H5: Channel B contribution decays as the election approaches

Sponsor-paid slant in a Brazilian electoral poll has a future cost: the election eventually arrives and returns a public, granular check on the poll's headline numbers. The cheaper-to-fabricate margin is Channel B (residual / fabrication) — the part of the slant that the disclosed signal structure does not justify. As the election nears, verification looms, and Channel B should be suppressed. The quantity of interest is the sponsored_by × days_to_election interaction after Channel A controls: a negative interaction (slant larger far from the election, smaller near it) is the prediction.

Evidence strength: Mixed by AN-005 (2026-06-02). Without Channel A controls in place yet, the sponsored_by × days_to_election slope is +0.037 pp/day (SE 0.019, p = 0.054) — implied β ≈ +13 pp at 6 months out, +7 pp in the final week. Direction matches the verifiable-disclosure prediction; magnitude is economically large; significance is marginal and the spec does not yet separate Channel A from Channel B.

Theory

The framework is Polls as verifiable disclosure (theory.md §"Polls as verifiable disclosure"). Crawford & Sobel (1982) cheap-talk baseline + Milgrom & Roberts (1986) verification-cost extension applied to the election-day-verification structure. The sender chooses slant magnitude τ subject to a cost c(τ) × P(detection), where P(detection) is decreasing in time-to-election. Detection cost is multi-channel: reputational (a contradicted pollster loses future media commissions), regulatory (LE.33.§4 criminalizes fabricated polls — detenção 6 mo – 1 yr + multa), and commercial (the sponsor's party loses narrative control if the poll is publicly refuted) [institutions.md §"Compliance and sanctions"]. Channel A (theory.md §"Polls as Bayesian persuasion") is essentially time-invariant because methodology choices are committed at registration; the verifiability channel acts on Channel B.

Prediction

The sponsored_by × days_to_election interaction, after Channel A methodology controls, is positive (slant rises with distance to the election). Polls in the final 1–2 weeks should show β^B ≈ 0; the action is at the longer horizon, where detection is far off and discounted.

Competing predictions

Bandwagon-incentive demand. The verifiable-disclosure story is a cost-side claim. The bandwagon channel (theory.md §"Polls as bandwagon triggers") makes the opposite prediction on the same interaction: voter-side bandwagon returns are larger near the election, so the sponsor's demand for slant is higher near the election. The two effects pull in opposite directions on days_to_election × sponsored_by. The observed sign discriminates; a null is consistent with both operating and roughly cancelling.

Channel A residual. Even if methodology choices have to look more defensible closer to the verifiable event, design-driven slant could also decay with proximity. The clean Channel B identification requires Channel A controls (H3) to be in place. AN-005 does not yet absorb the methodology bundle, so the decay slope mixes Channel A and Channel B attenuation.

Prior research

Brazilian press postmortems of polls that missed late mayoral results predominantly target late polls — the discrepancy is newsworthy because verification is immediate — but no formal estimate of the time gradient exists. ABEP industry-side commentary names auto-financiamento as a "cover" channel for irregularidades, institutional confirmation that non-verifiable polls accumulate slant [stories.csv #077]. LE.33.§4 criminalizes fabricated polls (detenção 6 mo – 1 yr + multa) [institutions.md §"Compliance and sanctions"]; the penalty bites post-election, sharpening the time-decay logic. The cheap-talk-with-verification template is Crawford & Sobel (1982) sharpened by Milgrom & Roberts (1986).

Evidence

Analysis	Bearing	Key takeaway
AN-005	Mixed	sponsored_by × days_to_election slope = +0.037 pp/day (SE 0.019, p = 0.054). Implied β rises from +7.00 pp at 7 days out to +13.41 pp at 180 days out — about half the headline β decays over the 6 months before the election. Direction matches verifiable disclosure; marginal significance and absence of Channel A controls keep this Mixed rather than Confirmed.

Open tests

Decomposition after Channel A controls

The clean Channel B test re-estimates the sponsored_by × days_to_election interaction after absorbing the declared methodology bundle (the H3 specification — see H3). The Channel B prediction is that the interaction survives the methodology controls; if it collapses, the decay is Channel A, not Channel B. Blocked on the poll_methodology LLM extractor queued in pipelines/politica/docs/todo.md.

Sample-split estimator

A two-bin estimator (last-month polls vs early polls) rather than the continuous interaction would sharpen interpretability: directly report β_late and β_early with a difference test. Suggested follow-up in AN-005.

Sponsor-route heterogeneity

Interacting the decay slope with sponsor route (CPF, committee, party) would tell us which sponsor types let slant decay most. Hypothesis from AN-005: Route B (committees) — professional staff with verification awareness.

2022 cycle extension

p = 0.054 is borderline. The 2022 cycle would add time-to-election variation and candidates, with negligible additional Channel A ambiguity since the design carries over.

Supporting analyses