Estimate scorecard

Notes on the record

• G-100-25324-425560: The product's first real alert subscriber's case, recorded as /perm-case-status showed it on August 28, 2026. The next day the estimator gained a measured employer-initial shift, worth +2 days for this employer; the prediction of record is the earlier one and the shift does not move it outside its window.
• G-300-25324-425356: The first predictions from the decision-pace model, which became the lead model on 13 September 2026: cases ahead from the live census divided by DOL's measured 625 decisions a calendar day, band from that rate's own p10/p90. Recorded across four filing months on purpose - a model that is only ever scored near the frontier is never tested at the horizon where it can be most wrong. Recorded from the LIVE PAGE, not from a harness: an earlier draft of these entries came from calling the calculator directly and was 3 to 5 days out, because the page also applies the measured stage adjustment and the employer-initial shift. This interface says the anchor is the one the page printed, so it has to be.
• G-200-26015-564165: The first predictions from the decision-pace model, which became the lead model on 13 September 2026: cases ahead from the live census divided by DOL's measured 625 decisions a calendar day, band from that rate's own p10/p90. Recorded across four filing months on purpose - a model that is only ever scored near the frontier is never tested at the horizon where it can be most wrong. Recorded from the LIVE PAGE, not from a harness: an earlier draft of these entries came from calling the calculator directly and was 3 to 5 days out, because the page also applies the measured stage adjustment and the employer-initial shift. This interface says the anchor is the one the page printed, so it has to be.
• G-200-26075-707139: The first predictions from the decision-pace model, which became the lead model on 13 September 2026: cases ahead from the live census divided by DOL's measured 625 decisions a calendar day, band from that rate's own p10/p90. Recorded across four filing months on purpose - a model that is only ever scored near the frontier is never tested at the horizon where it can be most wrong. Recorded from the LIVE PAGE, not from a harness: an earlier draft of these entries came from calling the calculator directly and was 3 to 5 days out, because the page also applies the measured stage adjustment and the employer-initial shift. This interface says the anchor is the one the page printed, so it has to be.
• G-300-26166-017385: The first predictions from the decision-pace model, which became the lead model on 13 September 2026: cases ahead from the live census divided by DOL's measured 625 decisions a calendar day, band from that rate's own p10/p90. Recorded across four filing months on purpose - a model that is only ever scored near the frontier is never tested at the horizon where it can be most wrong. Recorded from the LIVE PAGE, not from a harness: an earlier draft of these entries came from calling the calculator directly and was 3 to 5 days out, because the page also applies the measured stage adjustment and the employer-initial shift. This interface says the anchor is the one the page printed, so it has to be.

How this is scored

An estimate is recorded with the anchor and the window exactly as the case page printed them on the day. The error is the signed number of days between the anchor and the day DOL's index first showed a final status, read from the sweep that finished September 16, 2026. A decision inside the printed window counts as inside; one outside counts against the estimate whatever the anchor said.

Entries are never edited. When the model changes after a prediction is recorded, the change is noted beside the entry and the original prediction is still the one scored, because a ledger that rewrites its own predictions scores nothing. The estimator's backtested error over decided cases, about 50 days at the median, is on the accuracy guide; this page is the forward test, one real case at a time.

Case numbers are public federal records. Nothing here identifies the person who watched the case.