AN-097: Type-2 pollster portfolio audit

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.

Hypothesis: H13: Shell-contratante polls show larger residual β
Confidence: yellow
Type: descriptive

Design

Sample: 6 Type-2 pollsters from AN-096, all 76 distinct CNPJs sponsoring polls registered by them in 2024
Specification: For each Type-2 pollster: 1. Pull all sponsor rows in poll_sponsor_2024.parquet with that institute. 2. Apply source/assemble/poll.classify_sponsor_row to each row (CNAE-based MEDIA / POLLSTER detection + name-regex fallback). 3. Aggregate # distinct protocols per sponsor_type. 4. For the other_firm bucket specifically, list the top sponsors with razão_social and capital social — AN-094-style hand classification of whether each looks like a shell.
Comparator: AN-094 PROBABLE_SHELL signature (R$0 capital, MEI individual format, no editorial brand)

Script: source/analysis/an-097-type2-pollster-portfolios.py
Target: build/intermediate/an-097-type2-pollster-portfolios.csv
Status: interpreted · 2026-06-17
Created: 2026-06-17

Question

AN-096 identified six pollsters with the asymmetric bipartite pattern: paired tightly to one AN-094 shell from the shell side (≥80 % of the shell's polls go through this pollster), but operating diverse sponsor portfolios from the pollster side (8–33 distinct sponsors). Two readings of that asymmetry are possible:

Real-pollster + side-channel shell: most of the pollster's volume comes through legitimate channels (real media outlets contracting polls, candidate committees, self-sponsoring), and the AN-094 shell is one of several registration channels used to handle some clients' polls without naming them.
Shell-network pollster: most of the pollster's volume comes through multiple shells, of which the AN-094 entry is just the highest-frequency. The pollster itself is best read as a shell-network operation that happens to have one prominent registered shell partner.

The two predict different volume distributions across the pollster's portfolio.

Results

Portfolio composition

Counts are distinct protocols per sponsor_type (after applying the AN-094 rebuilt classifier). Protocols can have multiple sponsors, so columns can sum > n_polls.

Pollster	Polls	Sponsors	Media	Cand-linked	Self	Pollster-other	Other_firm
IPPI Pesquisas (MA/PI)	107	33	60	17	12	0	18
J J Coelho	51	7	2	0	12	28	9
Instituto Franca (BA/SE)	110	21	67	7	6	0	30
Ivani Mota de Araujo (RO)	60	13	17	0	18	0	25
Severino de Araujo (PB/PE)	41	9	16	0	3	20	2
Tony Brand (MG)	33	12	5	10	0	0	18

Other_firm tail audit (hand classification)

Beyond the AN-094 main pair, are the other_firm sponsors of each pollster additional shells (matching the R$0 capital / MEI individual / no editorial brand signature) or just real small businesses?

IPPI Pesquisas (33 sponsors) — the other_firm tail is mostly party-directorate misses: MDB Piauí state directorate (7 polls, R$0 capital — a real party directorate that the despesa_partidaria join missed), Republicanos Piracuruca municipal (2 polls — same pattern), plus small businesses with normal capital (Made Propaganda R$42k, JB Empreendimentos R$140k, IBS da Costa Construtora R$700k). The 18 other_firm protocols outside Estação I are mostly classifier misses for party-side sponsors, not additional shells.

J J Coelho (7 sponsors) — the other_firm tail is 3 additional MEI individual shells: 41.720.320 Etevaldo Ramos Soares (5 polls, R$3k capital), 50.276.763 Edy Carlos Lourenço de Sousa (3 polls, R$3.5k capital), 40.132.341 Carlos Augusto Monfort (1 poll, R$2k capital). All three match the AN-094 MEI-shell signature exactly (MEI individual format, sub-R$5k capital, no editorial brand). Combined with ABC Publicidade (29 polls), shells account for 38 of 51 J J Coelho polls = 75 %. Self-sponsoring accounts for the other 12. The pollster has essentially no real-media or candidate-committee clients.

Instituto Franca (21 sponsors) — the other_firm tail is dominated by Tres Marias (26 of the 30 other_firm protocols come from Tres Marias alone, the AN-094 shell). The remaining 4 are scattered: H G Empreendimentos (1 poll, R$300k), MMG Construções (1 poll, R$210k), an MEI individual (1 poll, R$100 capital — likely a 4th shell), and one individual-advocacy firm. The majority of the pollster's 110 polls (67) go through real media — IPPP's portfolio is mostly legitimate.

Ivani Mota de Araujo (13 sponsors) — Gledson Lopes dominates the other_firm tail (24 of 25 protocols). The 1 remaining is Enrique Alliana / E Alliana LTDA (1 poll, R$10k capital — borderline). 18 self-sponsoring + 17 media protocols vs 24 shell = roughly even split between legitimate and shell.

Severino de Araujo Alves (9 sponsors) — very small other_firm tail (2 polls from Global Z Consultoria, R$100k — looks legitimate). The 20 pollster_other classifications are mostly Hyago Cavalcante / Loading Marketing (the AN-094 shell, classified pollster_other because it declared 7320300 CNAE). So Severino's portfolio is: real media (16) + Hyago shell (20) + self (3) + 2 small. Closer to "real-pollster + side-channel shell".

Tony Brand (12 sponsors) — Assoc. Marketing MG dominates the other_firm tail (17 of 18 protocols). The 1 other is Aposvale (Vale-pensioners' association, R$0 capital — strange but not obviously shell). Real candidate-linked (10) + real media (5) + Assoc. Marketing shell (17) = the pollster splits its volume across legitimate and shell channels roughly evenly.

Typology

Pollster	Shell volume share	Reading
IPPI Pesquisas	49 % (52/107)	Real pollster + side-channel shell: 60 media + 17 candidate-linked majority
J J Coelho	75 % (38/51)	Shell-network pollster: ABC + 3 MEI shells = majority. Almost no real clients.
Instituto Franca	24 % (26/110)	Real pollster + side-channel shell: 67 media majority
Ivani Mota de Araujo	40 % (24/60)	Mixed: self + media + shell roughly equal
Severino de Araujo	49 % (20/41)	Real pollster + side-channel shell: 16 media + 3 self + Hyago shell
Tony Brand	52 % (17/33)	Mostly shell: 17 Assoc. Marketing + small tail of real candidates and media

Interpretation

Two of six Type-2 pollsters are mostly shell operations (J J Coelho at 75 %, Tony Brand at 52 %). The AN-094 paired shell is the largest single component but not the only one — J J Coelho specifically has at least 4 shells (ABC Publicidade + 3 MEI individuals) accounting for 38 of 51 polls. This is the "shell network" reading: a pollster whose registered sponsorship pipeline is overwhelmingly cover entities, with the AN-094 entry just being the highest-volume one.
Three of six are real-pollster + side-channel shell (IPPI Pesquisas, Instituto Franca, Severino de Araujo). Their majority volume comes through real media outlets and candidate committees; the AN-094 shell is one of several channels they use. The pollster itself has a legitimate editorial business; the shell is a discrete service for clients who don't want their funding traced.
One is mixed (Ivani Mota de Araujo): self-sponsoring + real media + the AN-094 shell, with roughly even volume splits. Could be read either way.
The MEI-individual-shell sub-pattern is a recurring registry signature. J J Coelho's 3 additional shells (Etevaldo Ramos Soares, Edy Carlos Lourenço de Sousa, Carlos Augusto Monfort) all use the NN.NNN.NNN [INDIVIDUAL NAME] MEI format, all have capital R$2k–R$3.5k, and all sponsored polls only with this pollster. This is a distinct operational signature from the AN-094 corporate-shell (LTDA / EIRELI with descriptors like "Publicidade" / "Empreendimentos"). The MEI sub-pattern hides shells one level deeper than corporate shells — they look like individual entrepreneurs in registry data.
The AN-094 top-25 audit understated the total residual-tier shell count. The four additional J J Coelho shells (Etevaldo, Edy Carlos, Carlos Augusto, plus presumably more in the long tail) are below the top-25 cutoff (they have 1–5 polls each, vs the AN-094 cutoff at 17 polls). Auditing the long tail — particularly the 779 singleton sponsors — would surface more MEI shells. AN-094's projected 8–12 % footprint may be a lower bound.

Follow-ups

J J Coelho deep dive (extension): this pollster's 75 % shell-dependence makes it the most interesting Type-2 case. Are its 3 MEI shells the same individual (different CNPJs registered to one operator)? What's its 2020 cycle structure — does the Publi. QC / IPOP self → shell pattern apply? Suggested script: source/analysis/an-097a-j-j-coelho.py.
MEI-shell pattern search across the long tail (extension): scan all 1,210 other_firm CNPJs for the MEI signature (8-digit prefix + individual name in razão social + capital < R$10k). Estimate the share of singleton residual-tier sponsors that are MEI shells. Suggested script: source/analysis/an-097b-mei-shell-tail.py.
Party-directorate classifier fix (infrastructure): IPPI Pesquisas's tail revealed that the despesa_partidaria-based Route C miss includes MDB Piauí state directorate (7 polls), Republicanos Piracuruca municipal (2 polls). These are real party-funded polls being classified as other_firm. The classifier could be augmented with name-regex hits for known party abbreviations + "ESTADUAL/MUNICIPAL" tokens.

Caveats

The Type-2 set was defined as "shell side tight, pollster side loose". The 4 Type-4 shells (both-moderate-or-diffuse from AN-096) probably have different portfolio structures that the typology doesn't predict. A Type-4 audit is a separate exercise.
Pollster-self sponsoring (where the pollster itself registers as contratante on its own polls) is counted as "legitimate" here but is actually the IPOP/Publi. QC 2020 pattern — disclosure failure under a different name. So a high pollster_self share could itself be a different kind of shell pattern. Two of the Type-2 pollsters (IPPI 12, Ivani Mota 18) have non-trivial self-sponsoring; worth flagging.
The other_firm hand-classification here is light-touch (CNPJ data + capital + name pattern) and doesn't include web verification.