How it is built
A lending decision engine with 3 model attachments, an evaluation harness, trace-level observability and 10 control tests. Each layer below uses what the layer under it produces.
Controls tested
10 objectives run against the system, 14 findings, evidence attached to each
Layer 3 supplies the traces, scores and explanations these tests read.
Quality and transparency
1,067-case labelled set, Langfuse traces, explanations for 3 audiences
Layer 2 names the 3 call sites that get instrumented and measured.
AI declared
3 attachments, each with an owner, a classification and a fallback
Layer 1 is the decision the models attach to.
AI Security is out of scope
Supervisory framing splits AI assurance 3 ways: AI Quality, AI Transparency and AI Security. This system covers the first 2.
AI Quality
Covered. Accuracy, precision and recall against a 1,067-case labelled set, plus the hard cases the categoriser gets wrong.
AI Transparency
Covered. Drill-to-source figures, a grounded rationale checked before display, and a human-review route under GDPR Article 22(3).
AI Security
Not covered.
AI Security covers attacks on the model itself. Prompt injection puts instructions inside content the model retrieves. Data poisoning corrupts training or fine-tuning data upstream. Model and prompt extraction lifts the asset itself. Membership inference recovers facts about the training data. Denial of wallet drives cost instead of downtime. Excessive agency limits what a compromised agent is permitted to do.
Most of that has no surface here. The system puts no retrieved content into a prompt, fine-tunes nothing and trains nothing, and runs no agent holding payment entitlements. Denial of wallet is the one live exposure. The primary provider has an 8,000 tokens-per-minute limit and traffic above it fails over to a paid provider without a signal. Recorded as F6.
The architecture page covers the platform itself: the interface boundary between the account servicer and the lender, and how the system would extend next.