How our data works.

Plain English. No legal jargon. Where the data comes from, how we handle gaps, how confident our scores are, and what we never claim.

What we never claim

The hard rule: we describe patterns in public data. We do not call any specific person a fraudster, a liar, or a criminal. The bestiary is a catalogue of behavioural patterns, when a real director's public-register data matches one, we surface the match, not a verdict.

• No accusations. We say "the pattern matches", not "this person is guilty".
• No private allegations. Everything we surface is from public sources you can verify yourself.
• No automated decisions. We give you the data; you decide whether to extend credit, sign, or walk.
• Right of reply. Anyone can dispute a public-register entry directly with the relevant body (Companies House, HMRC, Insolvency Service).

Where the data comes from

Every score we compute starts with public, free, UK Government sources. We don't buy private data or run private opinion as fact.

• Companies House register (every UK Ltd company, free, official)
• The London Gazette (statutory notices: strike-off, administration, winding-up)
• HMRC deliberate-defaulter list (quarterly publication)
• Insolvency Service disqualified-directors register
• Individual Insolvency Register (IIR, IVAs, bankruptcies, DROs)
• Companies House charges register (every secured loan)
• Money Claim Online + tribunal public decisions
• UK Payment Practices Reporting register (large-company payment behaviour)
• Connected-ledger data, anonymised invoice-payment behaviour from RecoupIQ users who opt-in

When data is missing, what we do

Small UK Ltd companies often file abbreviated accounts that omit fields larger companies must disclose. We have three honest options for each missing value: skip the score, infer cautiously, or refuse to score that company at all. Here's how we choose.

• For purely descriptive fields (e.g. address, registered office): we display "not available" rather than guess.
• For computational fields where we need a value to compute a score (e.g. one missing year in a three-year trajectory): we use multivariate imputation (a statistical technique called MICE, Multivariate Imputation by Chained Equations) which estimates the missing value from patterns across thousands of similar companies. Every imputed value is flagged as imputed.
• For composite scores: when more than 30% of input fields are missing or imputed, the overall score is suppressed. We say "data too thin" rather than show you a misleading number.

What MICE actually does (in plain English)

When a small company files abbreviated accounts, certain fields are missing, but the fields that ARE present give us a strong signal. MICE works like an experienced auditor's gut: 'Companies with this turnover, this director, this trade code, and this filing pattern usually have a balance like X.' It fills in the most plausible value based on thousands of similar UK companies.

• It is not a guess. It is a statistical estimate with a confidence range.
• It is not used as evidence. Imputed fields are never cited as fact in our reports.
• It can be wrong. We always show you both the imputed value and the confidence, and you can choose to skip imputed fields entirely.
• It improves with scale. The more UK companies in the reference cohort, the tighter the confidence range becomes.

Honest about what's running today

We classify every detection model in our stack as one of three states. The badge appears on every archetype page so you know what's actually scoring vs what's still being calibrated.

• Live, the model is running in production right now, scoring real UK companies daily. Currently 10 of 21 archetypes.
• In progress, the data is ingested and partial scoring exists, but a known calibration gap is being addressed (per the May 2026 audit). Currently 6 of 21. Each in-progress archetype shows a specific note explaining what is pending.
• Building, designed and statutorily backed, but no production scoring yet. Currently 5 of 21. Free scans do not include these archetypes yet; they ship in a future release.
• Why we tell you this: a confident score on the wrong basis is worse than a missing score. The honesty layer is the answer to "is this for real?".

What we'd like to be wrong about

No detection system catches everything. We have known false-positive and false-negative profiles for each model, published in the May 2026 ML audit. Two important admissions:

• Fuzzy name matches can hit common surnames. Where confidence is below 0.6 we say "match candidate, verify manually" rather than asserting identity.
• New companies (under 12 months) are difficult to score. The data is too thin for the trajectory and capital-bleed models. We say so rather than score them anyway.
• Imputed fields can drift in cohorts where filing rules are changing (e.g. ECCTA introduced new disclosure requirements from late 2025). We re-train the imputer quarterly to catch this.
• When in doubt, we suppress the score. A blank is better than a wrong number.

Your right of reply

If you are a director or company that believes your public-register data is wrong or out of date, the route is to amend it at the source, that is the body of record.

• Companies House, for filings, addresses, director appointments. Free.
• HMRC, for tax-default disclosures. The PDDD list is updated quarterly.
• Insolvency Service, for disqualification orders, IIR entries.
• Once the source register is updated, our scoring follows automatically (typically within 24-48 hours of next ingest).
• If you believe a RecoupIQ scoring is materially wrong, contact us at [email protected], we investigate every report.

Honest scoring beats confident-but-wrong every time.