Predictive alerts paired with a ledger that checks every lead-time promise against what actually happened, after the fact.
Closed source, inquirePrecausal combines three mechanisms: kinematic anticipation, a constant-velocity rollout that raises zone-entry, convergence, and crowd-buildup alerts ahead of time; ablation-replay counterfactuals, a real forked re-run of the predictor with a candidate cause surgically removed, not narration; and a falsification ledger that resolves every prediction to fulfilled or falsified against what actually happened.
The falsification ledger is the honesty mechanism this product is built around: it does not just log predictions, it grades them after the fact, and, as the evidence below shows, that grading can surface real weaknesses rather than only confirm the product's own story.
Predicted a zone-entry alert 6.25 seconds ahead of an intruder reaching the zone. Both ablations (presence removal, motion freeze) correctly killed the alert on replay. The ledger later resolved the same prediction as fulfilled, with 6.5 seconds actual lead time, and ledger precision 1.0 on this single scenario.
| Alert kind | n | Precision | Recall |
|---|---|---|---|
| zone_entry | 45 | 0.667 | 1.000 |
| convergence | 35 | 0.657 | 1.000 |
| convergence_multi | 10 | 0.500 | 1.000 |
| crowd_buildup | 45 | 0.356 | 1.000 |
Overall false-positive rate across 75 true-negative scenarios: 6 out of 75, or 0.080. The project's own results file states plainly that crowd_buildup is the weakest kind: of its 29 falsified predictions, 11 genuinely occurred but later than the predicted window, and 18 never happened at all, because the underlying trend reversed, a real structural weakness of trend-slope forecasting, not a scenario-generation artifact.
63 fulfilled, 87 falsified, 50 no-alert. Falsified predictions outnumbering fulfilled ones here is consistent with the newer battery's finding that aggregate precision is well under 1.0 by alert kind. Two independent evaluation passes agree the core prediction claim is currently weaker than the single scripted demo suggests, and both are disclosed rather than hidden.
An outsider can download the published benchmark dataset and try the demo Space to see the kinematic-anticipation and ledger behavior. The 210-scenario battery and its per-kind precision numbers are reported from the project's own results file and are not independently re-runnable from outside, since the source code is not public.