AN-083: IPOP makes no observable predictions

IPOP and Alcateia (the two firms named in paper.tex §Setting as 2024 Goiás shell operators) systematically register polls without publishing usable relatórios. Alcateia uploaded 0 of 41 relatório PDFs to TSE; IPOP uploaded 17 of 68 (25 %), and ALL 17 contain methodology + demographic tables but ZERO vote-intention data. Mainstream comparator INSTITUTO GAZETA uploaded 121 of 145 (83 %), and the 8 we LLM-extracted have 100 % vote-intention coverage (3.4 scenarios × 21.6 cands each on average). The accuracy comparison the user asked for cannot be computed — IPOP's polls produce no observable predictions in the disclosure regime.

Hypothesis
H13: Shell-contratante polls show larger residual β
Confidence
green
Type
descriptive
Design
Sample
68 IPOP + 41 Alcateia + 145 INSTITUTO GAZETA polls registered in GO 2024
Specification
Three-stage funnel: (1) registry-based count of polls per firm from pipelines/politica/build/clean/poll_2024.parquet; (2) PDF availability via streaming inspection of GO.tar.zst on bi-dropbox-ro:data/TSE/2024/pesquisa_eleitoral/relatorios/; (3) vote-intention disclosure via LLM extraction on the 17 IPOP + 8 INSTITUTO GAZETA PDFs available.
Comparator
GO baseline + INSTITUTO GAZETA (in-state mainstream firm)
Script
source/analysis/an-083-ipop-relatorio-content.py
Target
build/table/an-083-ipop-relatorio-content.csv
Status
interpreted · 2026-06-17
Created
2026-06-17

User question (2026-06-17): "Can we check if IPOP made worse predictions compared to other pollsters or compared to polls sponsored by media?"

Answer: The standard accuracy comparison can't be computed for IPOP because IPOP's polls produce no observable predictions in the TSE disclosure regime. This is itself the finding.

Results

Disclosure funnel — three named firms vs GO universe

Firm Registered PDFs on bi-dropbox PDFs with vote intentions
GO universe 1,305 1,031 (79 %) —
INSTITUTO GAZETA (mainstream control) 145 121 (83 %) 8 of 8 sampled (100 %)
IPOP CIDADES E NEGOCIOS 68 17 (25 %) 0 of 17 (0 %)
ALCATEIA OUTSOURCING 41 0 (0 %) — (none to extract)

What's actually in IPOP's 17 available PDFs

Hand inspection of GO016032024 (São Luís de Montes Belos): 1,515 lines of methodology + sample stratification tables (sex, age, income, education repeated across geographic strata) and sponsor data (FacUnicamps, R$ 2,500). Zero vote intention tables. Zero candidate names. The "Prefeito" field is the only string matching voto|intenç[aã]o|candidato|prefeit|estimulado|espont[âa]neo in the document, and it's the cargo-field label on the registration card at top.

Quantitative confirmation across all 17 IPOP PDFs: vote-related string matches range 1–1 (exactly 1 per PDF, the cargo label). Mainstream control (INSTITUTO GAZETA) on 8 sampled PDFs: 12–53 matches per PDF, mean 26.

LLM extraction confirmation

GPT-4o-mini run on all 17 IPOP PDFs returns scenarios: [] for every single one, with extraction notes saying: "The PDF does not contain any vote intention results for candidates. Only methodological details and sample demographics are provided."

Same extractor on the 8 INSTITUTO GAZETA PDFs returns mean 3.4 scenarios × mean 21.6 candidates per scenario — uniformly populated.

Interpretation

Three nested ways the firm evades observability:

  1. Sponsor channel. FacUnicamps (a private faculdade) is the contratante and pagante for every IPOP 2024 poll, decoupling the registration footprint from any candidate / party CNPJ that the project's Routes A–D can trace.
  2. Relatório upload. Only 25 % of IPOP's registered polls have a PDF uploaded to bi-dropbox at all (vs 79 % universe baseline, 83 % for the mainstream GO comparator). Three- quarters never enter the relatório system.
  3. Relatório content. The 17 PDFs IPOP did upload are methodology + demographic tables only. None contain vote intention results. The disclosure regime has no record of what IPOP claimed about candidates in those races.

The three layers compound. Any one of these defeats the project's accuracy infrastructure independently:

After all three filters, the project's cand_poll.parquet contains 0 / 68 IPOP 2024 polls. We cannot compute IPOP's mean |error|, leader-rank error, or anything that requires knowing what the poll claimed.

Alcateia is the same case, only stricter. Of its 41 registered polls, 0 PDFs are uploaded. No content audit needed to conclude the firm is unobservable.

What this implies for the headline of AN-082

AN-082 found other_firm polls understate the eventual winner by −1.83 pp within race × week (p = 0.010). That regression is on the 1,553 other_firm polls in the project's sample — and IPOP + Alcateia together contribute zero of those 1,553 protocols. The AN-082 signal is measuring the visible 21 % of the shell pattern; the IPOP-Alcateia core of the shell pattern is literally invisible to the regression. Whether the full shell universe has β at the AN-082 magnitude, double it, or zero, the sample we observed cannot tell us.

Implications for the iceberg framing

The paper's iceberg framing (memory: "Both evade Routes A–D. Load-bearing §2 example for the iceberg framing.") is reinforced by this finding but extended: the iceberg has not three but three nested layers of evasion, of which Routes A–D's tracing failure is only the first. The deeper two — relatório non-upload and relatório content-emptiness — are not addressed anywhere in the existing literature on Brazilian poll disclosure to my knowledge. The disclosure regime nominally requires the relatório; in practice the enforcement gap is content-level, not just registration-level.

Caveats

Follow-ups

  1. Cross-check Alcateia by pulling its scrape directly from the TSE divulgação portal. bi-dropbox's 0 / 41 might be a scrape gap; the TSE portal itself might have the PDFs. If the TSE portal also has 0, the finding is that Alcateia evades upload entirely. If TSE has them and bi-dropbox missed them, we need to re-scrape and re-audit.
  2. Audit the 274 GO polls (≈21 %) with no PDF on bi-dropbox by firm. Is the rest of the missingness concentrated in a handful of firms with the same shell pattern, or is it evenly distributed across the GO pollster universe? If concentrated, expand the AN-083 frame to a list of "non-disclosing firms."
  3. Replicate on 2020 IPOP. The 2020 Operação Leão de Neméia case had 357 polls. Do the 2020 IPOP polls show the same relatório-content-emptiness pattern, or did the firm upload real (fabricated) numbers in 2020 and switch to content- emptiness after the prosecution? The 2020 vs 2024 comparison would tell us whether content-emptiness is a defensive adaptation specifically in response to enforcement.
  4. Generalize beyond Goiás. Are there firms in other UFs with similar disclosure-failure patterns? An automated audit running the same funnel (registered → PDF on bi-dropbox → vote-intention extraction rate per firm) on the full universe would surface comparable cases without manual curation.
  5. Wire to the paper. Currently §Setting names the IPOP 2024 case. AN-083 turns "FacUnicamps as cover sponsor" into a three-layer disclosure-evasion finding with hard numbers: 25 % PDF upload, 0 % vote-intention content. Worth a half-page in §Setting with the three-layer table.

Artifacts