Red Flag RadarForensic public-record signals
Start a pilot

Methodology

Red Flag Radar reads several US public registers and turns each into structured, point-in-time events for a company. Every signal carries a rule version, a confidence score kept separate from severity, an evidence hash back to the source, and an Ed25519 signature verifiable at /v1/pubkey. Fields are descriptive, never conclusive. See how it works for the product overview.

Descriptive, not predictive — by design. Red Flag Radar surfaces verified public-record events for human credit and risk assessment. It does not generate default probabilities, credit scores or recommendations, and makes no claim that its events anticipate any outcome. Every figure it produces is a fact already on a public record, resolved to a company, dated, and shipped with its evidence — which is what makes it reviewable by a risk or model-governance team rather than something they must validate as a model.

Each source follows its own availability model. Filing and recordation events (USPTO, SEC) are dated as of their public record date, no look-ahead. Historical SBA credit outcomes are retrospective labels — the FOIA feed carries a current status per loan — with true point-in-time vintage capture only from 2026 onward (each vintage snapshotted as it publishes). We do not claim a perfect reconstruction of what was publicly observable each quarter before then.

The registers

Entity resolution across registers

The registers do not share a company identifier, and each refers to a company by name — an unreliable key in both directions: one company files under many spellings (IBM, Intl Business Machines Corp, International Business Machines Corporation), while unrelated companies share a name (two Acme Holdings, in different states). Matching on the name alone both misses real links and invents false ones.

Instead, every register record is resolved to a canonical entity. The backbone is the public DOL Form 5500 filings, which tie a company to its EIN (the federal employer tax ID) and record every name spelling it has filed under. A record from another register is attached to that backbone by name + state, grounded on this EIN-linked name corpus — not by bare string equality. A name that resolves to a single EIN is linked; a name shared by several EINs that the state cannot disambiguate is sent to a human review queue, never guessed (precision before recall). Every link carries a method, a confidence, and a valid-from / valid-to date, so mergers, tax-ID reassignment and ticker reuse cannot leak look-ahead or survivorship through the graph.

Identity is then reported on two honest axes: a name match (how well the string matched) is kept separate from entity confidence (whether it is one real company). A name-only match with no EIN, state or graph anchor is a candidate pending review, never a resolved entity. Because the backbone is EIN-anchored, a name + state query can still resolve to a proven entity and surface every register that entity appears in — a signal present in two registers on the same proven company, not a name coincidence. Coverage and matcher precision are measured on audited samples, not asserted — see coverage & precision.

IP signal detectors

IP_SECURITY_INTEREST_RECORDED

Fires when a recorded patent assignment's conveyance is a security interest (patents pledged as loan collateral). Releases and terminations are excluded. The flagged entity is the assignor (the patent owner granting the pledge).

PORTFOLIO_ASSIGNMENT_OUT

Fires when a plain assignment transfers a large block of patents out in one recordation (sale / M&A / distress). Quarantined to review before publication.

A recorded filing is a fact, not a judgment about the entity, its patents, its loans, or its securities.