Forensic screens

Parameters, response columns and types are generated from the SQL catalog and render under Endpoints → Screens in the sidebar. This page carries what a schema cannot: what each screen measures, what else looks the same, and how to narrow one that is too wide.

These are signals, not verdicts. A row is a fund or a series whose public filings crossed a stated threshold. It is not a score, a rating, a rank of suspicion or a finding, and none of those is served. Every pattern below has an innocent explanation that produces exactly the same row. Confirm against the fund’s own filings before repeating anything a screen returns.

Overview

The same screens the dashboard runs on /suspicious, /dormant and /fidc, served from one definition: each screen_* function calls the screen the dashboard reads, so the page and the API cannot disagree about who crossed a line.

Function	Family	Crosses the line when…	Dashboard
`screen_zombie_growth`	FIDC	delinquency > `p_min_delinq_pct` of NAV while NAV > `p_min_aum`, one aging month	/suspicious
`screen_evergreen_aging`	FIDC	the > 1080-day bucket peaks above `p_min_longtail_pct` of delinquency and moves < `p_max_variation_pp` over the window	/suspicious
`screen_delinquency_drivers`	FIDC	every FIDC, first vs last month, the move classified into five drivers	/fidc
`screen_captive_vehicles`	FII	NAV peaked above `p_min_aum` while quotaholders stayed below `p_max_investors`	/suspicious
`screen_overdue_securit`	CRI / CRA	past filed maturity, newest filing not terminal, volume > `p_min_volume`	/suspicious
`screen_dormant_funds`	FI	every month of the window filed, zero subscriptions and zero redemptions	/dormant
`screen_dormant_trend`	FI	the dormant screen counted at every month-end	/dormant

Three more screens look at how funds file rather than what they hold — restatements, late and silent filers — and have no dashboard page. They follow the same contract and live on Filing-behaviour screens. The machine-readable twin is screens in catalog(): per screen, the defaults, the allowed ranges, the output filters and a meaning that says what else produces the same pattern.

Every row says how it got there

Two columns close every row, whatever the screen:

Column	Meaning
`screen`	which screen produced the row — `zombie_growth`, `evergreen_aging`, …
`params`	the exact arguments it was evaluated with, keyed by argument name — replay the call with them

{ "cnpj": "11111111000191", "fund_name": "…", "period": "2026-06-30",
  "nav_mm": 2.0, "delinquency_pct": 15.0,
  "screen": "zombie_growth",
  "params": { "p_period": null, "p_min_delinq_pct": 5, "p_min_aum": 1000000 } }

A row read without its params is a claim without its premise. Keep them together when you pass a result on.

Calling one

With no arguments, every screen reproduces its dashboard page.

curl -s -X POST "$SILO_URL/rpc/screen_evergreen_aging" \
  -H "apikey: $SILO_ANON_KEY" \
  -H "Content-Type: application/json" \
  -d '{}'

# The same screen, stricter: 18 months, an 80% long tail that moved under 5 p.p.
-d '{"p_lookback_months":18,"p_min_longtail_pct":80,"p_max_variation_pp":5}'

Screen	Defaults (the dashboard’s)	Allowed
`zombie_growth`	latest aging month, 5%, R$1mm	`p_min_delinq_pct` 0..100 · `p_min_aum` ≥ 0
`captive_vehicles`	3 months, fewer than 10, R$50mm	`p_lookback_months` 1..36 · `p_max_investors` 1..1000 · `p_min_aum` ≥ 0
`evergreen_aging`	12 months, 70%, 10 p.p.	`p_lookback_months` 3..36 · both percentages 0..100
`overdue_securit`	R$100k	`p_min_volume` ≥ 0
`dormant_funds`	3 months	`p_lookback_months` 2..12
`dormant_trend`	3 months, 36 months of history	`p_lookback_months` 2..12 · `p_history_months` 1..60
`delinquency_drivers`	latest complete month, 12 months, 6 observations, R$1mm, 1 p.p.	`p_months` 2..24 · `p_min_months` 2..`p_months` · both deltas ≥ 0

A threshold outside its range, or null, raises 22023 naming the range. It is never clamped: a screen evaluated at a threshold you did not ask for is a different screen with your label on it.

Too many rows: narrow, don’t page

Screens do not page. Above one 1,000-row page they raise 22023 (Conventions → Row caps) — a screen is a short list or it is the wrong screen. Two of them exceed a page at their defaults, and take output filters that narrow the result without changing the screen (echoed in params like the thresholds):

screen_dormant_funds — p_dormancy (empty_shell or parked_capital) and p_min_nav (keep classes whose last NAV is at least this). The parked-capital population alone was 8,257 classes in September 2026: pin p_dormancy and walk p_min_nav bands.
screen_delinquency_drivers — p_driver, one of the five classes. Every FIDC with enough observations gets a row, stable included.

An unknown filter value raises 22023; it is not ignored.

What each screen cannot tell apart

Screen	The same row also comes from…
`zombie_growth`	a distressed-credit mandate; a fund in orderly wind-down; one late payer in a small book. `delinquency_pct` is delinquency / NAV and can exceed 100.
`evergreen_aging`	a slow judicial recovery; a policy of not writing off. `months_observed` counts the aging months actually filed.
`delinquency_drivers`	anything that moves two numbers: no sector, no debtor, no guarantee. `value_up_rate_masked` is NAV growing with delinquency; `denominator_only` is NAV shrinking, not new delinquency.
`captive_vehicles`	exclusive and family-office FIIs, which are legal and look exactly like this.
`overdue_securit`	a series that was extended, renegotiated or simply not yet re-filed. The filing is stale; the screen cannot say why. `status` is served as filed. The series number is not a column, so two series under one `instrument_code` read as two rows with the same identity.
`dormant_funds`	exclusive and closed structures (`parked_capital`); a vehicle registered ahead of its launch (`empty_shell`). A month with unreported flows or quotaholders disqualifies a fund rather than counting as zero. FI only — the other families file no monthly flows.

Windows and freshness

dormant_funds, dormant_trend and delinquency_drivers anchor on the family’s latest complete period (coverage().complete_through), never on today, so a partially-filed month cannot read as “no flow”.
captive_vehicles and evergreen_aging look back from today; a month CVM has not published yet is simply absent.
delinquency_drivers refuses a window that starts before 2025-01: FIDC delinquency is not filed before then (the regime break). Shorten p_months or move p_end.

Limits

No tier ceiling: the page is the ceiling, identically for every caller. The binding limit is the query budget — 3 s anonymous, 8 s signed in. dormant_trend over its full 60-month range is the heaviest call here; if it times out anonymously, sign in or shorten p_history_months.

Overview

Data

Get started

Guides

API reference

Endpoints

Forensic screens

Overview

Every row says how it got there

Calling one

Too many rows: narrow, don’t page

What each screen cannot tell apart

Windows and freshness

Limits

​Overview

​Every row says how it got there

​Calling one

​Too many rows: narrow, don’t page

​What each screen cannot tell apart

​Windows and freshness

​Limits

Overview

Every row says how it got there

Calling one

Too many rows: narrow, don’t page

What each screen cannot tell apart

Windows and freshness

Limits