H12: High-β polls are over-represented in poll-related Representações

H12: High-β polls are over-represented in poll-related Representações

If markets approximate the same posterior as the econometrician, polls that look most slanted on residual β should be the ones that draw legal challenge. The prediction is that protocols in the top decile of estimated residual β-hat are more likely to be the subject of a poll-related Representação (EJ assuntos containing PESQUISA ELEITORAL) than the bulk of the distribution — and that plaintiffs' filings should name the dimension on which the slant operates (methodology, sample, sponsor identity).

Evidence strength: Refuted by AN-072v2 (2026-06-16). Poll-level OLS within race × week gives β = −3.9 pp (SE 1.5, p = 0.010, n = 8,943 protocols, 7,274 race-week clusters) on the fraud-flavored assunto bucket. The candidate-level companion AN-072 (+3.15 pp, p = 0.011) runs the predicted direction but operates on the wrong unit and reflects race selection, not poll-level scrutiny. The statutory reading in institutions.md § "Court interpretation" independently shows the adversarial channel cannot reach declared-methodology slant.

Theory

The framework is Polls as Bayesian persuasion (theory.md § "Polls as Bayesian persuasion (supply-side / Channel A)") together with its supply-side complement on verifiable disclosure (theory.md § "Verifiable disclosure and ex-post detection"). If receivers approximate the econometrician's posterior, transparent slant should be downweighted at consumption — and where slant is egregious, a litigant should petition under LE.96. Litigation rates become a market test of perceived bias: a poll that looks tilted should draw a representação if the receiver-side discount mechanism is real and the legal regime is the venue where that discount gets expressed.

Prediction

Probit/logit of a poll-related injunction indicator on protocol-level β-residual estimated from H3 yields a positive coefficient. As a mechanism pass, an LLM classification of petitions' alleged-dimension fields should cross-tabulate against structural methodology features of the same protocol — i.e., when plaintiffs sue, they should name the dimension on which residual β is loading.

Competing predictions

Statutory channel is too narrow. LE.33.§4 reaches only pesquisa dissociada do registro (Channel B / fabrication); LE.33.§3 is a formal-checklist regime; LE.34.§3 is an audit of planilhas against published numbers (institutions.md § "Court interpretation and poll-related litigation"). None of these reaches declared- methodology slant within a disclosed plano amostral, so Channel A β should be invisible to the adversarial system regardless of how perceptible it is. Under this reading the coefficient is bounded toward zero by design — not by absence of perceived bias.

Race-level selection. Candidates who self-sponsor concentrate in lawsuit-heavy races, but the lawsuits filed in those races may target other polls (the independents the candidate is competing against). The candidate-level coefficient then runs positive mechanically while the poll-level coefficient runs zero or negative.

Prior research

The litigation channel is observed across cycles and races in Brazilian electoral coverage: Russomanno–Datafolha SP 2020 [stories.csv #131; #132], PT–Datafolha Bahia 2022 [stories.csv #078], Chiarelli–Ibope Ribeirão Preto 2008 [stories.csv #089], and Boulos–Datafolha SP 2024 [stories.csv #133]. Each pleading names a specific dimension — plano amostral, sample weighting, subcontracting, registration timeline — consistent with the verifiable-disclosure framing. The procedural framework is documented in institutions.md § "Court interpretation and poll-related litigation". In the 2020 EJ dataset, 2,376 of 19,032 injunctions (~12.5%) are poll-related [summary.md § "Connection to electoral-justice"; projects/electoral-justice/source/clean/proc_2024.py topic classifier].

Evidence

Analysis Bearing Key takeaway
AN-072 Refutes at correct unit Candidate-level OLS with race FE: candidates with any self-sponsored poll are +3.15 pp more likely to be involved in a fraud-flavored PESQUISA case (p = 0.011, base rate 7.9 %). Direction matches the prediction at the candidate side, but the unit is wrong — the hypothesis is about polls, not about candidates.
AN-072v2 Refutes Poll-level OLS within race × week: candidate-sponsored polls are −3.9 pp less likely to be the subject of a fraud-flavored case (p = 0.010, n = 8,943; base rate 4.3 %). Tighter FE drives the coefficient more negative. Sponsored polls survive more legal challenge than peer independents in the same race-week.

Reconciled reading: candidates who self-sponsor operate in lawsuit- heavy races (selection), but the lawsuits target the other polls in those races — independent or media-aligned ones — not the candidate's own. Candidate-paid polls plausibly undergo more legal vetting (registration, defensible methodology) precisely because they are visible; news-outlet polls aligned with one side draw fraud filings from the opposing campaign more often than direct candidate sponsorship does; and "DIVULGACAO FRAUDULENTA" is a publication-side claim that fits media-published polls better than candidate- commissioned ones.

Open tests

Residual β-hat instead of sponsorship indicator

The two AN-072 specs use sponsorship status as treatment, not the protocol-level β-residual the original prediction calls for. The cleaner test is a probit of sued_fraud(p) on the H3-estimated residual β-hat per protocol, holding race × week fixed. Blocked on the H3 protocol-residual extraction.

Plaintiff-side dimension mechanism

LLM classification of petition free-text against structural methodology features in the same protocol. The Channel A statutory gap (institutions.md § "Court interpretation") implies the distribution should pile on formal-disclosure failure (LE.33.§3) and LE.33.§4 fabrication — not on weighting, frame, or wording. Confirming that pattern would substantiate the "no cause of action" reading and discriminate it from the "hard to prove" reading (docs/thinking/poll-lawsuits-as-evidence.md).

Coverage extension

AN-072v2 case→protocol regex recovers only ~9 % of fraud cases (source/intermediate/case_protocols_2024.py). Expanding mov-text coverage or relaxing the regex would shrink the false-zero rate; the estimate is currently biased toward zero, so the true negative magnitude may be larger.

Cross-refs: H3 race-week FE; H10 methodology-flexibility-a; verifiability-decay (Channel B).