One engine. Fifty-seven predictions. One counterfactual panel on every one.
The whole drug-development lifecycle is one prediction surface. Talaria designs it as fifty-seven output-adapter heads over a single shared substrate — each head a calibrated read on a different decision, from target validation to patent challenge. This atlas lays the surface out by lifecycle stage and marks, honestly, what ships today and what is still on the bench.
Pre-deployment · not for clinical use1 shipped · 12 in the Pareto build queue · the rest rostered.
Fifty-seven heads, laid out by lifecycle stage
Filter by status or stage, and toggle the impact / moat / white-space tiers. Focus or hover any head to inspect it. Head #49 is an international-regulator bundle — one chip that opens to four sub-heads.
The moat is the panel, not the number
A probability you cannot act on is a number. Every Talaria head ships a counterfactual panel — the design levers that move the result — carried as a contract, not a courtesy. Behind it sits an audit ledger on every prediction, an adversarial Judge, CATE for heterogeneous subgroup effects, and a label-scarcity-first design that treats scarce honest outcomes as the binding constraint. That stack is the same for all fifty-seven heads.
One substrate. Seven base learners. Fifty-seven heads.
The architecture note that makes the surface tractable: one shared substrate, 57 output-adapter heads, 7 base learners (Lock 8 is at cap, 7/7). A new prediction is a calibrated head plus a counterfactual panel — not a new model. New capacity goes above the base learners, never as an eighth learner.