Features

Model registry

Every AI model your firm uses, registered, versioned, and named — so the regulator never has to ask which model produced which decision.

The model registry is the system of record for every AI model that has ever produced an artefact your firm has acted on. Each model is identified by a provider and version. Whenever a piece of advice is submitted for review, the provider and version of the model that drafted it are captured against the review job. The registry is built from those captures.

Why it matters

Consumer Duty (PRIN 2A) requires firms to know which tools they used to produce a customer outcome. In practice, this means: when something goes wrong with a customer's advice, you need to be able to answer “which model produced this, when, on what version, and which other customers got advice from the same version?” Without a registry the answer is “we'll have to dig through the database.” With one, it's a single API call.

What gets recorded

  • provider — the model vendor (e.g. openai, anthropic, google, internal)
  • version — the version string the provider returned (e.g. gpt-4-2024-11-20)
  • firstSeenAt — first time this firm produced advice with this provider+version
  • lastSeenAt — most recent advice produced with it
  • jobCount — total review jobs attributed to this version
  • outcomeBreakdown — approved / approved-with-modifications / rejected counts

How models get registered

There's no separate “register a model” call. Models are auto-registered the first time they appear on a submitted job. When you POST a review job, include the model provenance:

json
{
  "documentType": "SUITABILITY_REPORT",
  "documentUrl": "https://files.example.com/report.pdf",
  "modelProvider": "openai",
  "modelVersion": "gpt-4-2024-11-20",
  "clientReference": "client-12345"
}

The first time Bedrock sees that (provider, version) tuple for your firm, it appears in GET /v1/firm/me/models. Every subsequent job updates the counters.

Listing models

bash
curl https://api.bedrockcompliance.co.uk/v1/firm/me/models \
  -H "X-Bedrock-Key: bk_live_..."

Returns a ModelSummary[]. See the schema in the Firm API reference.

Evidence produced

  • Per-job model provenance attached to every review job
  • Aggregate model summary in the registry endpoint
  • Drift signals (see Drift detection)
  • Activity timelines for each model (see ModelTimelinePoint)

FCA mapping

  • PRIN 2A.5 — firms must monitor and review the support given to retail customers
  • SYSC 8 — outsourcing of operational functions and recordkeeping
  • SYSC 9 — record-keeping requirements

See also