Select a fund to see what each detector reported.
A signal is a lead to review, not a finding of wrongdoing.
Filings from 2026-06-01 to 2026-06-30 · 22 reporting days
run full_202606 · all · analysed 2026-08-14 15:35:21
Leads, not verdicts. A flag is a statistical pattern, not an allegation. Strong flags means strong signals in at least two independent evidence families; Review means one strong or several weaker families. Neither means fraud. Most signals have an innocent explanation. To report an error, write to pedro@deloslabs.xyz.
Incomplete coverage — 1 skipped for missing inputs: market_data. Those detectors report nothing either way; absence of a signal from them is not a clean result.
Network and issuer checks: cross fund issuer, manager network anomalies. These ran outside the per-fund matrix and require counterparty/network review.
Select a fund to see what each detector reported.
A signal is a lead to review, not a finding of wrongdoing.
The pipeline reads CVM's public filings (daily NAV/flows, portfolio composition, fund registry) and runs about 19 independent statistical detectors per fund. None are machine-learned — every one is a threshold or z-score check, fast and free of training data, but only as good as its threshold for a given fund type.
| Family | Detectors | What it flags |
|---|---|---|
| Flow / NAV | flow z-scores, PL drops, redemption runs, flow-vs-performance divergence | statistically unusual captação/resgate or PL moves |
| Valuation | Benford's Law, valuation smoothing, window dressing, price divergence | numbers that look fabricated or smoothed rather than market-marked |
| Portfolio | phantom assets, concentration, reconciliation, cost basis | assets that don't exist in registries, or portfolios that don't add up |
| Structure | circular flow, layered funds, shell networks, cross-fund issuer, peer comparison | patterns matched against the documented Banco Master/REAG scheme |
A synthetic-data eval harness (evals/) injects labeled fraud
into 4 of 40 funds per signal and checks recall and false-positive rate
against the other 36 clean funds. Recall is 100% on every signal tested;
false-positive rate on the clean synthetic funds ranges 0–10%. That
is real evidence the core logic works — but it is a synthetic universe,
not real CVM data, where legitimate funds have structural quirks
(feeder/master fund families, illiquid credit, small AUM) a synthetic
fixture doesn't reproduce.
An early full-manager run flagged 90 of 94 funds (96%) at a major,
reputable administrator — almost all on concentration_violations,
because a feeder fund (FIC) legitimately holding ~100% of its portfolio in
its single master fund's cotas was scored against the same 20–25%
single-issuer limit that applies to a direct-portfolio fund. That is
standard, CVM-permitted architecture, not a diversification failure.
The concentration detector now recognizes a top holding that is itself
another fund's CNPJ and excludes it from the regulatory-limit, HHI and
top-5 checks — the same treatment phantom_assets_by_fund
already gives ordinary fund-of-fund holdings. A fund concentrated in
something that is not a fund cota is still flagged normally.
Strong flags means strong signals (HIGH/CRITICAL) in at least two independent evidence families. Review means one strong family, or several weaker ones. Neither means fraud — every detector here produces a lead, not proof, and needs corroboration this toolkit cannot see: subscription ledgers, bank transfers, counterparty documentation.