Midterm 2026 poll-weighting method and recall-vote test

One polling reading per race, with poll count and fieldwork dates. The reading is a documented weighted-average heuristic where eligible, otherwise the latest primary-verified published poll, explicitly labeled as a single poll. The research also asks whether polls using recalled 2024 presidential vote read differently from other polls. Optional debate, ad and coverage descriptions are separate from the numerical polling estimate. Every number in this note is computed by analysis/weight.py and read from the sealed release bundle; this note formats, it never recalculates. The tracker web app renders the same bundle.

What this is not: a forecast, a pollster rating, or an identification of any causal effect. Section 2 says what the recall comparison can and cannot show, and that framing sits beside every number that depends on it.

Two clocks

• analytical_as_of: one timestamp per release, America/Los_Angeles with a numeric offset. Every eligibility test, recency age and freeze uses it; it is never read from the live clock after being set. The current release's value is in its manifest.json and in the tracker footer.
• Operational time: every retrieved_at, scored_at, recorded_at, checkpoint and budget record carries the actual time of the action.

1. Weighting spec (heuristic, documented, not an inverse variance)

weight = f_track × f_recency × f_sample × f_pop × f_sponsor × f_disclosure

Exclusion cutoff, not a floor. A weight below 0.05 removes the weighted_average flag with a reason; retained and excluded counts, retained weight per race and the no-cutoff sensitivity are all reported. The no-cutoff sensitivity uses the retained stratum, or the same deterministic stratum rule on otherwise eligible positive-weight polls when no poll clears the cutoff. It is reported separately and never substituted for the headline. Weighted mean = Σ(w × margin) / Σ(w) over weighted_average-flagged estimates in the race's averaging stratum; NA if none. The weighted margin itself stays "not yet computed" until Stage 2c passes; before that the tracker shows the unweighted stratum mean labeled preliminary.

Scenario lines. A uniform 2.5-point miss toward Republicans is the control-flipping scenario; a 5-point miss is the 2016 and 2020 scale scenario. The 5.0 in the Stage 3 probability mapping (Section 8) is a standard deviation, not a flip threshold, and is labeled separately wherever it appears.

Deferred by design: house effects (leave-one-firm-out against time-matched peers within 7 days, at least two independent peer firms and three independent waves; never applied to the recall comparison); herding (no RealClearPolling average reconstructed); correlated error (a disabled sketch; probabilities only in Stage 8/Stage 3).

Imported track-record source

Silver Bulletin source, January 2026 update, 2026-01-14; file Pollster_Stats_Full_2026.xlsx, downloaded 2026-10-25T21:17:55.561941+00:00, SHA-256 14f5aeba96e21c3535fca225869595557347d2a1005dbe0303fc082cf6261d61. Metric Simple Average Error, percentage points; cohort Silver Bulletin database, all race types and years in the file. Mean-Reverted Bias is descriptive only.

Sparse histories are unstable; a one-poll history is not reliable evidence of persistent quality. Unmatched or incompatible identities retain PROVISIONAL 1.0, which is not an endorsement. Partner ratings are not borrowed. Publisher ban flags and identity reasons remain in each method record's notes; the threshold rule itself does not substitute for the integrity quarantine. No historical bias correction is applied. The project's threshold factors have not been shown to outperform a simple average.

2. The recall-vote test: hypothesis and limits

Hypothesis under test. Some midterm surveys may model an electorate that differs from the voters who turn out. Recalled-2024-vote weighting is one possible contributor. A shift toward Republicans is a hypothesis, not an established finding, a preset correction, or a reason to automatically prefer NO polls. Direction depends on targets, nonvoters, recall error and the likely-voter screen. The same numeric eligibility and weighting rules apply to YES, NO and UNVERIFIED surveys.

Validation still required. The current estimator is a weighted-average heuristic. It has not been shown to beat a simple average on held-out elections. A future evaluation must freeze rules before scoring and compare against that baseline. January 2026 Silver Bulletin ratings include 2025 outcomes, so using those ratings to claim an independent 2025 accuracy test would leak outcomes.

Separate research. Campaign spending may be a diagnostic cross-check, but campaigns also react to polling. Spending remains outside the numeric estimate unless held-out testing establishes incremental value. Rhetoric scores are optional descriptive research; no accuracy improvement has been established and they never enter the polling margin.

What the comparisons are. Descriptive differences between groups of estimates classified by whether the applicable methodology states recalled 2024 vote was used in weighting or calibration. None of them identifies the effect or its mechanism. Mode, timing, recruitment, screens, targets, nonresponse, pollster effects and selective publication remain confounded. Any association with mode or alignment is measured in these data before it is mentioned. This sentence sits next to every contrast in Section 5 and in the tracker.

Classification (Section 4 contract). YES: the applicable methodology states the survey used recalled or self-reported 2024 presidential vote in weighting or calibration. NO: explicit non-use, or a complete applicable weighting-variable list that omits it. UNVERIFIED: everything else (generic "past vote", "when appropriate", unspecified year, a general page without demonstrated applicability, a questionnaire that merely asks past vote). Firms with mixed classifications across method records are MIXED in summaries. UNVERIFIED estimates never enter YES-NO contrasts and are counted beside them.

3. Scope and candidate identities

Stage 1 Senate (race_id): 2026-TX-SEN Talarico/Paxton, 2026-IA-SEN Turek/Hinson, 2026-ME-SEN Jackson/Collins, 2026-AK-SEN Peltola/Sullivan, 2026-OH-SEN-SPECIAL Brown/Husted, 2026-MI-SEN El-Sayed/Rogers, 2026-NC-SEN Cooper/Whatley, 2026-GA-SEN Ossoff/Collins. Deferred Senate: KS, NE, FL-SPECIAL, SC, MS, NH. Stage 1b governors (2026-ST-GOV): OH Acton/Ramaswamy, IA Sand/Lahn, MI Benson/James, GA Bottoms/Jackson, WI Crowley/Tiffany. Deferred: AZ, NV, FL, KS. Candidate identities are confirmed against each release; a mismatched matchup is quarantined (exclusion_reason MATCHUP_MISMATCH), never relabeled. Ranked-choice races: ME and AK only; every other race is round=NA.

4. Data contracts

Shares and margins. Shares in percentage points on the published denominator; margin = D-slot share minus R-slot share, never renormalized; undecided separate from other. denominator_class: FULL (all named candidates plus undecided on the base), TWO_WAY (forced choice), LEANERS (undecided leaners allocated). pop: LV, RV, A. Stratum = (race, round, denominator_class, pop).

Rounds. FIRST and FINAL estimates are stored and kept; rounds are never combined. Round precedence, RCV races only, applied before any other ranking for a given purpose: FINAL if any estimate eligible for that purpose is FINAL; otherwise FIRST, flagged rcv_first_round_fallback=1; a race with neither is NA for that purpose.

Source verification. Poll schema 1.2 records primary_verified as YES or UNVERIFIED. Rows whose dates, population or toplines have not been confirmed on the primary release remain in the inventory but enter no average, sensitivity or recall contrast. A two-candidate question without a documented RCV tabulation does not establish FINAL; its round stays UNVERIFIED and it is also excluded from calculations. The earlier Alaska convention treating those questions as FINAL was withdrawn on 2026-10-25; the original data are preserved in timestamped snapshots.

Averaging stratum (per race, for weighted averages and Stage 3): after round precedence, the stratum with the most weighted_average-flagged estimates; ties broken by denominator_class FULL, LEANERS, TWO_WAY, then pop LV, RV, A. It does not depend on recall classes: a race with only YES, only NO or only UNVERIFIED estimates still gets an average.

Contrast stratum (per race, for contrasts (ii) and (iii)): after round precedence over recall_contrast-flagged estimates, the stratum with the most such estimates that contains at least one YES and one NO; same tie-break; null with a reason if no stratum has both, and the race is then absent from (ii) and (iii).

Eligibility windows. HOUSE_ONLY: field_end 2026-09-01 to 2026-09-30 (stored, not averaged or contrasted). BOTH: field_end 2026-10-01 through analytical_as_of. Required ordering: field_start ≤ field_end ≤ release_date ≤ analytical_as_of. A release after analytical_as_of stays in the file for the next release.

Identity. pollster_id, method_id, sample_id, poll_id, estimate_id, tracking_series_id are deterministic functions of the data and stable across releases. Partnerships keep a joint identity separate from the sponsor and the member firms. Every reported variant is a row; the principal estimate per (poll, race, round) is LV over RV over A, FULL over LEANERS over TWO_WAY, unprompted over leaner-allocated, and only the principal carries selected=1.

Trackers. A tracking series is one pollster publishing rolling or repeated waves of the same instrument for the same race, sponsor and population. Deduplication applies only within one tracking_series_id: keep the latest wave, walk backward keeping only waves whose field periods do not overlap a kept wave; dropped waves get selected=0, exclusion_reason TRACKER_OVERLAP. Independent polls, including two pollsters with identical field dates, are never deduplicated against each other.

Integrity quarantine. Before any extraction: every Wedgewood Polls release (admission 2026-10-20 after the FiftyPlusOne investigation that most releases had no real data) and Median Strategies (August 2026 admission), plus any other documented fabrication or withdrawal found. Primary-domain hosting never overrides an authenticity failure. Quarantined estimates stay in polls.csv with selected=0 and exclusion_reason QUARANTINE, enter no calculation, and each is covered by an exclusions.csv record.

Eligibility flags (owned by weight.py, per estimate, with reasons): principal (selected and not quarantined); recall_contrast (principal, in-scope race, dates, pop, D and R, denominator, round, recall_weighted in {YES, NO}, window BOTH; missing n, unknown sponsor, undisclosed targets and the weight cutoff do not remove it); sponsor_restricted (recall_contrast and sponsor_class NONPARTISAN or BIPARTISAN); weighted_average (principal, BOTH, valid n, sponsor_class known, every factor input present, weight ≥ 0.05; independent of recall_weighted).

Inventory, not census. polls.csv is an inventory from the union of RealClearPolling, 270toWin and primary releases, with field dates, n, population and toplines confirmed on the primary release where it was reachable (rows that rest on an aggregator alone say so in notes). It is never described as every public poll. The per-race coverage ledger is in Section 11.

Method evidence. Classification per pollster per cycle when the 2026 methodology statement is boilerplate across releases (one October release verified; applicability=CYCLE_BOILERPLATE); per survey where statements differ or where a release is a race's sole eligible poll (applicability=SURVEY). Primary evidence is the conducting pollster's page, the commissioning partner's page, or their linked documents, with owner, date, locator and retrieval time recorded. Stop rule: three distinct attempts for the specific method version, then UNVERIFIED; a later or general version is never substituted.

5. Recall-vote contrasts (descriptive; see Section 2)

Methods registry

Filled from derived.json › section5_contrasts at each release. Four contrasts on recall_contrast-flagged estimates with unadjusted margins:

• (i) pooled unweighted YES minus NO margin difference across all strata: a mixed-population, composition-sensitive description;
• (ii) race-balanced within each race's contrast stratum: average within pollster and group, then equally across pollsters within group, YES minus NO, then equally across races that have a contrast stratum;
• (iii) as (ii) on sponsor_restricted estimates;
• (iv) as (i) restricted to LV and the modal denominator_class: the matched-population sensitivity.

Each contrast reports the UNVERIFIED count beside it, the races included and the independent waves. Per-race values for TX, IA, ME, AK, OH-SPECIAL and MI appear where both groups exist in that race's contrast stratum, else NA with the reason.

Every comparison in this section is descriptive. None identifies the effect of recalled-vote weighting or its mechanism. Mode, timing, recruitment, screens, targets, nonresponse, pollster effects and selective publication remain confounded.

5.1 Results (release 20261025T121938-0700-stage2-df9d158-6ec76c0a50)

UNVERIFIED principal estimates in the BOTH window, counted and excluded from every contrast: 45.

Descriptive only. This number does not identify the effect of recalled-vote weighting or its mechanism; mode, timing, recruitment, screens, targets, nonresponse, pollster effects and selective publication remain confounded.

Per race (TX, IA, ME, AK, OH-SPECIAL, MI), within the contrast stratum:

Descriptive only. This number does not identify the effect of recalled-vote weighting or its mechanism; mode, timing, recruitment, screens, targets, nonresponse, pollster effects and selective publication remain confounded.

6. Governor-Senate gap (illustration, not benchmark)

For IA, OH, MI and GA: only pairs sharing an explicit sample_id, pop, denominator_class and round; gap = governor margin minus Senate margin. 2022 comparators from certified November general results (GA Kemp vs Walker, OH DeWine vs Vance, NH Sununu vs Hassan, NV Lombardo vs Cortez Masto), cited in Section 11.

2022 November general-election comparators. Illustration, not a benchmark. Full valid-vote denominators, including certified write-ins and Nevada None of These Candidates; D minus R, without two-party renormalization.

Georgia uses an unprocessed archive of the official state export, plus the state's qualified write-in certificate. Nevada uses an archived copy of its official Supreme Court canvass. Ohio and New Hampshire use current official PDF returns. Source locators, vote counts and provenance are frozen in derived.json.

7. Weighted margins (Stage 2c; heuristic, track-record stand-in)

Not yet computed in Stage 1 and 1b releases; the preliminary unweighted stratum mean is shown instead and labeled. After Stage 2c passes, each race's weighted margin, retained and excluded counts, retained weight and the no-cutoff sensitivity are listed here, labeled heuristic and stand-in.

Weighted margins are heuristic, track-record stand-in values (Section 1) once Stage 2c passes; until then the preliminary column is an unweighted mean of the averaging stratum.

Public polling-read coverage

An unavailable average is not evidence that no published poll exists. The public card uses the qualifying weighted average when available. Otherwise, weight.py selects one principal, non-quarantined, primary-verified estimate in the BOTH window, with supported population, denominator and round, a finite published margin, primary source link and field/release dates no later than the cutoff. RCV FINAL-before-FIRST precedence applies first. Within the chosen round, prefer newest field end, then latest release, then denominator FULL / LEANERS / TWO_WAY, population LV / RV / A, and stable estimate ID. This deterministic fallback does not average or rescale shares. It never enters averages, recall contrasts, scenarios or Stage 3 probabilities. The 0.05 cutoff and all eligibility rules remain unchanged.

The card says “Latest published poll,” displays its population, field end, age as of the snapshot and primary source, and says “Single poll · no qualifying average.” Missing sample size remains unknown. Adult respondents are not relabeled as voters. No eligible verified source means no public read.

The October 25 coverage correction adds three source readings: Alaska Senate, Data for Progress final round, fieldwork ending October 16, Peltola +2; Michigan governor, Marist likely voters, ending October 18, Benson +6; Wisconsin governor, YouGov adults, ending October 9, Crowley +4. Michigan's distinct likely-voter sample size is unestablished: the published methodology's 1,185 registered voters cannot be substituted for an LV n. Wisconsin's sponsor classification remains UNKNOWN. These facts do not qualify those polls for the weighted average. All ten existing weighted averages and every other analytical section are unchanged from the prior sealed Stage 2 release.

8. Stage 3 contract (executed 2026-11-02 only)

On 2026-11-02: a new analytical_as_of no later than 2026-11-02T23:59:59-08:00; Stages 1 and 1b rerun as a merge; Section 10 rerun on the 2026-11-01 edition; Stage 2c rerun against the same rating_source version recorded in the Stage 2 release; then weight.py --release --freeze writes a release containing predictions.csv (schema 1.0: race, vector, p_d_slot, weighted_margin, stratum_round, stratum_denominator, stratum_pop, n_retained, heat_race_set, frozen_at). Vectors: YES_ONLY and NO_ONLY for races whose contrast stratum has at least one weighted_average-flagged estimate in each group; ALL_ELIGIBLE for every race with a Section 7 margin on its averaging stratum. Mapping fixed now: p = Φ(m / 5.0), Φ the standard normal CDF, m the weighted margin on the relevant stratum (RCV: FINAL only; a race with only FIRST is excluded with reason), clamped to [0.03, 0.97]. 5.0 is a heuristic standard deviation, not an estimate and not a flip threshold.

The freeze is one immutable release: the CSVs, derived.json, predictions.csv and this note's snapshot, with hashes. A later rerun never overwrites it; a correction is a new release carrying supersedes and reason, and the Grading page shows both with the original marked ORIGINAL.

Results: results.csv (schema 1.0: race, resolution_status, winner_slot, certified_margin, denominator_note, result_url, resolved_at; resolution_status in CERTIFIED, PROJECTED, RUNOFF_PENDING, RCV_PENDING, UNRESOLVED). Grading uses CERTIFIED rows only; November 4 is a status check. Brier = mean over graded races of (p − y)², y = 1 if the D-slot candidate won. Signed margin error = frozen weighted_margin minus certified margin on the same denominator and round, NA where the certified pair does not match the frozen stratum. Heat grading per Section 9 on heat_race_set.

9. Candidate heat via the LITMUS Live meter

Selection at analytical_as_of. DEBATE: the latest general-election debate between the two slot candidates with air date ≤ analytical_as_of and captions from C-SPAN or the broadcaster (BROADCASTER), or failing those YouTube auto-captions (AUTO, flagged). AD: the two most recently published ads on each candidate's official campaign YouTube channel with publish date ≤ analytical_as_of, captions OFFICIAL or AUTO. Moderator questions are excluded. Unit = one debate answer or one ad; unit_sha256 over the transcript text. Missing material gets a MISSING_TRANSCRIPT row with the reason; a race is never dropped because one source type is missing.

Scorer, pinned. The LITMUS meter at the commit, prompt SHA-256, model id and configuration recorded in heat.csv and the manifest (the 2026-10-10 era-check configuration with the opinion floor removed; Max's ruling: persuasive is not hyperbolic, score hyperbole only). Changing any of the four starts a new scorer version and a new release. Cache: heat-cache/<key>.json, key = sha256(unit_sha256 | scorer_commit | prompt_sha256 | model | config_sha256); a cached key costs nothing, a changed input pays again, never reuses.

Aggregation. Per candidate and source_type, mean ph_score over SCORED units with n. Debate gap = D-slot debate mean minus R-slot debate mean, defined only when both have at least one SCORED debate unit; ad gap likewise, labeled exploratory (the meter was validated on debate answers and historical speeches, not ads). Hotter candidate = higher mean; a difference under 0.1 is a TIE.

Grading rule, locked. Race set = races with a defined debate gap at the Stage 3 freeze (heat_race_set in predictions.csv). Outcome = the certified loser. Metric = count and share of graded races where the hotter debate candidate lost, ties reported separately and excluded from the share, beside a 0.5 line labeled "no-information reference; not a test of predictive validity". Ad heat graded the same way, separately, labeled exploratory.

Budget. Ceiling $3.00 operational. Each call reserves units × unit_rate × 1.5 in run-state first; if spent plus reserved would exceed $3.00 the remaining units get SKIPPED_BUDGET rows and scoring stops. Failures record their cost. Work left after a cap is listed under Deferred; completion is never implied.

Replay 1 recommendation for the LITMUS go-to-market: the in-scope debate with the widest between-speaker debate gap and caption_source BROADCASTER or OFFICIAL.

Units scored: 63; by status: {'SCORED': 63, 'MISSING_TRANSCRIPT': 45, 'FAILED': 7, 'SKIPPED_BUDGET': 50}; spend $2.80 of the $3.00 ceiling. Races with a defined debate gap: 2026-AK-SEN, 2026-GA-SEN. A difference under 0.1 is a TIE.

Replay 1 recommendation: 2026-AK-SEN (absolute debate gap 0.86; in-scope debate with the widest between-speaker debate gap and caption_source BROADCASTER or OFFICIAL).

The meter scores how something is said, never the position taken; persuasive is not hyperbolic. Ad readings are exploratory: the meter was validated on debate answers and historical speeches, not ads.

10. Coverage temperature via the LITMUS connector

At the Stage 1b gate and again on 2026-11-02: for each in-scope race, find_stories with the candidate surnames; expected edition 2026-10-25 now and 2026-11-01 at the freeze. One coverage.csv row per race and call whatever the result, with edition_status (OK, STALE_FLAG, OLDER_THAN_EXPECTED, ACCESS_FAILED) and match_status (MATCHES, ZERO, PARTIAL, ERROR). Two-snapshot table, descriptive context only; no claim that coverage heat predicts anything.

Descriptive context from the LITMUS paper's edition; no claim that coverage heat predicts anything. Automated reading of the writer's framing, not a verdict on any person or outlet.

2a. Retrospective check: 2025 Virginia and New Jersey governors (Stage 2)

Polls with field_end 2025-10-14 to 2025-11-03 released before polls opened November 4, same selection, integrity and dedup rules; only pollsters already in the 2026 registry are classified, from methodology evidence applicable to that 2025 survey, never backfilled from 2026; others listed, not classified. Certified results from the Virginia Department of Elections and the New Jersey Division of Elections. signed_error = poll margin minus certified margin, negative = miss toward Republicans. Reported: counts and mean signed error by state and group, the within-state YES minus NO contrast averaged equally across states having both groups, and a one-final-poll-per-firm sensitivity. A two-election descriptive check, not causal identification.

Status: PASSED. signed_error = poll margin minus certified margin; negative = miss toward Republicans.

Within-state YES minus NO averaged equally across states with both groups: +4.50 pt (states: 2025-VA-GOV); one-final-poll-per-firm sensitivity +4.50 pt. Two-election descriptive check on a primary-verified subset, not causal identification or held-out validation of the weighting heuristic. the january 2026 ratings incorporate 2025 results and are not used to score a supposedly independent 2025 prediction. Each state needs a verified YES and NO survey. Ambiguous past-vote language, modeled or imputed vote choice, and recall questions without explicit weighting evidence remain UNVERIFIED.

2b. Turnout context, not validation (Stage 2)

North Carolina only: the ncsbe.gov absentee file, accepted ballots, deduplicated per the official layout, by voter_party_code, early in-person and returned mail separated; 2026 shares versus the 2022 final file, labeled non-contemporaneous. Florida deferred. Georgia dropped (no party registration). Compared to poll composition only where a poll publishes party registration for a comparable population.

Party registration, not vote choice. Non-contemporaneous comparison of the 2026 current file against the final 2022 file; not validation or a prediction. Excludes explicitly counted ambiguous identities.

2026: official file, source Last-Modified 2026-10-25T09:22:59Z; latest accepted return date 2026-10-24; 6 repeated accepted rows removed; 2 ambiguous identities excluded from every method total. SHA-256 7d8b5f6bedaa2ad9705f83a5106ab5f2536b7fe6812581dd60c2f1c41420a648.

2022: official file, source Last-Modified 2022-11-19T10:33:42Z; latest accepted return date 2022-11-17; 5 repeated accepted rows removed; 3 ambiguous identities excluded from every method total. SHA-256 c6e85dde21c953d0a757c83a0191865516b020f1fd156e1a1f50c14cf7d2d126.

Retrieved 2026-10-25T14:37:16-07:00. Official layout; identifier definitions. One accepted record per election and statewide NCID; missing NCID falls back to county plus voter_reg_num. Identical accepted repeats count once. Conflicting accepted method or party for the same identity is excluded from every method total. This source ambiguity is not evidence about voter conduct.

Poll comparison: not comparable. p.8 Q26 asks voter registration; n=600 statewide likely voters, not an accepted-early-ballot subgroup. It cannot be compared as though both represent the same population. Harper questionnaire. Original saved PollPrint PDF for #28047: crosstabs label Party Affiliation; comparable registration-by-accepted-vote-method composition not established. Party identification is not substituted.

11. Coverage ledger and sources

Inventory from the union of RealClearPolling, 270toWin and primary releases (Wikipedia and FiftyPlusOne used as finding aids, never as the verification source), with field dates, n, population and toplines confirmed on the primary release where it was reachable. Not every public poll.

Gaps recorded by the research pass (verbatim from the per-race logs)

• 2026-IA-GOV: Cook/NR/GSG public date/sample summary visible; 54/43 subscriber toplines only in RCP/270; no shared sample inferred. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-IA-GOV: Siena governor summary read; NYT exact n606/field dates from finding aids only after direct/PDF/archive blocks. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-IA-GOV: September governor wave archived summary from finding aids; primary current page only October. No cross-ballot pairing established. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-MI-GOV: Governor summary from RCP/270toWin; primary NYT crosstabs blocked after three paths. Do not infer9undecided. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-MI-GOV: Washington Reporter secondary account; original memo absent. Other not inferred. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-MI-GOV: Washington Post/Schar finding-aid toplines; cached PDF is a blocked HTML response, original questionnaire not read. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-MI-GOV: Detroit News governor figures from finding aids; article paywall prevents primary ballot verification. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-MI-GOV: Aggregator-only50/40, MI483LV subset of1430multi-state sample; five other/two undecided shown secondarily but not verified or stored.
• 2026-OH-GOV: Cook subscriber toplines inaccessible; 49/44 from RCP/270toWin. Primary date/sample summary only; other/undecided not inferred. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: BGSU October PDF blocked after direct PDF, browser-UA and archive attempts; RCP/270toWin 50/44. Other/undecided not read on primary. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: Siena governor summary 49/45; NYT exact n/dates blocked (direct, fetch, archive), therefore primary verification incomplete. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: Governor chart numbers inventoried from finding-aid table; primary Senate article does not expose readable governor numeric chart. Primary authenticity of governor values not established. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: Governor chart numbers inventoried from finding-aid table; primary Senate article does not expose readable governor numeric chart. Primary authenticity of governor values not established. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: Governor chart numbers inventoried from finding-aid table; primary Senate article does not expose readable governor numeric chart. Primary authenticity of governor values not established. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-OH-GOV: Governor chart numbers inventoried from finding-aid table; primary Senate article does not expose readable governor numeric chart. Primary authenticity of governor values not established. Toplines or field/sample details not confirmed on primary; aggregator-only and analytically excluded.
• 2026-WI-GOV: Aggregator-only49/44. X original/login page, archive/login and sponsor/firm sites yielded no readable release or methodology. Other/undecided not inferred.
• 2026-AK-SEN: RealClearPolling blocks curl and WebFetch (DataDome 403); inventory taken from the live page in Max's logged-in Chrome (2026-10-25) and the Wayback snapshot of 2026-10-13 for per-row percentages.
• 2026-AK-SEN: FiftyPlusOne has no Alaska Senate page (three slug guesses 404; its senate index lists only Kansas, Michigan, Texas), so it contributed nothing as a finding aid.
• 2026-AK-SEN: NYT/Siena Alaska toplines returned 403 to curl, three Wayback captures and WebFetch; read in Max's Chrome, where the page is dated Oct. 3, 2026 (URL path shows 10/04 after redirect).
• 2026-AK-SEN: Public Policy Polling (Oct 14-15, 578 RV, Peltola 48 / Sullivan 47): no primary release exists; the only source is Andrew Solender's tweet of a 'Dem-aligned' poll circulating among groups. Sponsor, ballot composition (whether Heikes/DJ Sullivan/undecided were offered) and methodology are unknown; round=FIRST is an assumption; sponsor_class recorded UNKNOWN.
• 2026-AK-SEN: Rasmussen's September release (Sept 13-14, 1,188 LV) was overwritten at the same URL by the October release; it was verified from the Wayback Machine capture of 2026-09-25 of the pollster's own page.
• 2026-AK-SEN: Rasmussen October poll (Oct 12-15, 775 LV, Peltola 44 / Sullivan 42) is not listed on RCP, 270toWin or Wikipedia as of 2026-10-25; included from the primary release.
• 2026-AK-SEN: InsiderAdvantage first-choice shares (48.2/42.6) come from the pollster's Top Line Tabs image on X (@InsiderPolling, 2026-10-11); the release page text states only the final round (50/48). Mode not stated for the Alaska release.
• 2026-AK-SEN: Round convention used in this file: FIRST = question naming all four ballot candidates (first preference / single-choice ballot); FINAL = Peltola-vs-Sullivan estimate, whether an RCV simulation final round or a two-candidate question. Two-candidate questions that retain undecided (NYT/Siena, Quantus Q3, co/efficient Q7) are recorded as FINAL with denominator FULL; stage 2 should decide whether a forced two-way and a two-candidate-with-undecided question belong in the same bucket.
• 2026-AK-SEN: Data for Progress recall_weighted left UNVERIFIED because the survey statement says only 'recalled vote' (no year/office); the instrument's recall items are 2024 presidential and 2024 U.S. House, and the Q1 recall topline (Harris 40 / Trump 53 / did not vote 3 / other 4) matches the certified 2024 two-party split, which is consistent with weighting to 2024 presidential recall. A reviewer may upgrade to YES.
• 2026-AK-SEN: Intermediate RCV rounds (DFP Q16/Q17, ASR Round 2 in both waves, AARP Rounds 1-2) have no slot in the round enum and are recorded in notes only; DFP's hypothetical transfer round (50/50) and the Quantus topline-summary 'two-way share 50.3/49.7' are likewise notes, not rows.
• 2026-AK-SEN: ASR first-round rows report no undecided (ranking question; shares sum to 100); ASR September Round 3 is labeled 'hypothetical' by the pollster because Peltola passed 50% in Round 2.
• 2026-AK-SEN: AARP final-round 53/47 excludes the 11% not ranking either finalist (two-way denominator).
• 2026-AK-SEN: NYT/Siena volunteered 'another candidate' shares reported as '<.5' and '<1' are recorded as null.
• 2026-AK-SEN: Quantus: Heikes share in the full-ballot-with-leaners treatment and the undecided share after the two-candidate leaner push are not printed as toplines (null). RCP lists Quantus n=737 (voters-only ballot base) while 270toWin/Wikipedia list 758 (completed interviews); both are correct per the PDF.
• 2026-AK-SEN: RCP labels the AARP poll 'Fabrizio/Anzalone' (Impact Research is the Anzalone firm); recorded under the names the release uses.
• 2026-AK-SEN: Release dates for Trafalgar (2026-09-25, from PDF filename and Wikipedia cite date; 270toWin shows 9/26) and co/efficient (2026-09-17 per 270toWin/Wikipedia; page says only 'Released September 2026') are not stated to the day on the primary.
• 2026-AK-SEN: No Wedgewood Polls or Median Strategies poll of the Alaska Senate race was found on RCP, 270toWin, Wikipedia, or in web search; quarantine_candidates is empty for this race.
• 2026-AK-SEN: AAPOR Transparency Initiative membership was not verified for any pollster (ti_member='unknown' throughout).
• 2026-AK-SEN: Alaska Survey Research posts no October/September 2026 Senate report on its website; the X threads are the primary release.
• 2026-GA-SEN: RealClearPolling blocked (DataDome 403) on three distinct attempts (curl with browser UA on both URL orderings, WebFetch, Wayback/Google cache/api path); RCP inventory not obtained. Wikipedia's aggregation row says RCP's average covered Sept 15 – Oct 17, 2026 as of Oct 20.
• 2026-GA-SEN: FiftyPlusOne's Georgia page is JS-rendered; its public API (race_id=16091) was used as a finding aid only; Quantus Oct and Rasmussen rows were not seen in the API pages fetched (not verified absent).
• 2026-GA-SEN: Wedgewood Polls (Oct 2-3, 600 LV, 55/45): quarantine candidate; release date not stated (public by 2026-10-08 per Newsweek); data read from a Wayback snapshot of the firm's Google Sites page; the live domain now redirects to a Google Drive sign-in.
• 2026-GA-SEN: Median Strategies: no Georgia Senate poll found on 270toWin, Wikipedia, the FiftyPlusOne API or web search; nothing to quarantine.
• 2026-GA-SEN: InsiderAdvantage Sept 22-23: 'Top Line Tabs Posted Separately' not located; undecided taken from the release text only; MoE not stated in the release (Wikipedia lists ±4.0).
• 2026-GA-SEN: InsiderAdvantage Oct 12-13 (800 LV) Georgia poll reported only the governor race (ANF Oct 14; Wikipedia governor table); no Senate topline found, so no row recorded.
• 2026-GA-SEN: Big Data Poll: undecided/other for the initial RV and LV ballots are not in the release text and the MarketSight crosstab app could not be read; the RV with-leaners figure listed by FiftyPlusOne (55.4/44.6) is not stated in the release and was not recorded.
• 2026-GA-SEN: YouGov Sept 15-18: sponsor not identified (PDF carries none; filename prefix 'ttw'); release date taken as 2026-09-23 (first public coverage and FiftyPlusOne entry), the PDF is undated.
• 2026-GA-SEN: Trafalgar: PDFs are undated; release dates 2026-09-13 and 2026-10-10 come from FiftyPlusOne entries/url slugs, Wikipedia citation dates and 270toWin; Trafalgar's Sept news post (thetrafalgargroup.org/news/ga-sen-general-0913/, fetched 200) and Oct post carry no visible publication date.
• 2026-GA-SEN: Quantus Sept with-leaners row: residual undecided/other not stated. Quantus Oct with-leaners row: 'another candidate' share not shown.
• 2026-GA-SEN: Trafalgar Sept includes Allen Buckley (Libertarian) at 1.3%; AJC reported (July 2026) that Libertarians will not appear on Georgia's 2026 ballot, so the ballot tested differs from the certified ballot; recorded as other.
• 2026-GA-SEN: AAPOR Transparency Initiative membership verified only for ActiVote (its own page); all other ti_member values are 'unknown'.
• 2026-GA-SEN: pollster_alignment 'R' for Trafalgar, InsiderAdvantage, Big Data Poll, Quantus and Rasmussen follows Wikipedia's table labels; FiftyPlusOne's API marks none of them partisan.
• 2026-GA-SEN: Republican nominee confirmed as Mike Collins on every release recorded; no mismatched matchups found.
• 2026-IA-SEN: RealClearPolling live pages returned 403 (WebFetch and curl, both URL variants); the RCP table was read from an archive.org copy captured in late October 2026 (RCP average dated 9/29-10/22), so RCP 'listed_on' reflects that snapshot.
• 2026-IA-SEN: Decision Desk HQ polling page returned 429 on three attempts; not used. FiftyPlusOne's Iowa poll list is JS-rendered (paid API); only its average was readable, so FiftyPlusOne could not be used as a finding aid beyond the Wikipedia aggregate row.
• 2026-IA-SEN: NYT/Siena Iowa poll (Sept 22-Oct 1, 606 LV): NYT toplines/methodology blocked (403 x3); only the headline 48-47 was read on Siena's site; undecided, field dates, n and weighting unverified on primary (recall_weighted UNVERIFIED).
• 2026-IA-SEN: Cook Political Report / New River / GS Strategy Group (Oct 9-13, 602 LV): toplines and methodology are members-only and the site is 403; dates/n confirmed on an archived summary, shares taken from Wikipedia/RCP/270toWin (verified_on_primary=false).
• 2026-IA-SEN: InsiderAdvantage Oct 21-22: topline tabs not posted on the firm's site (said to be on X); Laehn/other/undecided and ballot composition unconfirmed (denominator_class FULL assumed; undecided null).
• 2026-IA-SEN: co/efficient Sept 14-16: the live page now carries the October wave; shares rely on the firm's own X post as rendered in a search snippet plus Wikipedia (verified_on_primary=false); dates, n and methodology confirmed on an archived copy.
• 2026-IA-SEN: Marist Oct 8-11 likely-voter n is not published (only MOE +/-6.6); aggregators' '1,100 LV' is the RV n.
• 2026-IA-SEN: YouGov Iowa polls (Sept 3-8 and Oct 6-12) state no sponsor in the PDFs; sponsor and sponsor_class left null/UNKNOWN.
• 2026-IA-SEN: Release dates for co/efficient Oct (10/19), YouGov (9/11, 10/14) and Cygnal/ITR (9/15) are taken from the 270toWin listing date, not stated on the primary; Rasmussen Sept 28-30 release date (10/02) rests on a news snippet saying 'released Friday'.
• 2026-IA-SEN: Quantus Insights: the linked presentation/crosstabs were not fetched; weighting statement ('weighted to Iowa likely-electorate benchmarks') gives no variables (UNVERIFIED). Trafalgar's methodology page lists modes only (UNVERIFIED). Rasmussen's general methodology page lists 'age, race, gender, political party, and other factors' (UNVERIFIED). Fox News' statement hedges with 'generally applied to' (UNVERIFIED despite omitting past vote). Suffolk and Cygnal publish no weighting variables (UNVERIFIED). NPA says only 'key demographics' (UNVERIFIED).
• 2026-IA-SEN: Wedgewood Polls: only Iowa Senate release found was fielded Aug 27-29 (out of window), recorded under quarantine_candidates without a row. Median Strategies: no Iowa poll found.
• 2026-IA-SEN: Global Strategy Group (Turek campaign) Aug 27-31 internal and Abacus Data Aug 26-28 have field_end before 2026-09-01 and were not collected. No Public Policy Polling or other internal polls with field_end on/after 2026-09-01 were found.
• 2026-IA-SEN: AAPOR Transparency Initiative membership was not checked against the AAPOR list this session; ti_member is 'unknown' except for firms whose membership is well established (Marist, Emerson, Suffolk, YouGov, SSRS, Siena).
• 2026-IA-SEN: Inventory is the union of RealClearPolling (archived), 270toWin, Wikipedia's table and the primary releases named above; it is not every public poll.
• 2026-ME-SEN: RealClearPolling blocked (403 x3: curl Chrome UA twice, WebFetch once); inventory is the union of 270toWin, Wikipedia (finding aid) and primary releases only. Describe as an inventory from those sources, not every public poll.
• 2026-ME-SEN: NYT/Siena (Sep 15-22) not verified on primary: nytimes.com 403 on three URLs plus WebFetch failure; 270toWin's link points at the Michigan toplines. Figures from Wikipedia/270toWin/Bangor Daily News; other-candidate share, governor sharing and weighting unknown.
• 2026-ME-SEN: UNH (Sep 17-21) not verified on primary PDF (403 on four attempts); landing-page abstract confirms a statistical tie and release 9-24-2026 only. Figures from Wikipedia/270toWin; weighting unknown.
• 2026-ME-SEN: Washington Post/Schar (Sep 30-Oct 5): WaPo PDF and page unreachable (curl 000 x2, WebFetch 403 x2). LV 50/46, n=800 and dates confirmed on the Schar School page; other/undecided, any RV variant, governor sharing and weighting not read.
• 2026-ME-SEN: InsiderAdvantage Oct 13-14: primary post gives only Jackson 49 / Collins 47; undecided (4% per aggregators) left null because no ME tabs were posted on the site.
• 2026-ME-SEN: YouGov Sep 2-8 (ttw_me_20260902.pdf): sponsor not stated in the document; sponsor null / sponsor_class UNKNOWN.
• 2026-ME-SEN: MPRC Sep 25-29: deck says only '684 Maine voters'; pop=LV taken from the releasing organization's write-up (Maine Beacon / Maine People's Alliance); 270toWin shows RV.
• 2026-ME-SEN: co/efficient: page states 'Released September 2026' without a day; release_date 2026-09-30 from 270toWin.
• 2026-ME-SEN: Trafalgar Oct 21-23: release_date 2026-10-25 inferred from the report slug and page date (PDF created 2026-10-24 17:20 PDT).
• 2026-ME-SEN: CBS/YouGov: 270toWin lists 1,114 RV; the document reports the Senate ballot among 1,136 weighted likely voters (full sample 1,144 RV). Recorded per the document.
• 2026-ME-SEN: AAPOR Transparency Initiative member list page returned 404; ti_member recorded 'unknown' for every pollster.
• 2026-ME-SEN: No RCV simulation or final-round estimate was published by any pollster in this inventory; every row is round FIRST (single-choice ballot).
• 2026-ME-SEN: Wedgewood Polls and Median Strategies: no Maine Senate release found on 270toWin, Wikipedia, FiftyPlusOne (readable portion) or WebSearch; quarantine_candidates is empty for this race.
• 2026-ME-SEN: Fox News RV row (n=1,689) and YouGov LV row (n=1,335) and the Quantus leaner variant are additional variants read on the primary documents and not listed as such by both aggregators.
• 2026-MI-SEN: RealClearPolling blocked (DataDome 403) on three distinct attempts (Safari UA curl, Googlebot UA curl, WebFetch) plus a web.archive.org capture with no table; the RCP inventory could not be unioned, so listed_on never includes RCP and any RCP-only poll would be missing.
• 2026-MI-SEN: Washington Post-Schar School (Sep 10-14, 803 LV, 48/45): PDF unreachable (curl timeouts x2, WebFetch 403, web.archive.org 403, archive.ph screenshot only); toplines rest on 270toWin + Wikipedia; undecided, denominator class, methodology and governor sharing unknown.
• 2026-MI-SEN: GBAO (Sep 19-22, 800 LV, 48/44): memo PDF 403 at politico.com/f/; numbers from the archived Politico Playbook text only; sponsor ('internal poll') not named; undecided and methodology unknown.
• 2026-MI-SEN: Torchlight Strategies (Sep 15-19): no pollster document; numbers from the Washington Reporter exclusive only.
• 2026-MI-SEN: InsiderAdvantage Oct 9-10 and Oct 21-22: minor-candidate and undecided shares not on the release pages (tabs on X not reached); denominator class inferred from the Sep tabs wording.
• 2026-MI-SEN: Marist October likely-voter n not stated in the release or tables (RV n=1,185).
• 2026-MI-SEN: Z to A Research: population (LV/RV) not stated in the X thread; recorded null though aggregators list LV.
• 2026-MI-SEN: Tyson Group (both waves) and Quantus: undecided (Tyson) and minor-candidate totals (Quantus) not stated; Quantus 'report' with methodology not linked.
• 2026-MI-SEN: Big Data Poll: undecided shares for the LV initial, LV leaned and RV ballots are not in the release text.
• 2026-MI-SEN: Glengariff (Aug 31-Sep 3): Detroit News article paywalled (402); governor-sharing recorded false because the reached WDIV page covers the Senate question only.
• 2026-MI-SEN: Mitchell Research weighting 'to 2022 and 2024 election results' is ambiguous between actual results and recalled vote; classified UNVERIFIED.
• 2026-MI-SEN: FiftyPlusOne poll list is JS-rendered (not in HTML) and DDHQ returned 429, so those finding aids contributed nothing beyond the FiftyPlusOne average.
• 2026-MI-SEN: No Median Strategies Michigan Senate poll was found; quarantine entry recorded with zero rows.
• 2026-MI-SEN: YouGov Sep 22-Oct 4 sponsor not stated in the PDF; recorded null.
• 2026-MI-SEN: Nothing from 270toWin or Wikipedia tested a Republican other than Mike Rogers; no matchup mismatches found.
• 2026-NC-SEN: RealClearPolling live page returned 403 on three curl attempts and WebFetch; inventory taken from the Wayback Machine snapshot of 2026-10-23 18:47 UTC (listed_on 'RealClearPolling' refers to that snapshot). Any RCP row added after Oct 23 is not captured.
• 2026-NC-SEN: Decision Desk HQ polling page blocked (429 / Vercel checkpoint); FiftyPlusOne NC page renders only the average (poll list is client-side JS). Neither could be used as a finding aid beyond the average.
• 2026-NC-SEN: Elon University Poll Oct 20 wave: topline/methodology PDF not located (index page lists only September PDFs; six filename guesses 404). LV n therefore null (aggregators show 561) and registered-voter Senate figures, if published, are not captured. recall_weighted YES rests on the release's sample-matching sentence, not a full weighting paragraph.
• 2026-NC-SEN: Change Research methodology page returned 429 three times; recall_weighted UNVERIFIED. Toplines read via the Google Doc /mobilebasic view (export endpoints 401).
• 2026-NC-SEN: Catawba College/YouGov Sep 4-15: third-party and undecided shares not stated on the release page and no topline link found; left null (Wikipedia shows Bray 5 / another 1 / undecided 13, unverified).
• 2026-NC-SEN: Big Data Poll Sep 27-29: third-party/undecided shares not in the release text (crosstabs hosted on MarketSight, not retrieved); methodology paragraph says 'in Texas' (apparent copy error in the NC release).
• 2026-NC-SEN: YouGov Sep 15-18 (ttw_nc_20260915.pdf): sponsor not identified anywhere in the PDF; release date 2026-09-23 taken from the 270toWin/RCP listing because the PDF is undated.
• 2026-NC-SEN: ECU Oct wave field dates: press release and TI disclosure say Sept 30 - Oct 3; the topline PDF header says Sept 29 - Oct 2. Release-page dates recorded.
• 2026-NC-SEN: SurveyUSA Oct: RCP lists field dates 10/1-10/7; primary says 10/7-10/11 (primary used). High Point: RCP lists 9/6-9/16; memo says Sept 9-16 (primary used). Quantus: aggregators list n=688; primary says 686.
• 2026-NC-SEN: InsiderAdvantage Sep 16-17 toplines exist only as an image on the site; weighting not stated for this survey (UNVERIFIED). Rasmussen Sep 22 found only on the RCP snapshot; its methodology is a general page naming 'other factors' (UNVERIFIED).
• 2026-NC-SEN: Elon University Poll Sep 10 release (field Aug 21-31, 2026) excluded by the window rule (field_end before 2026-09-01); the Sep 17 Elon release is the same survey (ballot amendments).
• 2026-NC-SEN: Harper Polling, Trafalgar Group, Fabrizio Ward/Impact (AARP), Catawba/YouGov: methodology statements name no weighting variables; UNVERIFIED. ActiVote classified NO from the AAPOR disclosure page the release links to (variable list omits past vote), noting its target tables are dated to Nov 2024.
• 2026-NC-SEN: Wedgewood Polls and Median Strategies: no North Carolina Senate release found on RCP (snapshot), 270toWin, Wikipedia, or web search for the Sep 1-Oct 25 window; quarantine_candidates is empty for this race.
• 2026-NC-SEN: This is an inventory from the named sources (RCP snapshot, 270toWin, Wikipedia finding aid, primary releases), not every public poll.
• 2026-OH-SEN-SPECIAL: RealClearPolling direct access blocked (403 on three routes); the RCP inventory comes from a Wayback capture and an r.jina.ai render of the live page retrieved 2026-10-25T18:07:59Z.
• 2026-OH-SEN-SPECIAL: NYT/Siena Ohio toplines (Sep 22-Oct 1, 616 LV) not readable on nytimes.com (403/captcha on four routes); D/R confirmed on sri.siena.edu, n/dates/undecided rely on RCP/270toWin/Wikipedia; recall_weighted UNVERIFIED.
• 2026-OH-SEN-SPECIAL: BGSU/YouGov October 2026 toplines (scholarworks article=1030) blocked by Cloudflare on every route; row relies on RCP/270toWin/Wikipedia for 48/42 and 9/29-10/9; other/undecided null; methodology UNVERIFIED (September statement not substituted).
• 2026-OH-SEN-SPECIAL: Cook Political Report / GS Strategy Group / New River Strategies: toplines and methodology are subscriber-only; 49/43 from RCP and 270toWin; undecided/other null.
• 2026-OH-SEN-SPECIAL: InsiderAdvantage 10/21-22 release: only the headline 47/44 is on the site; tabs were posted to X and not reached, so other/undecided null.
• 2026-OH-SEN-SPECIAL: Marist October likely-voter n is not stated in the release (RV n=1,176); recorded null.
• 2026-OH-SEN-SPECIAL: Trafalgar: 270toWin lists 1,200 LV and 45/42 vs. the PDF's 1,085 respondents and 45.4/41.5/6.0/7.1; PDF values recorded.
• 2026-OH-SEN-SPECIAL: Big Data Poll methodology paragraph names Texas (copy error on the Ohio release); undecided not published; RCP lists 678 LV and 9/25-27 vs. the release's 680 LV and 9/26-27; recall_weighted UNVERIFIED because applicability is not demonstrated.
• 2026-OH-SEN-SPECIAL: YouGov ttw_oh PDF names no sponsor; sponsor null / sponsor_class UNKNOWN.
• 2026-OH-SEN-SPECIAL: Suffolk weighting statement not reachable (marginals PDFs carry none; methodology page and part-1 PDF return landing HTML); UNVERIFIED.
• 2026-OH-SEN-SPECIAL: Fox News weighting statement hedges with 'generally' but enumerates variables (age, race, education, area) that omit past vote; classified NO on that list.
• 2026-OH-SEN-SPECIAL: FiftyPlusOne poll table is JS-rendered; used only to confirm pollster names present. Decision Desk HQ returned 429.
• 2026-OH-SEN-SPECIAL: Verasight: no weighting statement in the survey PDF; the firm's general page lists 2024 vote and the PDF's sample-composition table shows 2024 recall raw vs weighted, but neither is an applicable statement for this survey; UNVERIFIED.
• 2026-OH-SEN-SPECIAL: No Wedgewood Polls or Median Strategies release for this race appeared in any checked source; quarantine_candidates is empty for that reason, not for lack of checking.
• 2026-OH-SEN-SPECIAL: Governor-sharing flag for BGSU October, Trafalgar, Cook/GSG-NR and Siena is recorded false/true from indirect evidence as noted per row; not read on a primary for those four.
• 2026-OH-SEN-SPECIAL: Multiple-variant polls: Quantus additional variants (Q3 voters-only 47.8/47.3; Q3+Q4 two-way with leaners 47.6/47.2) are in the TWO_WAY row's notes because a second TWO_WAY/LEANERS row would collide on the row key.
• 2026-TX-SEN: Pulse Decision Science (Paxton internal, Sept 24-27): no primary memo located; row rests on a Texas Tribune reporter's X post (figures sum to 97). verified_on_primary=false.
• 2026-TX-SEN: Stratus Intelligence (Sept 22-24): memo text extracted from the Politico-hosted PDF confirms the 48-48 / 4 undecided topline but carries no methodology block; field dates, n=856 and sponsor come from Wikipedia only.
• 2026-TX-SEN: SoCal Strategies (Sept 12-13): primary release gone (Substack 404); figures from Wikipedia/Newsweek. Quarantine candidate.
• 2026-TX-SEN: Rasmussen Sept 28-29: n left null because the toplines page body contradicts its own headline ('Poll of 1,115 Likely Voters Conducted September 22-23, 2026'); aggregators say 1,128.
• 2026-TX-SEN: Marist Oct 8-11 LV row: likely-voter n not stated anywhere in the NOS; RCP/270toWin print the RV n (1,196) against the LV figures.
• 2026-TX-SEN: CBS News/YouGov Oct 8-16: LV n recorded as the ballot crosstab's weighted N (1,458); 'Someone else'/'Not sure' shares blank in the toplines, so other/undecided are null.
• 2026-TX-SEN: Big Data Poll and Verasight: third-candidate/undecided shares only partly printed in the release text (Big Data: none; Verasight: Brown only), so undecided is null on those rows.
• 2026-TX-SEN: Texas Public Opinion Research: full toplines live in Google Drive folders that need JS; weighting statement not found, recall_weighted UNVERIFIED despite three TPOR releases read.
• 2026-TX-SEN: Cook Political Report (GS Strategy Group / New River Strategies): article page is behind a Cloudflare JS challenge even in Chrome (no text exposed); the 23-page topline PDF was fetched directly but has no weighting statement.
• 2026-TX-SEN: Fox News/Beacon/Shaw classified NO on the explicit variable list 'age, race, education, and area', but the sentence is hedged ('When necessary, weights are generally applied'); flag if the hedge should move it to UNVERIFIED.
• 2026-TX-SEN: Univision/YouGov Aug 27-Sep 4: the only Senate topline in the crosstabs is 'Combined - Recoded' (leaners folded in); no initial-ask variant published.
• 2026-TX-SEN: AAPOR TI membership checkable only for AARP and ActiVote (visible on the member page) plus Verasight (self-stated); everything else 'unknown'.
• 2026-TX-SEN: FiftyPlusOne's per-poll Texas list is client-rendered and was not extracted; it served only as the average check (Talarico +2.6 as of Oct 25).
• 2026-TX-SEN: RCP 'University of Texas 8/5-8/13 1218 LV' row and TPPF/Overton Aug 24-26 row have field_end before 2026-09-01 and were not collected.
• 2026-TX-SEN: No mismatched matchup found: every release names Talarico (D) vs Paxton (R); Ted Brown (L) appears as a third option on most ballots and is carried in 'other'.

2022 certified-result comparators are computed from reviewed primary returns in weight.py and rendered in Section 6. No secondary-source percentages are injected by this ledger.

Continuation verification receipts

All245 Senate rows preserved; one documented conducting-partnership correction, with stable identifiers retained. Added78 governor estimates (OH22,IA19,MI19,GA14,WI4); primary verification is counted directly from polls.csv above, never inferred from aggregator categories. Every source/round-unverified estimate is excluded from analytics. Governor–Senate pairs require the same explicit sample,pop,denominator and round; independent releases never acquire a shared sample from dates/n alone.

Official2022 inputs and source locators are in research/comparators-2022.json at the calculator commit. Current official NH and signed Ohio PDF returns were read in real Chrome; Nevada official abstract was recovered from its Jan1,2023 archive; Georgia raw official export was recovered from the unprocessed OpenElections archive and its25 qualified governor write-ins independently read from the state certificate. All margins use full valid-vote counts.

Coverage: all13 requests returned staleOctober8 edition. Confirmed zero, partial candidate match, and failed access are distinct. Scores/ranges are parsed from the actual camelCase connector shape.

Validation: weight.py --self-test passed every required counterexample and all existing checks;8 unittest cases passed (actual connector shape, unknown-versus-zero, partnership identity, valid-vote denominator, cache invalidation and dry-run preservation). Source/method CSV merge was validated on a temporary bundle and rerun idempotently before the canonical merge.

Heat:63 SCORED,45 MISSING_TRANSCRIPT,7 FAILED,50 SKIPPED_BUDGET. Lifetime recorded spend$2.801933 of$3.00, reserved$0; meter server stopped. Governor debate captions lacked attributable speaker turns, so no answer or comparison was invented.

Stage 2 primary-source ledger

Primary-verified eligible surveys from the fixed 2026 registry found in the saved finding-aid inventory; survey-specific ambiguity stays UNVERIFIED. Newly encountered partnerships and outside-registry firms are listed without recall classification. SoCal is already quarantined and contributes no calculation. This is not an exhaustive census of all released polls.

SoCal Strategies remains quarantined from every calculation. The saved discovery inventory also lists A2 Insights, AtlasIntel, CNU, Clarity Campaign Labs, FDU, GQR, Kaplan, Quinnipiac, Research Co., Roanoke, Rutgers-Eagleton, Zogby and other outside-registry firms without a recalled-vote classification. Joint InsiderAdvantage/Trafalgar releases are separate conducting partnerships outside the fixed registry; member methods are not borrowed. Discovery inventories are finding aids, not a complete census or a source of numerical estimates.

Certified Virginia returns include write-ins in the full valid-vote denominator. Primary contest and CSV, certification date, state annual report p.8. New Jersey official list, candidate totals on pp.1–4, list dated December 3; a separate certification-signature date was not established and remains UNVERIFIED.

Primary-source recovery: co/efficient’s current legacy URLs render the new website; archived publisher HTML supplies the original survey-specific tables and complete weighting lists. Quantus NJ’s article is in embedded public blog JSON, which supplies its original linked Drive PDF. Quantus’s final VA tweet and original graphic were read through the platform’s public syndication endpoint; its UTC publication time converts to November 3 in Eastern time, before polls opened. Full URLs, hashes and extraction locators are preserved in the local source ledger and code research record. These are recoveries of primary documents, not secondary numerical substitutions.

North Carolina: official absentee data, layout, identifier definitions. The 2026 export was last modified before the fixed analytical cutoff. The aggregate-only results and explicit deduplication exceptions appear in Section 2b. No row-level voter data enters the publication bundle.

Ratings: Silver Bulletin January 2026 update, free XLSX source with exact metric/header/hash and explicit identity mapping. The source’s January ratings include 2025 outcomes; they are not used to manufacture an independent 2025 validation result. Provisional status and sparse histories are limitations, not endorsements.

Public-read correction: source facts above remain unchanged. Marist Michigan primary PDF was rechecked; 1,185 refers to registered voters, not an independently reported LV sample count. No distinct LV n was established. The latest verified fallback exposes the published result without manufacturing metadata or changing average eligibility. All current CSV input bytes and prior analytical fields are unchanged.

Deferred

• House effects sensitivity (leave-one-firm-out, time-matched peers within 7 days, at least two independent peer firms, at least three independent waves; never applied to the recall comparison).
• Herding (no RCP average reconstructed).
• Correlated error (disabled sketch; probabilities only in Stage 3).
• Project-specific track-record ledger: last eligible poll per pollster and race ending within 21 days of the 2022 or 2024 election, U.S. Senate and governor contests only, excluding state-legislative Senate; signed error = poll margin minus certified margin; MAE across observations; track_status LEDGER when built.
• Deferred races: KS, NE, FL-SPECIAL, SC, MS, NH Senate; AZ, NV, FL, KS governors.
• Florida turnout file (2b).
• Design effects in f_sample.

Heat summaries describe scored units only. Units without attributable captions or without budget remain explicitly missing or deferred; partial readings do not establish a complete debate comparison.

Heat units by status: {'SCORED': 63, 'MISSING_TRANSCRIPT': 45, 'FAILED': 7, 'SKIPPED_BUDGET': 50}. Recorded spend: $2.801933 of $3.00.

Primary-unverified poll rows and unsupported RCV-round labels remain inventoried and excluded from all calculations. Stage 2 uses a primary-verified retrospective subset with sparse verified recall classes; no weighted-heuristic accuracy improvement is established. Full held-out baseline evaluation, spending diagnostics and a project-specific historical ledger remain deferred. Stage 3 remains gated until November 2.