CTMM 2.2

Section 1

Overview

One staged model for 506 races, fitted on history and validated out of fold.

CTMM 2.2, the Call the Map Model, forecasts the 435 United States House districts, the 35 Senate seats and the 36 governorships on the ballot on November 3, 2026. For every race it publishes a central margin (Democratic minus Republican, in percentage points), a win probability, 50/80/95% intervals and, where the ballot is not a plain two-party pair, the probability that each named candidate wins. From the same set of simulated outcomes it publishes expected seats, seat distributions, the probability that each party controls each chamber and the joint outcomes across chambers.

The model is staged. A structural prior gives each race a starting margin from its terrain, history and incumbency. The national environment moves every race by a fitted elasticity to the generic ballot. Race polls, where they exist, pull the centre toward an evidence-weighted average. Expert ratings are excluded in every office, both in training and current prediction. Separate centre calibrations fitted on out-of-sample results address measured compression in Senate and governor races. Reconstructed House priors retain their fitted values instead of being truncated at 40 points. A fitted error distribution turns each centre into probabilities and intervals. Finally one joint simulation draws every race together, with national, regional and state shocks, so that seat totals and chamber control reflect the dependence between races.

Fitted coefficients are estimated within historical training windows, not chosen to produce a desired 2026 race margin. Structural choices, clipping bounds and candidate-transfer ranges are declared assumptions; they are not all fitted parameters. The historical record supporting the fits runs from 2008 to 2024. The validation record covers 2018, 2020, 2022 and 2024 at six forecast origins (60, 35, 21, 14, 7 and 0 days before the election), with the test cycle held out of parameter fitting. Section 13 reports that record; Section 15 states what four cycles cannot settle.

The frozen comparable-race poll-residual overlay from the earlier model is not a CTMM 2.2 input. Its stale donor residuals are not added to this model's own polling update. A September 8 sensitivity comparison found changes below 0.20 points from restoring that overlay, not a solution to safe-seat compression; a newly fitted borrowing model remains research. Historical v1 forecasts and their decompositions remain archived.

Two conventions run through the document. Margins are Democratic minus Republican, so a positive centre favours the Democrat; and a horizon h is the number of days between the forecast origin and the election, 53 days for the board this document accompanies.

Section 2

Forecast-origin data

Everything the model reads is dated on or before the forecast origin.

A forecast made on a given day may only use what existed on that day. Race polls require evidence of public availability, not only a field-end date; the generic ballot also uses readings available by the origin. Structural features use the appropriate prior election and geography. Expert ratings are excluded. The same input policy applies to current prediction and historical evaluation.

Structural inputs

Candidate nomination status and input eligibility are distinct. CTMM 2.2 has no blanket nominee-status veto: a poll must meet the actual source, ballot and date rules, but an empty admitted poll set does not mean polls were withheld. Reconstruction retains the geography operative on each date. No expert-rating source supplies a forecast input.

House districts: the district's prior two-party margin on its current lines (uncapped where the result is observed on unchanged lines or officially re-aggregated), the margin two elections back on unchanged lines, the district's 2024 presidential margin relative to the nation, incumbency and its strength, open-seat and defending-party signs, the district's prior over-performance against its state, a prior-uncontested flag, the midterm and White House party indicators, the previous certified national House margin, three American Community Survey demographics (educational attainment, median household income and median age) and the nine Census divisions. Senate seats: the seat's own prior margin and the state's most recent Senate margin (calendar rule; an appointed seat's prior is the state's last regular election), a prior-uncontested flag, relative presidential lean, incumbency, open-seat sign, defending party, the midterm indicator, the previous national margin and two interactions. Governorships: relative presidential lean, the prior gubernatorial margin by the calendar rule, the midterm indicator, signed incumbency and tenure, and the candidate-experience differential (the highest elected office each nominee has held, 0-3, Democratic minus Republican).

Evidence at the origin

Race polls. The source-admitted poll feed, each reading carrying its pollster, field end date, sample size and sponsor class; a Democratic- or Republican-sponsored poll is flagged partisan. On this board 73 of 435 House races, 28 of 35 Senate races and 30 of 36 governor races carry at least one admitted poll.

Independent of expert ratings. No publisher rating, consensus or rating-derived target enters the House, Senate or governor model. Archived earlier editions retain their original inputs; the current model does not read those ratings.

Generic ballot. The trailing 60-day mean of admitted public generic-ballot readings (Democratic minus Republican). At this origin: +6.2 from 93 polls.

Campaign finance. Leading Democratic and Republican receipts per Senate race are read from FEC filings and gated into a share in [−1, 1]. The finance leg is present in the code and switched off by the backtest in every office (Section 6), so no dollar figure moves a published centre.

Prediction markets, presidential approval and special-election results are not inputs: approval added nothing to a six-cycle environment regression and special elections have no origin-dated historical source; markets are shown on the site for comparison only.

Section 3

Structural foundations

A robust Lasso per office, fitted on uncapped certified history.

The structural prior is one penalised regression per office of the certified two-party margin on the features of Section 2. Nothing from a later evidence family reaches it: no polls, no generic ballot, no ratings, no finance. Features are median-imputed and standardised on the training rows; the regression is a Lasso whose penalty α is chosen by leave-one-cycle-out mean absolute error inside the training window from the grid {0.05, 0.1, 0.2, 0.4, 0.8, 1.5, 3.0}; and the loss is Huber (δ = 15 points), applied by iteratively reweighted least squares, so a landslide bounds its own leverage without being discarded.

The fitted penalties are α = 0.2 for the House (2,145 district-cycles, 2014-2024), α = 0.8 for the Senate (298 seat-cycles, 2008-2024) and α = 0.1 for governors (216 race-cycles, 2008-2024, odd-year elections included). The Senate list is deliberately compact: about three hundred rows cannot support thirty columns, and a wider list extrapolated a 2018 fold to a sixty-five point margin in New York. Races won by a candidate of neither major party are excluded from the fit (ME 2012, VT 2012, ME 2018, VT 2018, AK 2022, ME 2024, VT 2024 for the Senate; ME-02 2022 for the House; RI 2010 for governors).

The prior-margin feature enters uncapped. The House regression's largest coefficients are the prior margin (+12.31 points per standard deviation) and the relative presidential lean (+14.28); incumbency adds +3.03, and the midterm-against-the-White-House term -2.96. For governors the candidate-experience differential carries +6.19 and presidential lean +6.76. The full coefficient tables are in Appendix B.

Every later layer is fitted on the prior's leave-one-cycle-out centres, not its in-sample fit. A layer fitted on in-sample residuals would learn the prior's optimism; one fitted on held-out centres learns its real out-of-sample error, which is the quantity polling and calibration then have to correct.

The Senate state-history feature keeps the latest available regular contest two, four or six years earlier on its uncapped scale in both training and prediction; an uncontested contest stays missing. For MODEL_RECONSTRUCTED House inputs, a robust regression estimates the previous House margin from presidential votes on the same lines. Its full value supplies the main lag; bounded derivatives are recalculated from the replacement parent, retaining their historical training definitions. Observed and partial-official results keep their separate policies. Legacy producer files are preserved, not mistaken for this model's effective prior.

The reconstructed House prior is a + bP, where P = 100(D_pres - R_pres)/(D_pres + R_pres), in two-party margin points. The fitted intercept is -0.5887, slope 1.0679, and Huber epsilon 1.35. Robustness is selected by nested held-state prior-election errors. The coefficient fit uses 38 complete official 2024 reaggregations from North Carolina and Texas, not competitor forecasts. This limited donor geography is a generalization risk. Dated hindcasts use matching dated presidential vote totals; the general matrix's all-candidate margin is not silently substituted.

In the 2022 rolling-origin comparison, covering 402 races at six horizons, the House input replacement lowers overall MAE from 6.944 to 6.705 and safe-seat MAE from 8.473 to 7.927. Close-race MAE rises from 3.588 to 4.280; within the affected close subset it rises from 2.997 to 4.249. CRPS and interval scores improve, while Brier and log loss worsen slightly. With uncertainty fitted only on 2018/2020, 80% coverage rises from 69.61% to 72.47%, still below nominal. This is one election cohort and an explicit tradeoff, not an improvement in every statistic.

Section 4

National environment

A fitted elasticity to the generic ballot, mapped onto the realized national margin.

The structural prior already carries the average environment of its training cycles through its intercept and lag terms. The environment leg adds the news: how far the current generic ballot says the coming election differs from that average, scaled by how much each office's structural error has historically moved with the national result.

β_o is the through-the-origin regression of each training cycle's mean out-of-fold prior residual on that cycle's realized national House margin (a result read only for cycles strictly before the forecast cycle), clipped to [0, 1.5]; (a, ρ) is the regression of the realized margin on the election-eve generic ballot over the cycles that have one. The fits are β = 0.3068 (House), 0.5210 (Senate), 0.2996 (governors); a = -4.5215, ρ = 1.3984; ē = -0.0828 for the House and +0.5504 for the statewide offices. ρ above one says the generic ballot has understated realized margins over these cycles; a absorbs its systematic bias.

At this origin the generic ballot of +6.2 maps to an environment estimate of +4.2, and the leg adds +1.30 points to every House district, +1.88 to every Senate race and +1.08 to every governor race. The uncertainty of the environment is not carried here as a parameter: it enters the joint simulation as the national shock of Section 11, whose scale is estimated from the cycle-level residuals that remain after this leg.

Section 5

Polling

One poll's worth per pollster, a weight that vanishes without evidence, and a poll-bias posterior.

Polls are aggregated pollster by pollster and then weighted against the centre in proportion to how much evidence they carry. Three constraints are structural and tested: the weight goes to zero as the evidence goes to zero, so an unpolled race keeps its prior; one physical pollster contributes at most one poll's worth of evidence, so repeated releases refresh recency without multiplying influence; and sponsor precision multiplies the absolute evidence, so lowering every poll's precision lowers polling's influence rather than only redistributing it.

f(E) is either E/(E + k) or 1 − exp(−E/k), with k inflated in safer seats, k_eff = k (1 + κ|c|/20), so a lopsided race needs more evidence to move; w_max(h) = clip(w_eve − w_slope h/60, 0, 1). All of these, and the aggregation constants, are chosen per office by pooled out-of-fold absolute error over a preregistered grid. The executing values are in the table; the half-life of 60 days and the saturation constant k = 0.25 were selected in every office.

The poll-bias posterior

Polls have missed in the same direction across whole cycles. For each training cycle the election-eve aggregate's mean signed error is computed; the shift b is the mean of a normal-normal posterior with prior N(0, 2²) over those cycle errors, and its posterior standard deviation is kept. For the Senate the shift is -1.57 ± 1.02 points and for governors -2.53 ± 0.92: polls in these offices have read Democratic candidates too favourably, and the aggregate is shifted accordingly before it is blended. The House has no shift: by ablation it worsened both error and Brier score. The executing interval fit uses full-stack out-of-fold errors, including realized polling errors. The bias posterior is represented as a shared shock in the joint simulation, with national-shock variance adjusted to limit double counting; it is not added again to the new Senate marginal scale.

Section 6

Candidate information

Incumbency and experience in the prior; the finance leg present and switched off.

Candidate information reaches the centre through three routes. Incumbency, its strength and tenure sit in the structural prior of every office. For governors the prior also carries the candidate-experience differential: the highest elected office each nominee has previously held on a four-step ladder, Democratic minus Republican, the one candidate-quality term that improved the out-of-fold record. Nothing else about a candidate, no fundraising, no ideology score, no scandal flag, moves a published centre.

The finance leg is executed and evaluated in every fold. It is switched off in every office by the backtest's own ablation: for the House the coefficient was negative and the ablation difference indistinguishable from zero; governors have no as-of-origin finance source; for the Senate removing it improved the Brier score with a confidence interval excluding zero and cost a mean absolute error inside noise. It stays in the code as a refusable, tested component, and the stage c₂ is identical to c₁.

Candidate fields, who is actually on the ballot and under what rule, are a separate matter and are handled in the simulation (Section 8).

Section 7

Model independence

No expert ratings enter any forecast or fitted training layer.

Expert ratings were removed from all three offices by the owner's September 9, 2026 decision. The expert-free path excludes rating loaders and targets, then refits downstream centre calibration, error distributions and dependence using expert-free historical predictions. This is not a manual partisan shift or a claim that every validation score improves.

The historical expert-enabled model remains archived for comparison. The removal is an independence and input-policy choice; its costs and benefits are evaluated on election outcomes, not on whether current forecasts move toward another forecaster.

Section 8

Candidate fields and election mechanics

Same-party pairs, extra candidates, ranked choice, runoffs and independents, inside the same draws.

Most races are a Democrat against a Republican and the simulated margin decides the winner. Where the ballot is different, the same field draw decides candidate winners and party seats. Same-party concentration and ranked-choice transfer parameters remain declared ranges integrated over simulation batches; their historical evidence is limited. The extra-candidate logit scale is now fitted separately to historical poll-to-election error, as described in Appendix C. It is not adjusted to produce a desired win chance in a current race.

• Same-party general elections (California's top-two, where two candidates of one party advance). The party is fixed; the two candidates split the party's vote by a Dirichlet draw centred on the primary shares, with a concentration drawn log-uniformly on [8, 60].
• Extra candidates (an independent or third-party candidate beside the pair). A polled extra takes a logit-normal share around its polled mean, read from the race's own field polls (9 races at this edition: AK-AL, CA-06, CA-07, CA-11, CA-40, GOV-AK, GOV-ME, SEN-MS, SEN-MT). A poll is admitted when it offers every candidate of the current field that this race's polls can measure — a candidate no pollster has ever named cannot be required, and one who has withdrawn is dropped from the denominator rather than counted against those still running. Field tables never enter the Democrat-versus-Republican margin. An unpolled extra takes a multiple of the race's admitted 'other' share, the multiple drawn uniformly on [0.30, 0.90]. An extra sharing a pivot's party is subtracted from THAT pivot; a cross-cutting candidate is split between the two in a proportion drawn uniformly on [0.25, 0.75]. Where a race's own current-field polls measure both pivots, the two-party centre is recentred on that measurement — dispersion untouched — and the winner is the plurality of the resulting shares.
• Ranked-choice and top-four counts (Alaska, Maine). Rounds eliminate the lowest candidate and transfer its votes with affinities drawn per batch: same party [0.45, 0.75], the opposite major party [0.05, 0.30], non-major [0.15, 0.45]; a share of each transferred pool, drawn on [0.05, 0.30], exhausts each round. The count ends when a candidate holds a majority of the continuing vote.
• Runoffs. A race whose rule sends the top two to a later runoff is simulated on its two-candidate coordinate; extras present in the first round transfer as above.
• Independents. A race whose second pivot is an independent is simulated on its D-versus-R coordinate. Nebraska's Senate race, where the independent Dan Osborn is the alternative to the Republican, is not credited to Democratic control: the Senate control probability is reported both with and without it.

Field specifications come from the governed candidate-field records: certified top-two pairs, ranked-choice and top-four rules, extra candidates with any polled share, and the race's admitted 'other' share. On this board 22 races carry a field specification beyond the plain pair.

After subtracting an extra candidate's votes, a pivot share cannot fall below zero. The nonnegative candidate shares are then rescaled together to the available ballot total, 100% minus the remaining 'other' share. This common factor preserves plurality ordering and ensures each simulated ballot totals 100%; ranked-choice transfers start from that consistent ballot.

Candidate high/low bands are vote-share percentages. History bands are margins in percentage points. In a complete two-candidate ballot a candidate's share is (100 + margin)/2, so its share interval is half as wide as the margin interval. That shortcut does not apply to multiparty shares or ranked-choice rounds.

Section 9

Centre calibration

Smooth Senate tail and poll-drag correction; governor tail stretch and G1 calibration, all fitted on historical results.

After the evidence updates, governors receive the tail and G1 calibrations below. Senate races receive a smooth, party-symmetric correction for measured compression. House centres are not globally stretched: a fitted presidential-to-House estimator replaces reconstructed prior inputs before their dependent structural features are built.

Senate: a smooth tail and inward-polling correction

Nonnegative coefficients between zero and one are fitted by bounded least squares, giving equal total weight to historical result bands below 5, 5–15, 15–25, 25–40 and 40-plus points. Inner leave-one-cycle-out selection tries K = 5, 10, 15 or 20; T = 2 or 5; W = 8, 16 or 32. It minimizes error in safe seats (absolute result at least 25), allowing less than one additional point of overall or close-race MAE versus the existing stack. This explicit trade-off was authorized on September 8. The zero correction remains available. No competing forecast or desired 2026 margin enters fitting. Senate uncertainty is refitted after the correction.

The governor tail calibration

The earlier centre-family selector remains unchanged: House and Senate chose identity there, while governors selected the tail expansion with b_t = 1.1319. The separate Senate correction above follows that stage and uses its explicitly stated tail-focused objective.

G1: the governor slope and intercept

Out of sample the stack had read safe governor races too close: the signed error of races decided by 15 points or more ran -5.73 points toward the prediction, and safe-race error was 7.57 points against 2.95 in competitive races. G1 is the smallest correction that repairs most of it: with the same folds the safe signed error becomes -2.98, safe-race error 6.87, overall error 5.873 to 5.712 and root mean squared error 7.50 to 7.28. The cost is real and accepted: competitive-race error rises from 2.95 to 3.34, because a global slope moves a close race by about a point too. Nested leave-one-cycle-out (the calibration refit with each cycle held out) is what these figures report, so the improvement is not the fit admiring itself.

The fit is computed for every office and applied to governors alone. For the House it is nearly the identity (a = -0.1558, b = 0.9716). For the Senate it is a = -2.6362, b = 0.8922: the intercept is the Senate's historical Democratic centre bias, discussed in Section 15, and it is not applied; the separate smooth correction addresses compression without a global partisan offset.

On this board the largest G1 movement is HI (+31.5 to +37.1); Colorado's governor race moves from +17.4 to +20.9 with a win probability of 98.6% and an 80% interval of +10.5 to +31.2. Every published governor record carries both the pre-calibration and the published centre.

Section 10

Race uncertainty

A core+tail mixture for the House and governors; one coherent Student-t law for Senate probabilities and intervals.

A race's uncertainty describes forecast error e = result − centre. The House and governors retain a two-component normal mixture, with a race-specific core and a pooled tail. The Senate retains a Student-t family, now fitted on full-model out-of-fold errors. Its probability and every interval level use the same CDF. Candidate-field winners and party seats are resolved in the joint simulation; an analytical D-versus-R coordinate cannot give a Democratic win chance to a ballot with no Democrat.

F: the core+tail mixture (House, governors)

The mixture's variance for the illustrated race is 63.6 points squared, but its shape is not a normal's: a race with a 19-point centre keeps a win probability near 99% rather than certainty, because the tail component has real mass twenty points from the centre. Every published House and governor record declares its family, core scale, tail scale and tail share, and a release gate recomputes the probability and the 80% interval from the declared parameters and refuses the edition on any mismatch.

The Senate scale law

The executing Senate values are a = (+2.3272, +0.0127, -0.2819, -0.1598, +0.4760), ν = 8, z = 0.9493, b_p = 1.0000 (identity, not applied), and z_0.50 = 0.671, z_0.80 = 1.326, z_0.95 = 2.189. Held-cycle standardized residuals supply their 80th absolute-error percentile; z preserves that percentile while ν is selected by likelihood from {4, 6, 8, 12, 30}. Training uses the full-stack 2012–2024 out-of-fold record and excludes the evaluated cycle. Historical residuals already include realized poll errors; no separate polling-error variance is added again to this marginal scale.

Section 11

Correlated simulation

One joint draw of every race, with national, regional, state and residual components whose scales are themselves uncertain.

Races do not miss independently. A national mood that the polls did not see moves every race in the same direction, regions move together, and districts in one state share their state's swing. The joint simulation draws every race's error as the sum of four components, so that seat totals and chamber control carry the dependence a reader would expect from a wave.

The scales are estimated from the model's own out-of-fold election-eve residuals at the 60-day horizon, 2018-2024, with the election cycle as the independent unit: s_N,o is the standard deviation of cycle-mean residuals (3.56 House, 3.58 Senate, 4.01 governors), s_R,o the root mean square of region-by-cycle means after the cycle mean (1.37, 1.89, 3.35), ρ_o the same-state residual correlation (0.04, 0.80, 0.00), and λ_o the share of the national variance common to all offices, from the shrunk correlations of cycle-mean residuals between offices (1.00, 0.85, 0.91).

Four cycles cannot pin a standard deviation down, and the simulation does not pretend they can. Each of the 2000 batches of the simulation draws its own parameter vector: the national and regional scales from their posteriors under a cross-office hierarchical lognormal prior (the House national scale's 10th, 50th and 90th percentiles are 2.8, 3.7 and 4.9 points), and ρ and λ from Fisher-z posteriors with the cycle count as the effective sample. The published seat distribution therefore integrates over parameter uncertainty; a fixed-parameter sensitivity grid is reported beside it.

The poll-bias common shock and the tail indicator

The Senate and governor poll-bias posteriors of Section 5 are drawn once per batch in one shared direction, each office taking its own Student-t quantile of it (sd 1.02 with 8 degrees of freedom for the Senate; 0.92 with 10 for governors) and added to every race in proportion to its poll weight. Because the national scale was estimated from residuals that already contain the realized shift error, its variance is reduced by the common shock's share, floored at a quarter of its value, so nothing is counted twice.

For House and governor races the tail indicator of F is drawn per race per draw with probability π = 0.0580. The shared national, regional and state components are unchanged, being sized from the core scale; the tail's extra variance enters only the race-specific residual, so that conditional on the indicator the race's marginal is exactly N(0, s_core²) or N(0, s_tail²) and the unconditional marginal is the published mixture.

The simulation runs 100,000 joint draws. A reconciliation over ordinary House and governor races (456 races, 200,000 draws) finds a maximum difference of 0.0025 between simulated and standalone win probabilities (Monte Carlo standard error 0.0011 at even odds) and 0.11 points in the 80% quantiles. Senate dependence begins with correlated normal scores; their normal-CDF ranks are transformed through the fitted Student-t inverse CDF. The resulting Senate marginal uses the same scale and tail shape as its published probabilities and intervals, including an exactly tied race. Candidate-field rules subsequently determine candidate winners.

Section 12

Chamber control

Seat totals and control probabilities as functionals of the same draws.

A draw is a full election night. The House total is the number of the 435 simulated districts the Democrat wins; a majority is 218 seats. The Senate total adds the 34 Democratic-aligned seats not on the ballot to the simulated wins; control needs 51. The governor count is the number of the 36 contests the Democrat wins, of which 19 is a majority of those on the ballot.

On this board the House expectation is 229.5 Democratic seats (draw mean and marginal sum differ by 0.00), with a 73.8% probability of a Democratic majority and a 90% seat range of 204 to 261; the Senate expectation is 50.3 Democratic-aligned seats, control 48.0% (47.0% without crediting Nebraska); governors 19.5 of 36. Cross-chamber outcomes (both chambers to one party, split control) are read from the same draws, and the tipping-point and seat-distribution products are regenerated from the persisted draws by the exporter, never from a second simulation.

Section 13

Historical validation

Leave-one-cycle-out over 2018-2024 at six horizons, on the executing code.

The validation protocol is nested rolling-origin, leave-one-cycle-out: for each of the four full-tier cycles every fitted layer (prior penalty, environment, poll curve, centre calibration, G1, F and the Senate law) is refit without the evaluated cycle, the held-out cycle is scored at all six horizons with the evidence available at each origin, and the scores are pooled. Expert ratings are absent throughout. The numbers below are those of the executing code, not of a research replica. Older cycles (2012-2016) are not pooled with the full tier.

Three things to hold against the table. The House and Senate probability slopes are near one (0.99 and 1.34); the governor slope of 2.16 means governor probabilities remain under-confident, an artefact of the office's small folds (35 races in 2018, 11 races in 2020, 36 races in 2022, 11 races in 2024). The Senate's mean signed error of +2.75 points is the Democratic centre bias of Section 15. And 80% coverage is 0.79, 0.81 and 0.80: the House and governor intervals are honest to within two points of their nominal level, the variation across margin bands matters as much as the overall coverage.

The race distribution, by ablation

The distribution is judged by margin scores (CRPS and weighted interval score), winner scores (Brier and log loss), coverage and the shape of its bands across safer and closer races. The 2.2 non-saturating extension explicitly trades small aggregate score losses against better-calibrated House tail and middle-range widths: compared with the seven-term mixture, held-cycle House CRPS worsens 0.0105 and WIS 0.0115, while 80% coverage above a 40-point predicted margin improves from 70.4% to 75.2%. This is not an improvement in every score or every cycle. The table compares the executing distribution with the retained Student-t reference law.

Section 14

Hindcast and issued history

Reconstructed points are faded and labelled; issued forecasts are stored once and never rewritten.

The site's history charts show two kinds of point. From 2026-08-01 to the day before launch the series is a reconstruction: today's CTMM 2.2 method scored at each historical date, not a forecast published then and not an out-of-sample backtest. The replay starts with dated repository captures and supplements polling records collected later only when a publication date, archive-entry date or earlier observed capture establishes availability by the origin. Fieldwork must also have ended by that date. A field end alone is not evidence that a poll was already public. Unknown availability excludes an individual poll, not the whole day's reconstruction. Expert ratings are excluded from the reconstruction, as they are from the current forecast. House district/prior rows come from the historical matrix rather than today's boundaries. Capture dates can lag a real-world announcement; per-date input manifests record those vintages and limitations. Every reconstructed point is labelled reconstructed = true and issued = false and drawn faded.

From the launch marker, 2026-09-09, issued points come from forecasts that reached publication; the current page appends its own reading without splicing in earlier-model history. Inputs are checked several times a day, and a change in model-relevant evidence can produce a new live forecast at any of those checks. Separately, one permanent daily snapshot preserves that scheduled archival reading and its content hashes. Daily archiving is not a limit on live updates, and neither operation requires rebuilding the website. Historical snapshots are immutable; corrections create new records rather than rewriting an issued forecast. The reconstruction contains 40 daily points through 2026-09-09.

A reader comparing two points that straddle the launch marker is comparing a reconstruction with an issued forecast; the chart marks the seam and the site's movement surfaces say so. Movement across a model-version date is recorded in the model-revision register as a change of method, not as news about the race.

Section 15

Limitations

What four cycles cannot settle, and what the model states rather than corrects.

• Few comparable cycles. The full validation record is four elections. Every pooled figure in Section 13 hides a spread across them (2020 was a bad Senate year for polls; 2022 a bad House year), and the dependence scales of Section 11 are estimated from four cycle means. The simulation integrates over that uncertainty; it cannot remove it.
• Residual Senate bias remains. Out of fold, 2018-2024, the Senate's mean signed error is +2.75 points (positive means too Democratic). The smooth correction targets safe-seat compression, not a desired partisan result. A global Senate G1 intercept and slope are not applied. Improvement in tail error does not establish that every aspect of calibration improved.
• Governor probabilities are under-confident. The calibration slope of 2.16 says a stated 80% has historically won more often than 80% of the time. The folds are small (governor races come 11 to 36 at a time) and a slope estimated on them would be an unstable correction; the under-confidence is left in.
• Safe-seat error remains. The signed safe-seat error is -0.3 points for the House, +1.0 for the Senate, and -3.0 for governors; negative values indicate compression. A smoother central margin does not by itself establish accurate uncertainty bands in the most lopsided races.
• Unusual fields rest on declared ranges. Same-party pairs, extra candidates, ranked-choice transfers and exhaustion have no historical panel large enough to fit. Their parameters are declared ranges, the probabilities integrate over them and sensitivity intervals are published, but a range is a statement of ignorance, not a measurement.
• Rare mechanics are approximated. Runoffs are simulated on their two-candidate coordinate; independents on the D-versus-R coordinate; an unpolled extra candidate takes a share from the race's 'other' vote. In a race like Alaska's top-four count the person the simulation favours can differ from the sign of the two-party coordinate, and the site shows both rather than reconciling them by hand.
• Expert ratings are excluded. No third-party rating or consensus target enters current forecasts. Earlier expert-enabled editions remain archived, not rewritten.
• Pre-launch history is reconstructed. Points before the launch marker were computed afterwards on today's poll bank. They are the best available account of what the model would have said, and they are not forecasts anyone could have read on those dates.
• Refused features stay refused. Presidential approval, special elections, poll trend and disagreement, pollster house effects, latent-environment models, splines and interactions in the priors, statewide indices and latent candidate quality were each tested on the historical record and did not improve it out of fold. They are documented as refusals and are not in the model.