Research
Dated histories become models and decision tools
We reconstruct a world from its records, learn how its state changes, and turn the model into tools for forecasts, triage and conditional scenarios. Agents choose actions; later outcomes tell us how those decisions worked.
Explore an event stream · Follow the model pipeline
- Software: the September scaling study covers 28.40 million events from 20 repositories.
- Clinical registry: AACT 2.0 combines trial and sponsor history for outcome forecasts.
- Agents: ranked activity and independent reader panels across recorded corpora.
- Companies: source reconstruction, authored synthetic companies and playable agent evaluation.
Results and tools
Learn the history, then help choose the next step
Models trained on dated records forecast later outcomes, rank activity for review and let agents compare conditional scenarios.
| World | Tested result or working capability | What you can use |
|---|---|---|
| Software | July model: OpenClaw 28-day merge AUC 0.859; time-to-merge rank correlation 0.711, without retraining | Merge and delay forecasts, unusual-activity lists and proposed-split comparisons |
| Trial registry | Current six-month shortfall AUROC 0.819; termination 0.832 on development anchors | Trial and sponsor forecasts, attention lists and conditional scenarios |
| Agents | 86% independent-reader agreement on a selected 100-run Terminal-Bench panel from 7,832 held-out episodes | CLI scoring, candidate nomination and retrospective triage |
| Companies | Operator-assisted source reconstruction and synthetic-company authoring | Playable action consequences and agent evaluation under explicit authored assumptions |
The linked decks and World Model Atlas carry the comparisons, model vintages and detailed evaluations. Project Confirm’s FDA oncology results are a separate earlier study.
The record
Preserve what was known when
Each event carries its time, source and the object it changed. Registry archives are compared in publication order; company reconstructions keep source-claimed times and capture times distinct.
- 01
Recover source-linked histories of decisions and outcomes.
- 02
Check source-era changes and drift to keep the record consistent.
- 03
Train on earlier history and grade against later outcomes with matched-information controls.
A supplied scenario is a conditional sensitivity. An authored company mechanism is a synthetic assumption. The tools let people inspect both alongside the evidence.
Research decks
The worlds and their tests
Each deck explains the record, the usable tools and the tests behind the results.
- 01
- 02
- 03
- 04
Play poker against an agent · Explore a synthetic company · All worlds · World Model Atlas · review access