed by confirmed outcome alerts when definitive Phase 3 results arrive.