Filed numbers deviate from natural-frequency distribution

Rule id: benford · Version: 2026.04.27-v1

Microsoft Responsible AI Transparency artefact. One Model Card per ML signal RecoupIQ ships. Generated from lib/modules/intelligence/canonical-warning-adapters.ts, re-runnable with npx tsx scripts/responsible-ai/generate-model-cards.ts.

What this rule does

The leading-digit distribution of values in this company's filed accounts deviates from the Benford's-Law expectation. Used in forensic accounting as a low-cost anomaly screen.

When it fires

• Severity (when fired): Medium
• Confidence in detection (when fired): requires_verification

Severity is bounded by the magnitude observed. Confidence is bounded by the number of independent dimensions that agreed. The two are deliberately separate: a Medium warning at requires_verification confidence has very different triage weight from the same rule at confirmed confidence.

Inputs

Evidence the rule cites when it fires (illustrative, actual values vary per company):

• Benford anomaly detector, Chi-square statistic: 24.50 (above the 15.51 threshold for 8 degrees of freedom at 5% significance).

Known false-positive triggers

Documenting where the rule is most likely to mis-fire is a Responsible AI Transparency requirement. RecoupIQ surfaces these directly to buyers in the report panel so they triage informed.

• Small-set accounts (micro-entities with very few line items) routinely fail Benford screens by sample-size alone.
• Round-number conventions in published accounts (£thousands rounding) systematically inflate the test statistic.

Recommended next steps

Read alongside the autoencoder signal, a joint trigger from both is more informative than either alone. Treat as a screening flag for a deeper accounts review, not as a finding of misstatement.

Notes on language: this is process language only, review, verify, escalate. RecoupIQ does not give regulated financial advice ("do not pay", "refuse the contract") in any report. See lib/modules/intelligence/canonical-warning.ts for the validator that enforces this.

Intended use

• Pre-contract due diligence on a UK counterparty.
• Pre-credit-extension sizing for trade credit or invoice finance.
• Post-default forensic evidence pack (Practitioner Pack tier).
• Litigation / SBC adjudication evidence (Practitioner Pack Plus tier).
• M&A / investment diligence snapshot.

Out-of-scope use

• Not legal advice. Always consult a regulated insolvency practitioner, solicitor, or compliance professional before acting on a warning.
• Not a definitive determination of solvency, sanctions status, or director fitness. RecoupIQ reports on public-record signals; the registers themselves are the authoritative source.
• Not a credit score. The signals correlate with but do not predict insolvency probability in any calibrated sense.
• Not suitable as the sole basis for refusing service to a counterparty.

Data sources

Detailed lineage lives in the Methodology page linked at the top of this card. Source registers cited above are the immediate inputs.

Versioning + reproducibility

Reports stamp the rule id + version ([email protected]) in their footer. To reproduce an older report, pin the same version of lib/modules/intelligence/canonical-warning-adapters.ts from git and re-run the screen.

Right of reply

Subjects of any fired warning may submit a structured rebuttal via the Right-of-Reply endpoint linked from every report panel. RecoupIQ does not auto-suppress a warning based on rebuttal, but the rebuttal is recorded against the report and surfaced in any downstream version.