─── about
Calibration substrate for clinical-trial outcome prediction under data-scarcity.
─── pipeline
- 01
ctgov-fetch
Trial metadata, arm structure, outcome definitions pulled from CT.gov v2. Fixture cache for the four pinned demo NCTs; live lookup for the advanced path.
- 02
synthesis
Indication-specific dispatcher generates a TrancheSpec — outcome family, arm assignments, prior-precedent substrate. Today: scd-vaso-occlusive and sod1-als dispatchers; SMA, C9-ALS, and SOD1-presymptomatic next.
- 03
tabpfn-v2 inference
Tabular foundation model runs in-context over the synthesized substrate. No per-indication training; the synthesis carries the inductive bias.
- 04
ltt loss control
Learn-Then-Test multi-alpha calibration. Empirical risk bounded below user-specified alpha across the chosen loss surface — what “calibrated” means here, precisely.
- 05
discount-rate output
The calibrated forecast surface, expressed as a discount-rate table over outcomes-based receivables — the deliverable a design partner takes away.
- 06
audit substrate
Every stage emits a typed event. Replayable; the TrancheSpec and the calibration alphas are visible end-to-end.
─── where the bounds fail
- n < 45
- Below this floor LTT loss control degrades; surfaced pre-stream as bounds:ltt in the validator.
- indication outside coverage
- No trained dispatcher for the inferred indication → structured refusal with the trained-coverage list visible. Not a degraded forecast.
- paramType mismatch
- Outcome measure cannot be coerced into a supported family (negbin / TTE / ordinal …). Disclosed at submit, not at stage four.
─── what this is and isn't
The demo wraps finnegas-v0; it does not introduce methodology. The methodology core was released 2026-05-16. The surface lets you exercise it on a real NCT, see the audit trail, and read the refusal when one is the honest answer. It is not a model picker, not a workflow chatbot, and not a notebook in a tab.
─── contact
Ahmad Elmowag — Founder. Finnegas is pre-seed; no commercial sponsors.