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.

User's intuition confirmed — the "media" bucket is heterogeneous. Within media-sponsored polls, trusted firms (DATAFOLHA, QUAEST, PARANÁ PESQUISAS, REAL TIME MÍDIA, VERITA — n=278) understate the margin by 3.16 pp LESS than other firms in the same race (p=0.002, race FE); major-media sponsors (GLOBO, FOLHA, ESTADÃO, large regional outlets — n=304) have mean |error| 1.0–1.4 pp LOWER than blog/small-digital sponsors. Trusted-firm self-contracted polls are the MOST accurate slice of the entire dataset (mean |error| 5.2 pp vs universe 8.1 pp). Universe-wide spec with both trusted-firm and bucket dummies: is_trusted_firm β = −1.15 on |error|, −2.02 on margin, +0.10 on calls-winner (all p < 0.002, race FE). The other_firm margin-error effect SURVIVES after controlling for firm tier (β=−1.95, p=0.002).

Trusted-firm advantage survives 3 of 4 alternative definitions for the two simplest outcomes (calls_winner_first +10 pp, mean |error| −0.9 to −1.7 pp) — robust to hand-picked, top-10-UF-spread, or low-|β| definitions. Volume-based (top-10 by # polls) is the OUTLIER and a bad trust marker — it picks up large state-level specialist firms with worse accuracy. The trusted-firm margin_error advantage (−2.02 pp in AN-085) is fragile: only survives under the hand-picked definition. CRUCIALLY: the bucket coefficients (is_other_firm β=−1.83 to −1.95 pp on margin_error; is_candidate β=−2.42 to −2.61) are stable across all 4 trusted-firm definitions — the shell-bucket finding doesn't depend on how we define trust.

Paper-ready 4-spec × 3-outcome × 3-bucket appendix table consolidating AN-082 / AN-085 / AN-089. Three readings emerge clearly: (a) `other_firm` is the most-robust sponsor-bucket signal — significant on margin_error in 3 of 4 FE specs (S0 −3.01***, S1 −1.87***, S2 −2.65***, S3 −0.91); collapses to ns only under firm + race FE jointly. (b) `candidate` margin_error coefficient FLIPS sign across specs (raw +0.02 → race FE −2.57*** → firm FE +1.01 → both FE −1.76*) — the sponsor effect is entangled with which races candidates commission polls in. (c) `mean_abs_error` is null across specs except raw `candidate` (+1.07***), absorbed by either FE. The headline message for the paper appendix is "race + firm sorting accounts for most of the cross-sectional sponsor differences; the residual within-firm-within-race slant is small but directionally consistent."

Adding log_sample_size as a control barely moves the sponsor-bucket coefficients under any FE-controlled spec (S1, S2, S3). The largest shift is under S0 (no FE) where log_sample_size soaks up race-level difficulty composition. Race FE absorbs ~95% of the sample-size variation. Net: the AN-090 spec ladder is robust to sample-size control. log_sample_size coefficient is −6.4 pp/log-unit on margin_error in S0 (p<0.001) but only −1.3 (ns) in S1 (race FE), confirming that sample size is mostly proxying for race difficulty (big races attract big polls).

Final paper appendix table. Five buckets (major_media reference, small_media, candidate, pollster_self, other_firm) × four FE specs × three accuracy outcomes, all with log_sample_size control. Under race FE (S1): candidate is 2.69 pp more accurate on margin than major-media (p<0.05); other_firm 2.00 pp more accurate (p<0.10). Under firm + race FE (S3): all four non-major-media buckets show 1.8 to 3.8 pp lower margin error than major-media within firm × race — but the major-media reference is thin (n=304, 4.3% of sample), so S3 selection is sharp and the magnitudes should be read with caution. S1 is the more interpretable spec.

VS Publicidade LTDA (the top-ranked AN-094 shell) routes 218 of its 254 mayoral polls (86%) through Publi. QC Pesquisas & Editoração — and Publi. QC self-contracted all 230 of its 2020 mayoral polls in São Paulo without an intermediary. The Goiás IPOP/FacUnicamps cycle-over-cycle shift recurs at this scale in São Paulo with a different cover entity. Owners are different people (Sanazar brothers for VS, Flávio Henrique da Silva for Publi. QC), so the relationship is contractual exclusivity rather than common ownership.

From the shell side, the dominant pattern is tight 1-to-1 with one pollster (8 of 14 shells route ≥80% of their polls through a single pollster; median 85%). From the pollster side, only 2 of those 8 pairings are reciprocally tight (VS Publicidade ↔ Publi. QC, FacUnicamps ↔ IPOP) — the other 6 pollsters operate diverse sponsor portfolios with 8–34 distinct sponsors per pollster, using the shell as one of several channels. One shell (Nivaldo Galindo) has the inverse asymmetry: shell uses 2 pollsters, pollster is 93% captive to the shell. No shell operates across many pollsters: there is no intermediary or partisan-aggregator pattern in the data.

|error| degradation on sponsored rows is concentrated in the firm tails. T3 (highest pro-sponsor β, 7 firms, 310 polls): +6.82 pp on |error| under race + cand FE (p<0.05). T1 (low or anti-sponsor β, includes CENSUS and EVA FRANCIELI): +3.54 pp (p<0.10) — slant in either direction inflates |error|. T2 (moderate β): null. Untested firms (no AN-016 β, 19,337 cand-poll rows, 80% of universe): −2.14 pp (p<0.10) on |error| but +5.73 pp (p<0.001) on SIGNED error — the bias is real and large but doesn't translate into measurable accuracy degradation. This is the cleanest demonstration that the market for poll accuracy cannot discipline pollsters: 90% of the universe of polls shows null-or-negative |error| degradation on sponsored rows even when the +7 pp slant is precisely measured.

σ_within(signed error) at the (cand × race) cell = 7.81 pp on the unsponsored sample. The +7 pp slant is 0.90σ of natural noise — sub-1σ. By folded-normal mapping, a +7 pp signed shift translates to a +2.35 pp |error| shift (matched well by the +3.42 pp empirical raw difference). 67.5% of error variance is between (cand × race) cells; only 32.5% is within. Only 47% of cells where a single sponsored poll's +7 pp slant would be statistically detectable at p<0.05 against the within-cell SD. Quantifies the "noise floor": the slant is real but smaller than typical poll-to-poll variation, so the market for poll accuracy cannot discipline pollsters.

Consensus deviation (poll share vs median of other polls of same cand in same race within ±14 days) confirms the +7 pp signed slant: +5.82 to +6.59 pp under race + cand + firm FE (p<0.001) using unsponsored-only consensus pool. This separates the slant from "late campaign movement" noise (vs-final-results contains both). The |deviation| magnitude effect under race FE alone is +3.31 pp (vs +1.35 pp using |error|) — 2.5× tighter signal at the race-FE level. But under cand FE the magnitude effect still collapses to null — the natural per-candidate consensus-deviation variance is large enough to mask the slant. Net: consensus-deviation is a better real-time bias detector than vs-final-results, but the fundamental noise-floor argument (AN-098) is unchanged.

Sponsor-blind detection of slanted polls is feasible but imperfect. Within-poll features computed from consensus deviation — signed_spike (max minus 2nd max cand-row deviation), poll_std_dev, mean_abs_dev — distinguish candidate-sponsored from unsponsored polls with AUC ≈ 0.64–0.69. The signed_spike best single feature: 17.2 pp for sponsored vs 10.1 pp for unsponsored, AUC = 0.636. Combined multivariate logit AUC = 0.685. Concrete policy mechanism: a publicly-computable "suspicion score" the TSE / journalists / regulators could use to triage polls for closer audit, requiring no sponsor identity information.

Out-of-sample predicted bias (from 5-fold CV gradient boosting trained on AN-100 sponsor-blind features → poll_has_candidate_sponsor label) shows other_firm polls have +1.6 to +2.4 pp higher predicted-bias probability than major-media polls under S0/S1/S2 (p<0.10 to p<0.01). The classifier was never trained on the other_firm label; it only learned the candidate-sponsorship pattern. The fact that other_firm polls light up as positively biased is independent algorithmic evidence for shell-style slant. The signal collapses under joint S3 (firm + race FE) — same composition story as the |error| version. AUC of out-of-sample classification = 0.69 (matches AN-100). Provides a publicly distributable per-poll suspicion score at build/analysis/poll_suspicion_score.parquet.

Modern ML pipeline (XGBoost/LightGBM, 37 features, 5-fold CV) achieves OOS AUC 0.91 for detecting candidate-sponsored polls — up from 0.69 (AN-101 sponsor-blind features only) and 0.69 (AN-100 within-poll only). Big jump comes from firm-level aggregates (firm_share_candidate_work is the #1 feature at 0.19 importance; firm_has_beta_estimate at 0.16). Gradient boosting (LightGBM 0.911, XGBoost 0.911) modestly beats logistic regression (0.901) and random forest (0.905). Within-poll consensus-deviation features still rank in the top 10 (error_concentration, poll_std_dev, mean_abs_dev) — the within-poll fingerprint is real. Other_firm bucket (residual, no shells) has predicted-bias coefficient +0.018 (p<0.01) under firm FE — sponsored-like pattern persists even with the powerful classifier. Shell bucket has −0.072 (p<0.01) at S0, null under FE — at AUC 0.91 shells STILL don't light up, sharpening the 'professional evasion' interpretation.

Strict-blind detection ceiling (no firm identity, no state, no municipality, no race aggregates — only poll results + election results + peer polls + dates + log_sample) achieves OOS AUC 0.742 with XGBoost on 28 features. Firm identity adds 0.17 AUC (AN-103's 0.911 ceiling). The strict-blind classifier confirms two structural findings: (i) is_other_firm survives joint firm + race FE on the spec ladder (+0.013, p<0.05) — the unaudited shell-like tail robustly shows a sponsored-poll fingerprint that does NOT rely on knowing the firm; (ii) is_shell is significantly NEGATIVE at S0 (−0.055\\*\\*\\*) and S1 (−0.023\\*\\*\\*) — the AN-094 identified shells defeat blind detection regardless of classifier sophistication, hardening the 'professional evasion' interpretation. Top features: log_sample, n_peers, poll_std_dev, max_abs_dev, days_to_election_sq — consensus-deviation features collectively dominate.

User's concern (race-proxy leak in AN-104) confirmed. Mode A (7 theoretical slant signals, no race-attention features) gives OOS AUC 0.614 — barely above chance. Mode B (28 AN-104 features demeaned within race) gives 0.693 — basically the AN-101 baseline. The 'genuine within-poll slant signature' detection ceiling is ~0.69 (within-race signal only); AN-104's 0.742 included ~0.05 AUC worth of race-attention proxy leak. Critical change: under within-race demeaning, is_other_firm survives only weakly (+0.006 ns at S3) — the 'shell-like tail signal' in AN-104 was partly race-proxy. Shells light up POSITIVELY under pure theoretical features (+0.016*** at S0 of Mode A) — opposite sign from AN-104's negative — because race-attention features no longer absorb shell variation. Honest §Policy framing: blind detection ceiling is 0.69, well below earlier claims; the genuine within-poll slant fingerprint is weak (~0.61 with theoretical features only).

Double-machine-learning (race-confounder residualization via cross-fitted XGBoost regression, then XGBoost classifier on residuals) achieves OOS AUC 0.717 — sits between AN-105 Mode B linear demeaning (0.693) and AN-104 raw (0.742). Non-linear race effects worth +0.024 AUC vs linear demean; remaining +0.025 to AN-104 raw is race-correlated signal that DML cannot remove (race confounders are imperfect proxies). R² of residualization shows the race-leakiest features are log_sample (0.825), n_peers (0.775), error_concentration (0.693), n_cand_rows (0.690) — those were carrying mostly race signal. Spec ladder shows is_shell coefficient flips back to NEGATIVE at S0 (−0.015\\*\\*) under DML — substantively cleanest 'shells mimic media' finding. is_other_firm at S1 (+0.008\\*) borderline positive but weakens further at S3 (+0.007 ns).

Magnitude measures (mean |error|, RMSE, max boost, spread) come back NULL across all four sponsor buckets within race × week. The signal is rank-disagreement, not noise. other_firm polls understate the eventual race winner by −1.83 pp vs media polls (p = 0.010) and call the winner #1 4.8 pp less often (p = 0.15); pollster_self polls call the winner +7.4 pp MORE often within race × week (p = 0.009). Goiás interpretation: 2020 IPOP scheme was pollster_self (357 polls); 2024 IPOP+Alcateia migrated to FacUnicamps shell, which lands in other_firm.

Within less-trusted firms (n=5,097, 419 firms) holding firm + race fixed: candidate-sponsored polls reduce margin_error by 1.94 pp vs media-sponsored (p=0.07); other_firm-sponsored reduce by 1.04 pp (p=0.11). Directionally consistent with the cross-sectional finding at ~2/3 magnitude. Same SIGN, smaller magnitude — most of the cross-sectional shell signal is CROSS-FIRM composition (different sponsors hire different firms), not within-firm slant. AN-088 already showed sponsor contributes only 0.1-0.2 % marginal R² beyond firm + race; this AN-089 result is the within-firm correlate of that. NOTE: in trusted firms (n=430, 5 firms) sponsor effects are statistically weak overall but point estimates on margin_error are LARGER (−9.4 / −8.8 pp at the firm-only spec, p≈0.07), suggesting trust × sponsor interaction worth flagging but with very thin n.

14 of the top-25 other_firm sponsors (668 of 914 polls, 73%) are probable shells with no journalism brand and no plausible non-poll business reason to commission 17–254 mayoral polls; 9 are real media misclassified by the regex (211 polls); 2 unclear (35 polls). The top single shell signal is VS Publicidade (254 polls, MS+SP, R$0 capital, no web footprint).

The 6 Type-2 pollsters (paired tightly to an AN-094 shell but with diverse sponsor portfolios on their side) split into two sub-types. Three are 'real-pollster + 1 shell side channel' (IPPI Pesquisas, Instituto Franca, Severino de Araujo — majority volume through media + parties + self). Two are 'shell-network pollster' where 50–75% of total volume goes through shell sponsors (J J Coelho 73% shell, Tony Brand 52% shell). One is mixed (Ivani Mota de Araujo). J J Coelho's network includes 3 additional MEI-individual shells beyond the AN-094 ABC Publicidade — so the pollster is a shell hub, not a real media operation.

Corrected headline table with 4 sponsor buckets + shell as separate category (per AN-094 audit of 14 probable shells, 229 polls / 668 cand-poll rows in analysis sample). |error| at cand-poll level: shell coefficient is null across all FE specs (+0.65 raw, ns; collapses under any FE). Predicted bias (GBM, OOS) at poll level: shell coefficient also null (~0.00 across specs). The non-identified other_firm tier (other_firm excluding 14 shells) shows persistent positive predicted bias (+0.024 raw p<0.01, +0.013 firm FE p<0.10), suggesting the within-poll fingerprint of shell-style slant lives in the UNAUDITED residual rather than in the top-25-identified shells. Two readings: (i) AN-094 shells are professional operators producing media-pattern polls to evade detection; (ii) 229 polls is underpowered for shell-bucket inference. GBM AUC = 0.673 OOS.

Of 14,887 mayoral 2024 polls, **12.4% are routed through cover vehicles** (shell CNPJs 7.7% + MEI-individual 4.7%) that obscure the candidate's connection; 84.9% are administratively recoverable as candidate-linked, media, or pollster-self; the remaining 2.8% are an *uncoded residual* — low-volume CNPJs (1–4 polls each) that the classifier cannot place, many of which are likely sub-threshold cover vehicles (small publicity firms, missed MEI, missed local media). The cover-vehicle share GREW from 3.8% in 2020 to 12.4% in 2024 — a 3.3× increase; the uncoded residual SHRANK from 6.4% to 2.8%, consistent with cover-vehicle activity consolidating into identifiable shell/MEI patterns. 15 pollster firms switched from pollster_self in 2020 to shell or MEI-individual in 2024 — the IPOP pattern at scale, covering 508 polls in 2024 alone. Universe-extended shell list: 89 CNPJs / 1,149 polls (vs AN-094 top-25 audit: 14 CNPJs / 668 polls). Calibration: rule recovers 9 of the 13 AN-094 PROBABLE_SHELLs present in 2024 (69% recall); the 4 misses are firms the source/assemble/poll.py CNAE upgrade promotes to media or pollster_other, so the 89-CNPJ / 1,149-poll count is a precision-favoring FLOOR on the universe shell footprint.

AN-102's null finding on shell-β is robust to the 2.6× sample expansion. With shell-CNPJ list extended from AN-094's 14 (n=620 cand-poll rows) to AN-121's 89 (n=1,612 rows), shell-β on |error| remains statistically indistinguishable from zero across every FE spec: S0 +0.03 (0.34), S1 −0.39 (0.36), S2 −0.29 (0.31), S3 −0.46 (0.31). Point estimates drift slightly negative as the SE tightens, but the t-statistic stays under 2 everywhere. AN-102's interpretation — 'cover-vehicle polls look like media polls in their per-poll error distribution' — survives. Implication: the cover-vehicle category does its work at the IDENTIFICATION margin (we can't link them to a candidate), not at a measurement margin (they don't produce detectably noisier polls); this is the iceberg-framing claim, not an additional mechanism claim.