Prediction performance and evidence

Created by Franz Bettag, applied AI engineering and AI security, Nürnberg/Fürth.

Generated at 2026-09-05T19:39:17.899738924+00:00. Methodology 1.0. Download the complete JSON ledger.

What these numbers measure

Retained public evidence only. Confirmation timestamps are source publication times, not verified technical reset times. Missing resets outside this monitored evidence universe cannot be measured. Model-release rumors are inferences, not reset predictions or confirmations.

Advance notice starts when this service first stored the incoming signal, never at an earlier tweet timestamp. A match requires the same stored event, audience and scope. A confirmed announcement is not independent proof that every account received a reset. Late discoveries do not count as predictions.

Only explicit retractions count as confirmed false positives. Expiry without corroboration remains unconfirmed. Observed resets without prior notice are misses within the retained evidence, not an estimate of all resets. Historical forecast windows were mutable and are therefore not backfilled. There is no defensible overall accuracy percentage yet.

Confirmed with advance notice

4

Observed resets without advance notice (observed misses)

8

Retracted false positives

0

Expired, unconfirmed signals

7

Pending signals

0

Model rumors, not scored as reset predictions

1

Median advance notice to confirmation publication

32857 seconds

Timestamped ledger

MAYBE/model rumors are inferred. YES is qualifying first-party reset evidence; only an effective-phase source confirms the outcome. Rows preserve this distinction.

Empty ledger means insufficient data, not zero errors or perfect predictions.

Engineering case study