These six headline figures were published, then withdrawn when their measurements failed. Each entry names the measurement, the defect, the date the withdrawal happened where the record states one, the editorial date of the record itself, and the replacement only when a dated calibration measure exists.
The source for this record is JOB_EVIDENCE_PACKET.md, published with this finding on 2026-07-23. That is an editorial date, not the date a withdrawal happened; where the packet does not state when a figure was withdrawn, this page says so rather than reusing the publication date. A withdrawn figure is historical documentation, never a current result.
withdrawn, no replacement
withdrawn: +18.38% pregame return figure computed on a market-following baseline
Measurement failure
The evaluation CSV had no prediction column. The grader selected direction from the devigged close, used a fixed conversion unavailable in the recorded source, and tuned filters in sample. That is a market-follow artifact, not a model measurement.
Withdrawal date
date not recorded
Editorial date of the record
2026-07-23
Replacement
No replacement measurement is published.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 4, Do-not-claim list: pregame return row
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the pregame return calculation and its market-following defect.
Related reading
Claim ledger (context only; snapshot predates these withdrawals)
superseded
withdrawn: 0.119 end-of-third-quarter Brier score
Measurement failure
Two fourth-quarter-derived features entered a model that was predicting the fourth quarter, and the cited file reported a different figure. The original score therefore contained future information.
Withdrawal date
date not recorded
Editorial date of the record
2026-07-23
Replacement
Replacement measurement: leak-free walk-forward end-of-third-quarter Brier score of approximately 0.141 (unitless), stated as an approximation in JOB_EVIDENCE_PACKET.md and published there on 2026-07-23.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 3, In-game win-probability Brier; Section 4, end-of-third-quarter row
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the leak-inflated end-of-third-quarter Brier score and its leak-free replacement.
Related reading
State-conditioned calibration (MLB and international soccer only)
withdrawn, no replacement
withdrawn: +54.57% / 78.11% in-play accuracy figure measured against a lagged L5 proxy ceiling
Measurement failure
The score used an L5 line proxy rather than a real closing reference. It described a soft proxy ceiling, not an externally validated measurement.
Withdrawal date
date not recorded
Editorial date of the record
2026-07-23
Replacement
No replacement measurement is published.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 4, Do-not-claim list: in-play proxy row
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the in-play proxy score, which used an L5 reference instead of a closing reference.
Related reading
Claim ledger (context only; snapshot predates these withdrawals)
withdrawn, no replacement
withdrawn: +8.94pp closing-line movement calculation
Measurement failure
The aggregate was circular: it used the same model-unused, devigged-direction corpus to define and grade the movement calculation. The denominator did not provide an independent comparison.
Withdrawal date
date not recorded
Editorial date of the record
2026-07-23
Replacement
No replacement measurement is published.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 4, Do-not-claim list: aggregate CLV row
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the aggregate closing-line movement calculation and its circular comparison corpus.
Related reading
Claim ledger (context only; snapshot predates these withdrawals)
withdrawn, no replacement
withdrawn: steals and blocks training R^2 about 0.79, with leak-free holdout R^2 about 0.06
Measurement failure
A leaky grid search inflated the training measurement before the holdout exposed the collapse. Corrective regularization now takes precedence over the stale tuned parameters.
Withdrawal date
date not recorded
Editorial date of the record
2026-07-23
Replacement
No replacement measurement is published.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 3, The self-caught leaks are the strength
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the steals and blocks training-versus-holdout overfit measurement.
Related reading
Claim ledger (context only; snapshot predates these withdrawals)
withdrawn, no replacement
withdrawn: assists conclusion after postseason stress testing
Measurement failure
The measurement was regime-dependent: it failed in the playoffs, and an in-series play-by-play replay confirmed that the earlier conclusion did not hold there.
Withdrawal date
2026-07-21
Editorial date of the record
2026-07-23
Replacement
No replacement measurement is published.
Evidence citation
No published evidence document. docs/JOB_EVIDENCE_PACKET.md is the cited withdrawal record, but it is not published on this site.
Section: Section 3, lines 193-197: assists withdrawal record
Editorial date: 2026-07-23 | Sport: NBA
Measures: Withdrawal record for the assists conclusion after its playoff stress test, dated 2026-07-21 in the packet text.
Related reading
Claim ledger (context only; snapshot predates these withdrawals)
The record keeps measurement failures visible. A replacement belongs here only when the published source names the calibration measure, its unit, and its date.