Institutional Background

Institutions

Shared institutional background lives in brazil-institutions (research/institutions/brazil/). Most relevant topic files for this project:

This file inlines only the project-specific institutional facts that ground the identification strategy.

TSE PesqEle / Divulgação Regime

Overview

Statutory mapping — registration fields → dataset columns

LE.33 itemizes the seven fields every pollster must register. Each clause maps to a column the project's analysis pipeline relies on. This mapping is load-bearing: it explains why the structured CSVs already carry sponsor identity (no PDF extraction needed) and why the four free-text blocks are precisely the levers of Channel A.

Statutory clause What it requires Data column / extraction target
LE.33.I Quem contratou a pesquisa NR_CPF_CNPJ_CONTRATANTE (contratantes CSV)
LE.33.II Valor e origem dos recursos VR_PAGO_CONTRATANTE, DS_ORIGEM_RECURSO (Fundo Partidário / Doações Eleitorais / Recursos Próprios / Outros)
LE.33.III Metodologia e período DS_METODOLOGIA_PESQUISA (free text → LLM), DT_INICIO_PESQUISA, DT_FIM_PESQUISA
LE.33.IV Plano amostral e ponderação (sexo, idade, grau de instrução, nível econômico, área física) DS_PLANO_AMOSTRAL (free text → LLM — the Channel A signal)
LE.33.V Sistema interno de controle e verificação DS_SISTEMA_CONTROLE (free text → LLM)
LE.33.VI Questionário completo Attached to relatório PDF
LE.33.VII Nome de quem pagou + nota fiscal NR_CPF_CNPJ_PAGANTE (pagantes CSV) — the contratante/pagante split

The statute names exactly five quota dimensions in LE.33.IV (sex, age, education, income, area). It does not mandate any of them, only that the plano amostral disclose its weighting on them. That is the textbook legal opening for Channel A: pollster picks which of the five to actually quota on, and which to omit.

Compliance and sanctions

What is legally challengeable — TSE jurisprudence

The statute provides three operational hooks; none reaches into methodology choice within a disclosed plano amostral. This is load-bearing for how we interpret the lawsuit corpus.

  1. Formal-disclosure failure (LE.33.§3, administrative multa). Standing test is checklist absence: did the registration name and quantify each of the seven LE.33 incisos? TSE has affirmed regional suspensions on the ground that the plano amostral "não indica ponderação atinente ao nível econômico" — substantive bias evaluation is not part of the test, only formal completeness of the registered form (TSE Temas Selecionados → Divulgação).
  2. Pesquisa fraudulenta crime (LE.33.§4, detenção 6mo–1yr + multa). Narrowly defined by TSE as "pesquisa dissociada do registro obtido — quando os responsáveis divulgam pesquisa diversa daquela registrada". The crime "somente se aperfeiçoa se a fraude ocorrer sobre dados coletados em observância às exigências legais" (TSE Temas Selecionados → Penalidade). In our taxonomy this is Channel B (fabrication / divergence from registered methodology). Channel A (declared-methodology slant) is outside §4 entirely.
  3. Party audit (LE.34.§1) + irregularidade nos dados publicados (LE.34.§3). Parties can confront published numbers against planilhas individuais and entrevistador identifiers via requerimento à Justiça Eleitoral; §3 imposes detenção + retificação obrigatória "no mesmo espaço, local, horário, página, caracteres e outros elementos de destaque". The procedural ceiling is "did the planilhas back the published numbers" — a verification audit, not a methodology audit. Obstrução do audit is itself a crime (LE.34.§2).

Cited precedent. TRE-RN extracts in its Temas Selecionados → AIME → Hipóteses de Cabimento digest the only acórdão specifically labeling a poll fraudulenta as AIME grounds: RE 1-20/2014, Rel. Francisco Eduardo Guimarães Farias — where candidates contracted and procured divulgação of a poll in violation of a prior judicial order. The finding is procedural- fraud, not methodological (TRE-RN digest).

Implication for this project's identification. Lawsuit exposure is NOT a valid perceived-bias proxy for Channel A. The regime is structurally incapable of reaching methodology choice within disclosed plano amostral. Empirical evidence is consistent with this design: candidate-sponsored polls draw less fraud-flavored litigation than peer independents within race × week (AN-072v2, −3.9 pp, p=0.010). The full reasoning is in docs/thinking/poll-lawsuits-as-evidence.md § "Disentangling 'hard to prove' from 'no cause of action'". For paper §sec:policy, the takeaway is that adversarial litigation cannot discipline Channel A by design — mandatory disclosure is the only lever consistent with the statutory scope.

Sub-bucket misclassification. A small share of the compliance-bucket cases are not actually poll cases — the PESQUISA ELEITORAL assunto_desc bundle catches propaganda- antecipada and related claims that touched poll data. The fraud-bucket count is cleaner. Quantification of the misclassification rate is a PROJECT-SPECIFIC — needs manual curation: estimate from a small sample.

Mayoral race rules (why the study scopes to prefeito)

Single-candidate-per-party rule (load-bearing for identification). Brazilian mayoral races are majoritário: each party fields exactly one chapa (prefeito + vice-prefeito) per municipality. There is no intra-party slate as in proportional races. This is what makes the CNPJ→party→prefeito map 1:1 within a race: a poll commissioned by Party X's CNPJ in city C unambiguously identifies Party X's prefeito candidate as the boosted party. Extending the design to vereador / deputado is not just bigger N — it's a different identification problem (within-party-slate ambiguity), which is why the project scopes to mayoral.

For the majoritário/proporcional split and the seat-allocation mechanics for non-prefeito races, see partidos-e-sistema-eleitoral.md in the shared reference.

Two-round rule. Per CF.29.II (and LE.3), second round applies only in municipalities with > 200k voters. ~95% of mayoral races settle in round 1. Relevant for the timing controls: the "final vote share" benchmark for the bias error is the decisive round. Restricting to 1-round munis simplifies; including >200k munis requires matching the poll's implied scenario (1st vs 2nd round estimulado) to the right benchmark.

Candidate registration deadline. Per LE.11, parties register candidates by 19h on August 15 of the election year. Polls fielded before this date may name candidates who never formally register — a small data-cleaning risk for the name-match step.

Brazilian polling industry

Major firms

Industry context

Statistician registration (CONRE)

Quota sampling — universal