Archive correction — 2026-10-08. This May 2026 finding describes an early track-record implementation and its historical diagnostic output. The figures below have not passed the current issue-time, model-version, target, and complete-window qualification gates. They are not verified production forecasting skill or current performance.
The original endpoint joined rows from sentinel_forecasts to sentinel_outcomes and reported 720 rows, precision 0.69, recall 0.39, Brier score 0.556, and a 56.4 percentage-point difference between average forecast probability and outcome rate. Those are preserved here as historical computed values, not validated predictions or mature outcomes. The join has not shown immutable issued probability and model-version receipts, a matching target definition, or complete observation of each seven-day window.
The former article called those numbers a live 30-day production track record and said Adaptive Conformal Inference (ACI) was correcting the difference. That interpretation is withdrawn. The ACI sample and interval target also lack the lineage needed for a 90% coverage claim. A threshold of 0.05 or a nightly refresh does not establish predictive skill.
Current status
The public empirical scorecard is on hold while its forecast and outcome lineage is rebuilt. The API may return an unvalidated status rather than numeric precision, recall, calibration, or interval coverage. Other model rows with training-time metrics must not be read as live empirical performance.
What would restore a scorecard
Keep immutable issuance receipts with probability, model version, target, and issue time; join them to independently sourced outcomes only after the full horizon matures; account for missing coverage and base rates; then score a held-out, forward-time set with the exact served prediction path. Publish the denominator and exclusions with any future metric.