Compliance and security.
The decision is built for US consumer lending, and defensibility is in how it is constructed rather than bolted on after. This is the regulatory and security posture behind that claim.
Built for US consumer lending.
Themisure is built against the frameworks that govern an autonomous credit decision in the United States: ECOA, Regulation B, FCRA, and SR 26-2 (successor to SR 11-7), with a SOC 2 path.
- ECOA and Regulation B
- The decision is built to give a specific, lawful adverse-action reason on every declined or counter-offered application.
- FCRA
- The adverse-action notice reflects the real basis of the decision, grounded in the policy and the applicant's facts.
- SR 26-2
- The decision is treated as a model under management: a model card, a shadow monitor, drift monitoring, and replay.
- SOC 2 path
- A SOC 2 path is part of the program. It is named as a path, not claimed as a completed certification.
Adverse action by construction.
Every decision carries a specific adverse-action reason, grounded in an explicit policy clause and the applicant's facts, so the ECOA and FCRA notice reflects the real basis of the decision. A model too opaque to give specific reasons cannot make adverse credit decisions; here, the reason is built at the same time as the decision, not reconstructed after it.
Fair lending, tested.
Decisions are tested for disparate impact against a four-fifths standard over the decision population. The testing is honest about sample size rather than asserting a result the data cannot yet support.
Model risk management under SR 26-2.
The decision is governed as a model: documented, monitored, and reproducible.
- Model card
- The decision model is documented as a model, with its purpose, its inputs, and its limits stated plainly.
- Shadow monitor
- An interpretable scorecard runs in parallel to the agents to flag drift. It watches the AI; it never gates a decision.
- Drift monitoring
- Drift between the agents and the shadow monitor is surfaced rather than hidden, so a change in behavior is visible early.
- Replay
- Any decision can be re-run and reproduced to the byte against the exact policy version that was live when it was made.
Policy versioning and the audit trail.
Every decision is pinned to the exact policy version that was live when it was made, so the decision and its rules are inseparable. Because the version is fixed, the decision is reproducible: re-run it and it returns the same result, to the byte, with the full agent trace and event timeline sealed as an immutable audit.
Security posture.
The decision engine is private. It is reachable only by the gateway, and only with a signed token, so the engine is not exposed directly. The demo data is synthetic and carries no real PII; sample and synthetic data is marked as such wherever it is shown.
For the vendor-security evidence a procurement review asks for, including subprocessors, the internal security audit, and standard questionnaire answers, see the trust center.