How CoverProof achieves Section 250 compliance

The six-step compliance process

Inside the classification pipeline

Every classification is the output of a versioned pipeline with explicit rules, thresholds, and reproducibility guarantees. Here is what runs on every row before a verdict is recorded.

The pipeline is the product. The AI is one component inside it.

On scope

• The pipeline reasons over the structured data your register actually contains — role title, function codes, seniority indicators. Individual-level context (day-to-day responsibilities, employment contracts) is the compliance officer's domain on review.
• The s.250(3) test is a functional test that courts apply to the facts of each case. CoverProof produces a structured risk-screening verdict against that test — one input to your compliance decision, not a legal opinion.
• CoverProof complements independent legal advice on your firm's specific Section 250 position. It is not a substitute for it.

Technical specification

Methodology questions

Same input → same output (and how to verify it)

Reproducibility is a load-bearing claim for an audit-grade compliance tool. If two classifications of the same individual on different days returned different verdicts without a methodology change, the evidence pack would be unverifiable. CoverProof controls this on five levers:

• Temperature 0. The model is invoked with temperature 0, so the same prompt produces the same response (subject to the next two levers).
• Model pinning. The model id is recorded with the classification (e.g. claude-sonnet-4-6). A model upgrade is a methodology change, not a transparent swap.
• Methodology version pinning. Every classification records the methodology version (e.g. s250-v12) it was produced under. The version pins the prompt content, the statute extract, the schema, and the verification gates.
• Prompt fingerprinting. The system prompt is SHA-256 hashed and stored alongside the methodology version. Any silent edit produces a mismatch detected by the worker on next run.
• Input hash cache. The user-facing inputs (name, role title, function codes, firm context) are SHA-256 hashed before the model call. An identical hash returns the cached verdict rather than re-invoking the model — provable equality, no second call needed.

How to verify on your data: in your gap-analysis review table, click any classified row to expand it. The row shows its input hash, methodology version, and model id. Re-importing the same SM&CR file does not produce new classifications — the file’s SHA-256 is detected and the import is short-circuited. The same row would yield the same verdict on demand.

Scope of CoverProof

An honest scope statement is more useful than a long feature list. Here is what CoverProof does, and where it deliberately stops — so a compliance director can place it correctly inside their existing process.

This methodology has not been reviewed by external legal counsel. CoverProof does not provide legal advice; firms should seek independent advice on their s.250 compliance obligations.

• Not legal advice. CoverProof is a structured workflow that documents your firm’s reasonable steps under Section 250. It does not, and is not authorised to, advise on the application of the law to your facts. Your qualified legal advisers do that.
• Not a regulatory determination. The AI classification produces a risk-screening verdict against the s.250(3) functional test. It is not a finding by the FCA, by a court, or by any other competent authority. Treat it as one input to your decision, not the decision.
• Not a substitute for the s.250 functional test. The statutory test is fact-dependent — actual responsibilities, actual decision-making authority, actual scope of activities. CoverProof reasons over role title, function codes, and seniority indicators because those are the data we have. The model cannot see the individual’s employment contract or actual day-to-day responsibilities.
• The s.250 test and SM&CR approval are distinct questions. Whether an individual meets the s.250(3) functional “senior manager” test, and whether they hold an FCA Senior Management Function, are two independent questions. Holding an SMF does not by itself settle the s.250(3) test, and lacking one does not by itself make a person s.250-exposed. Identifying an SM&CR-uncovered significant-role individual as a gap is a reasoned inference from the distance between those two questions — not an equation written into the statute. A court decides the functional test on the facts.
• Not per-firm tuned. Thresholds (confidence tiers, drift alert cutoff, self-consistency escalation) are fixed across all firms. Per-firm tuning would weaken the “uniform methodology” defence in any future challenge — it would let an adversary argue that a verdict turned on a configurable knob rather than on the statutory test.
• Not retroactively revised. A classification produced under methodology version v12 stays a v12 classification. If we ship v13, prior rows are not silently re-classified. A board evidence pack generated under v12 is still verifiable as a v12 record years later.
• Designed to sit alongside your compliance review. Every CoverProof classification flows through your team’s reviewer queue before any declaration is sent. The internal benchmark catches drift in the model on known scenarios; your reviewers catch anything specific to your firm’s register. The two layers together are the product.
• Not a confidentiality boundary for our sub-processors. Your data is processed by the sub-processors listed in the Trust Centre. The DPA documents the terms and the safeguards.

When the sources disagree

CoverProof reads three sources of truth about who is at your firm: the SM&CR register you upload, the live FCA Register Extract Service, and any in-product corrections the compliance officer makes. These rarely agree perfectly. CoverProof never silently reconciles a disagreement; the policy is explicit.

• SM&CR upload — your firm’s declared truth. What you uploaded is what your compliance officer has signed off as in scope today.
• FCA Register — what the regulator records. The FCA’s view of approved-persons status at the firm. Useful for cross-checking SMF approval, but it does not capture everyone who meets the s.250(3) functional test.
• CO in-product corrections — final. When the CO confirms, overrides, or excludes a row in the review table, that decision becomes the audit-trail record for the gap analysis.

When the SM&CR upload and the FCA Register disagree on who is at the firm, both populations are surfaced as in-scope for review. Names appearing only in the SM&CR upload are reviewed alongside names appearing only in the FCA Register. The CO reconciles either by accepting the row (confirm), substituting a different verdict (override), or removing it from the declaration list (exclude).

The reconciliation is recorded in the audit trail. A 2028 auditor reading the evidence pack can see which source each individual was first surfaced from and which decision the CO made.

Fixed thresholds, by design

Several numbers in CoverProof are configurable in the implementation but not in the product. We use the same thresholds for every firm and we do not let them be tuned at the firm or contract level.

• The confidence tier boundaries that drive review-required escalation.
• The 10pp disagreement-rate alert threshold on the drift monitor (see Quality & drift).
• The MEDIUM-tier self-consistency sample count (k=3) and temperature (0.3).
• The counsel-review trigger cutoffs (see below).

Why fixed: per-firm tuning would let an adversary later argue that a verdict turned on a configurable knob rather than on the statutory test. The uniform-methodology defence is part of the product’s value — trading it for configurability would be a downgrade, not a feature.

When the product recommends counsel review

CoverProof surfaces a non-blocking “Consider commissioned counsel review” banner on the gap-analysis page when one or more of three fixed conditions hold. The triggers are deterministic, evaluated on each page load, and disclosed here in full. Counsel review is always available; the product flags specific situations where it is worth the cost of asking.

• An AI low-confidence row was confirmed without override. The override path is how the CO records disagreement with the AI. Confirming a low-confidence row without using it means the CO accepted the AI’s uncertainty as the firm’s position. Counsel can advise whether that is the right call for the specific boundary case.
• Deadline pressure. More than five declarations have bounced, or more than 20% remain pending within 14 days of the 29 June 2026 deadline. The board pack generated in this state may understate residual exposure; counsel input ensures the pack accurately reflects the programme’s actual coverage.
• A low-confidence individual is being declared rather than excluded. The firm has staked part of its defence on an AI uncertain call. Counsel sign-off before the evidence pack is minuted is recommended.

The banner is advisory and never blocks declarations or evidence-pack generation. It does not replace your firm’s standing relationship with qualified counsel.

Primary sources

The statutory and regulatory materials this methodology relies on. Every claim about scope, applicability, or the s.250(3) test should be verified against these directly.

• Crime and Policing Act 2026, s.250 — legislation.gov.uk — the verbatim enacted text of the senior-manager attribution provision.
• FCA — Senior Managers and Certification Regime (SM&CR) — the regulatory framework that defines the SM&CR baseline against which the s.250(3) test is gap-analysed.
• FCA Handbook — SYSC — primary source for Senior Manager Function (SMF) definitions used in classification.
• FCA Financial Services Register — the authoritative public record of FCA-approved persons; used at classification time when FCA API access is configured.
• FCA Developer Portal — free authenticated API access to the Register, used by CoverProof for live SMF cross-referencing during gap analysis.