Will the trial succeed? Diviner started with a human expert crowd. Diviner AI asks whether the same discipline can be reproduced with LLM personas.
Diviner started with a human expert crowd: independent forecasts from experienced drug developers, aggregated into one probability. Diviner AI asks whether that same discipline can be reproduced with LLM personas.
Diviner began as a human venture: a network of veteran drug developers — people with decades in the industry — independently forecasting the outcomes of clinical trials, their individual judgments aggregated into a single probability. Structured collective forecasting applied to clinical-trial outcomes was novel on its own, and it taught us what the approach needed: genuine independence between forecasters, diverse professional perspectives, and disciplined aggregation.
Diviner AI automates the crowd. A panel of AI personas — each prompted to reason from a different professional vantage point — independently studies each trial and produces its own forecast; the panel's estimates are then aggregated, just as the human crowd's were. What once required recruiting, scheduling, and compensating an expert network now runs as a pipeline.
Each forecast rests on an evidence briefing the system researches for itself, glass-box style: it pulls from clinical-trial registries, SEC filings, medical literature, and news, with every query and every source logged. The personas reason from the assembled evidence — not from vibes.
Autonomous evidence gathering across registries, filings, literature, and news.
Findings become a structured, source-logged research briefing per trial.
AI personas independently forecast from the briefing — no peeking at each other.
Independent estimates combine into one calibrated probability.
The hard question for any AI forecasting system is: how do you know it works? The obvious test — run it against past trials with known outcomes — has a flaw: modern models have read the internet, so a model asked about a famous trial may simply remember how it turned out.
Diviner AI's answer is contamination-safe validation: every historical forecast is routed to a model whose knowledge provably predates the trial's real outcome. The model cannot recall an answer that did not exist when its training ended — so when its forecasts are scored against what actually happened, the score is earned, not leaked.