House leans Democratic.
Senate leans Democratic.

Among 282 races forecast at 20–80% probability, winner accuracy was 60.6%, Brier 0.233 and mean winner confidence 64.2%. Safe races make overall accuracy substantially higher.

On 619 matched races, the selected policy minus a plain weighted-poll baseline had Brier difference 0.0000 (negative is better). A whole-election-cycle bootstrap over 10 clusters gave an exploratory 95% interval of -0.0069 to 0.0063. This does not establish that the more complex policy is superior.

Does winning an independent seat imply caucus control?
No. MT is a multiway comparison, NE affiliation is unresolved; publish conditional caucus stress endpoints, not invented independent win/caucus probabilities.

Does a reused district number preserve its historical electorate?
No. Verified TX-09 boundary change retires the old district vote prior. Other remapped districts still require comparable reconstructed baselines.

Does overall accuracy hide competitive-race difficulty?
Report close-race and safe-race scores separately. Outcome-defined close subsets are retrospective diagnostics, not a deployable selection rule.

Are hundreds of races hundreds of independent error realizations?
No. Compare matched races using whole-cycle bootstrap; inspect the interval before claiming improvement.

Can fieldwork end be treated as public availability?
Only as a documented limitation where publication is unknown. Enforce verified release dates, retain first-observed source versions and publish a three-day delay sensitivity.

Can a generic-ballot archive be appended without source replacement?
No. Replace office/cycle coverage cells to prevent double weighting; keep older uncovered history.

Do registered Democrats voting early identify net Democratic votes?
No. Registration, vote choice, mode substitution and reporting lag differ. Early-vote data remains visible with zero central probability weight pending validated comparable history.

Are pollster labels independent organizations or samples?
No. An eight-way ablation separates alias pooling of frequency caps, counts and prior-cycle bias. The earlier-cycle-only outer policy does not establish an improvement, so production defaults remain unchanged. Publisher, sponsor, fieldwork and panel identities need separate provenance.

On the same 619 historical races, pooling aliases improved some retrospective comparisons. A policy choosing configurations using earlier cycles only had Brier 0.12123, versus 0.12156 for reported labels. Its exploratory cycle-bootstrap improvement interval crosses zero: -0.000969 to 0.000043. Production defaults remain unchanged.

Publisher, sponsor, fieldwork organization and sample panel are different roles. Shared organization keys do not establish statistical independence.

Senate poll-covered subsets · 6 held-out cycles · ranked count loss 1.093 · observed 90% interval coverage 100.0%. The fixed within-office correlation remains 0.30; this small sample does not establish full-chamber calibration.

Governor poll-covered subsets · 6 held-out cycles · ranked count loss 1.043 · observed 90% interval coverage 100.0%. The fixed within-office correlation remains 0.30; this small sample does not establish full-chamber calibration.

Historical poll-covered Senate/governor subsets with at least five contests; not full chambers, not caucus-aligned control or validation of unpolled priors. Few cycles and changing poll coverage limit generalization. No production rho change.

The production-style national seats–votes curve has 2 fully held-out cycles · Brier 0.042 · seat MAE 9.4. This sample is too small to establish robust control-probability calibration.

Only a few independent chamber outcomes are available, especially with the production post-2012 curve. These results do not validate historical district rating priors or certify the live market context layer.

Decay settings were compared on the same historical races. An outer selection rule could change settings using earlier cycles only, and required at least 0.002 Brier improvement before switching. It retained 30 days.

21-day half-life · 774 races · Brier 0.1106 · winner accuracy 82.2%

30-day half-life · 774 races · Brier 0.1099 · winner accuracy 82.6%

45-day half-life · 774 races · Brier 0.1092 · winner accuracy 83.2%

These are research comparisons, not measured 2026 accuracy.

Fieldwork end dates are not necessarily publication dates. On the same 749 races, adding a conservative three-day availability delay changed Brier from 0.1121 to 0.1133. Missing or revised release timestamps remain a validation limit.

A delay assumption is a sensitivity analysis, not recovered publication history. Backfilled/revised data can have longer delays. Live forecasts remain unchanged; actual as-published timestamps are the preferred future input.