Corrections Log
Every material correction to the model, the data, or a published forecast since launch. The point of this page is radical transparency: any forecast you read today links to the version of the model that produced it, and any prior call can be reproduced.
Active manual overrides (recal_overrides)
2026-09-16 | Live
Four 2026 races carry a manual override on the model's automated recalibration, applied race-by-race rather than by retraining. Each is reviewed individually and only overrides a single race's output, never the model itself.
- governor-2026-IA: Structural anchor 38.1% D and pre-recal 25.5% D are both R-favored (same side of 50) despite 32.42pp of tension between them, so the sign-flip test in selectiveRecal() never fires and full Platt sharpening was applied, serving 5.7% D. Matches the NM-Gov 2026 pattern exactly (see db/migrations/052). Confirmed via scripts/audit_ia_gov.js, 2026-08-25.
- governor-2026-ME: single-dim outlier, Risk
- governor-2026-NM: Recal sharpened 45.3% structural anchor (near-toss-up) + 31.8% pre-recal (mood/thermostat mix) to 87.1% R. selectiveRecal's sign-flip flag did not fire because both anchor and pre-recal sit on the same (R) side of 50%, so no disagreement was detected despite 13.5pp of tension. Two model-wide fixes attempted and rejected: (1) naive missing-dim/forward-filled gate regressed LOCO Brier +0.023 (2014 uniform thinness false-positives); (2) cycle-relative version of that gate passed the Brier/affected-count bar but does not flag NM-Gov itself (10/10 dims_present = cycle median exactly, forward-filled ratio 0.2 < 0.5 threshold). Held un-sharpened here specifically pending a better structural fix.
- senate-2026-ME: single-dim outlier, Risk
CivicTrust registration_rate inverse-signal bug (v11c -> v11d)
2026-06-01 | Methodology
The Census CPS Voting Supplement question PES2 (voter registration status) is only asked of self-reported non-voters. Our earlier ingestor computed regPct = pes2_yes / (pes2_yes + pes2_no), which inverted the signal: states with high turnout produced low registration rates. The corrected formula counts voters as presumed-registered: regPct = (voters + non-voter_yes) / total. Combined with switching turnout source from CPS turnout_delta to MIT EL absolute senate turnout, and replacing the binary state_lean with a continuous presidential margin, v11d shipped at test_brier 0.135. The prior v11c (0.114) was a broken-data ceiling and has been retired from the publishable selector along with all pre-fix variants.
Information_candidates feature clip (v9 -> v10)
2026-05-31 | Methodology
The Information dimension's candidate-quality sub-feature was missing a sign-flip on incumbent contests, producing P(D) drift on uncontested seats. v10 added a clip and re-calibrated. Brier moved 0.137 to 0.135 on the 2024 holdout.