Calibrated forecasting on n<45 trial designs.
Refuses honestly where bounds fail.

A live demo wraps the v0 pipeline — CT.gov fetch → synthesis → TabPFN-v2 inference → LTT loss control → calibrated discount-rate output → full audit substrate. Try it on a real NCT.

─── what the demo demonstrates

1. 01

   calibrated synthesis

   CLIMB-121 → tranche-b spec, outcome family voc-count-per-year negbin, paired csscd-prior-control arm. Streams stage-by-stage through the audit substrate.

   → view in demo
2. 02

   honest refusal

   ENDEAR → indication outside trained coverage (sma-pediatric). Refusal trace surfaced pre-stream with the trained-coverage list visible. Not an error — the differentiated moment.

   → view in demo
3. 03

   audit substrate

   Every stage emits a typed event. The TrancheSpec the inference ran over is visible in the right rail. No black box.

   → view in demo

─── what's next

generalising the coverage frontier. The current demo refuses on three of four pinned NCTs because v0 has trained dispatchers for scd-vaso-occlusive and sod1-als only. The next release extends v0 to dispatch across SMA, C9-ALS, and SOD1-presymptomatic — same calibration core, broader indication surface. The honest-refusal pattern stays where coverage still doesn't fit.

─── not what you've seen

Not a workflow chatbot. Not a model picker. Not a notebook in a browser tab. A calibration substrate for design-partner conversations about pre-seed indications where n is small and priors are noisy.