Stage 1 Cross-Encoder · Full-Population Audit

Stage 1 Full Audit — Mismatches for Manual Review

All 4,188 stage1_confirmed matches reviewed against the same Ollama reviewer (mistral:latest) the production pipeline uses. 99.6% agreed. This report covers the 0.4% that didn't — with full data and, for the largest cluster, a traced root cause.

Date 2026-07-01/02
Reviewer mistral:latest (Ollama)
Population 4,188 stage1_confirmed rows (trademark_holder → company)
01

Headline numbers

Total audited
4,188
every stage1_confirmed row, not a sample
Agreement rate
99.6%
4,170 agree — consistent with the earlier 300-sample 99.7%
Disagreements
18
0.4% of audited population
Already resolved
1
Citrus CBD case — human-confirmed false positive, see §4
Open for review
17
11 share one root cause (§2), 6 scattered (§3)
Errors
0
clean run, no Ollama failures
02

Headline finding: a traced, systemic LEI parent-walk bug

11 of the 17 open disagreements are all "[Brazilian distributor name] ↔ Mars, Incorporated" — not scattered noise, one root cause. All 18 rows matching this shape (11 disagreed, 7 happened to agree with Ollama by coincidence) trace back to match_method='lei' at the exact same confidence, 0.88.

Root cause, traced end to end

companies.registration_number for "Mars, Incorporated" holds its real GLEIF LEI code (54930073HKEO6GZBSS03). GLEIF's lei_relationships correctly lists dozens of Mars's real global subsidiaries as children — including a real Brazilian pet-food subsidiary, "ROYAL CANIN DISTRIBUIDORA LTDA". The LEI parent-walk matcher (src/match/lei.js, hasSharedDistinctiveToken) fuzzy-matches a source holder name against GLEIF records and requires at least one shared token that isn't in GENERIC_CORPORATE_TOKENS before walking up to the parent. "distribuidora" and "ltda" — Portuguese for "distributor" and the Brazilian "Ltda." legal form — aren't in that list, which is English-centric (plus a handful of geography terms added by prior fixes). So any Brazilian holder whose name happens to contain "Distribuidora ... Ltda" fuzzy-matches "Royal Canin Distribuidora Ltda" on those two generic words alone, walks up to its ultimate parent, and gets proposed as a 0.88-confidence "Mars, Incorporated" candidate — regardless of what the rest of the name actually says.

This is the same class of bug this project has already fixed twice before (see project-pipeline memory / STATE_AND_CHANGELOG.md §6c, §6e — the Nestlé R&D cluster and the Taiwan/Japan/Korea geographic-token clusters). Both were fixed the same way: extend GENERIC_CORPORATE_TOKENS. This one slipped through because that list has never had non-English generic business terms — "distribuidora" (distributor), "ltda" (Ltd.), and likely "comercio"/"industria"/"atacado" (commerce/industry/wholesale) would all cause the same failure mode for Portuguese/Spanish-language holders.
Recommended fix (not applied in this report — flagging for a follow-up): add 'distribuidora', 'ltda', and similar Portuguese/Spanish generic corporate/ business-descriptor tokens to GENERIC_CORPORATE_TOKENS in src/match/lei.js. Low-risk, well-precedented, and would prevent this exact cluster (and likely similar undiscovered ones) from recurring on future rescans.

The 11-row cluster

TM HolderLocationStage 1Ollamareview_queue
PARATY ATACADO E DISTRIBUIDORA LTDA BR 0.9949 0 #274
PRIME DISTRIBUIDORA DE AÇAI LTDA BR 0.9936 0 #281
DISTRIBUIDORA 3 AMIGOS LTDA BR 0.9930 0 #269
AÇAI BENZADEUS DISTRIBUIDORA LTDA BR 0.9930 0 #271
DISTRIBUIDORA CIDADE VERDE LTDA. BR 0.9929 0 #265
DISTRIBUIDORA TOCANTINS LTDA BR 0.9927 0 #279
SCARDUA DISTRIBUIDORA LTDA BR 0.9923 0 #272
VIA APIA DISTRIBUIDORA LTDA. BR 0.9919 0 #273
VITA ANGRA DISTRIBUIDORA LTDA BR 0.9917 0 #278
MULTIGIRO DISTRIBUIDORA LTDA BR 0.9916 0 #280
QUEIROZ DISTRIBUIDORA LTDA BR 0.9915 0 #275

All 11 target the same company: "Mars, Incorporated" (id 7865). Ollama's reasoning was sound in every case ("only the generic word 'Company'/'Inc.' overlaps") — unlike the Citrus CBD case in §4, these rejections look like genuine catches, not lucky guesses.

03

Remaining 6 — scattered, individually reviewed

These don't share a common root cause — each needs independent judgment. Full data below for manual review.

B. Superfoods BV ↔ MB Superfoods DISAGREE

TM Holder

B. Superfoods BV
Nassaukade 147-F
Amsterdam NL

Company

MB Superfoods
Daelwijcklaan, 172, Utrecht, Utrecht 3554 HL, NL
NL
Aliases: mb superfoods
Sectors: Smoking Alternatives

Stage 1 score0.9992 Original methodfuzzy (0.945)
Ollama score0 review_queue#266 (pending)

Ollama's reasoning: "The distinctive core name 'Superfoods' does not fully overlap, and there is a city mismatch in the addresses."

Aura Medical LTD ↔ Aura Medical LLC DISAGREE

TM Holder

Aura Medical LTD
128 City Road London EC1Y 2NX
GB

Company

Aura Medical LLC
34 Franklin Avenue Suite 312 Brooklyn, NY 11205

Aliases: aura medical
Sectors: Manufacturing, R&D, Smoking Alternatives

Stage 1 score0.9991 Original methodexact (0.700)
Ollama score0 review_queue#276 (pending)

Ollama's reasoning: "The distinctive core name does not match (Aura Medical LTD vs Aura Medical LLC). Address signals are weak evidence against a match due to postal code difference."

ALDI ↔ Aldi, Inc. DISAGREE

TM Holder

ALDI
527 RUE CLEMENT ADER PARC D'ACTIVITE DE LA GOËLE, 77230, DAMMARTIN-EN-GOËLE
FR

Company

Aldi, Inc.
Registered Agent Address: C T Corporation System 208 S LaSalle St, Suite 814 Chicago, IL 60604-1101 United States
Illinois US
Aliases: aldi
Sectors: Wellness and Stimulation

Stage 1 score0.9975 Original methodexact (0.900)
Ollama score0 review_queue#277 (pending)

Ollama's reasoning: "Names differ significantly (ALDI vs Aldi, Inc.) and there is no clear abbreviation, acronym, or distinctive parent name overlap. Address signals show postal codes are different."

PT. Global Nutrition Indonesia ↔ PT. SPV International Indonesia DISAGREE

TM Holder

PT. Global Nutrition Indonesia
J1. Raya flternatif Sentul No. 28 RT.001 RH.001 Nanggeuer Cibinong, Bogor, Jaua Barat, 16912
ID

Company

PT. SPV International Indonesia
Komplek Union Industrial Park Blok A No. 16 Kel. Tanjung Sengkuang, Kec. Batu Ampar Kota Batam, Kepulauan Riau, 29453
ID
Aliases: indonesia international pt spv
Sectors: Manufacturing, Smoking Alternatives

Stage 1 score0.9911 Original methodfuzzy (0.788)
Ollama score0 review_queue#267 (pending)

Ollama's reasoning: "The distinctive core names do not match. While both are PT companies in Indonesia, only the generic word 'International' overlaps."

FIVE STAR BAKERY INC ↔ Five Star BV DISAGREE

TM Holder

FIVE STAR BAKERY INC
4272 N State Rd 7
Lauderdale Lakes FL US

Company

Five Star BV
Avenue Ceramique 221 6221KX Maastricht
NL
Aliases: five star
Sectors: Raw materials, Smoking Alternatives

Stage 1 score0.9901 Original methodlei (0.700)
Ollama score0 review_queue#264 (pending)

Ollama's reasoning: "Names differ in legal form and country of incorporation. The distinctive core name 'FIVE STAR' does not fully overlap."

FIVE STAR BAKERY INC. ↔ Five Star BV DISAGREE

TM Holder

FIVE STAR BAKERY INC.
4272 N State Rd 7
Lauderdale Lakes FL US

Company

Five Star BV
Avenue Ceramique 221 6221KX Maastricht
NL
Aliases: five star
Sectors: Raw materials, Smoking Alternatives

Stage 1 score0.9901 Original methodlei (0.700)
Ollama score0 review_queue#268 (pending)

Ollama's reasoning: "Names differ in legal form and country of incorporation. The distinctive core name 'FIVE STAR' does not fully overlap."

04

Already resolved (reference only)

CITRUS CBD INDUSTRIES INC. ↔ Citrus CBD LLC RESOLVED

From the earlier 300-row spot-check. Stage 1 scored 0.9995 (identical normalized names, "cbd citrus" on both sides). Ollama rejected it (score 0) — correctly, per human real-world research: the holder is a VC-backed California CBD brand ("Peels"), the company is an unrelated Oregon LLC at a private residential address. But Ollama's stated reasoning ("core names don't overlap") was false — the names are identical; it likely reached the right answer by over-weighting the address mismatch (a documented mistral bias), not genuine name analysis.

Resolved: match_method='human_rejected', confidence=0, review_queue #263 status rejected. Full record: output/stage1-mismatch-3107753.json.

05

Recommendations

  1. Review the 17 open review_queue entries (all status='pending') and resolve each to human_rejected or leave as-is if confirmed correct.
  2. Fix the traced LEI parent-walk bug: add Portuguese/Spanish generic corporate tokens to GENERIC_CORPORATE_TOKENS in src/match/lei.js.
  3. Consider a broader audit of GLEIF's other large multinational parents (companies with many global subsidiaries) for the same generic-token bridging risk in other languages.
  4. 99.6% full-population agreement is a strong result for the Stage 1 threshold as calibrated — no change to STAGE1_APPROVE_THRESHOLD (0.99) is indicated by this audit.