The model priced the risk. Something still has to answer for the decision.
Underwriting and claims decisions carry statutory constraints that do not move because a model produced the number. This scenario shows the constraint enforced at the moment the decision would take effect.
What you are already obliged to produce
Insurance regulators moved early on AI governance, and the obligations landed on the insurer rather than the vendor.
| Authority | What it obliges you to produce |
|---|---|
| NAIC AI Model Bulletin | A written AI systems programme with governance, risk management and documented controls over models used in regulated decisions. |
| Unfair discrimination statutes | Decisions must not rest on protected characteristics, including via proxies that reproduce them indirectly. |
| Solvency II governance | Effective governance including a clear allocation of responsibilities and an auditable decision trail. |
| IFRS 17 | Measurement and disclosure that hold up under audit, including the basis on which claims liabilities were recognised. |
Why the usual answer does not close
A fairness review sampled after the fact cannot tell a regulator that no impermissible decision was executed.
Proxy features are not always visible at review time; the question is whether an impermissible basis reached the decision, not whether anyone intended it to.
Claims automation scales the consequence of a wrong rule far faster than a quarterly review cycle can detect it.
The bulletin asks for documented controls over models in production. Documentation of intent is not the same artefact as a record of what each decision was actually permitted to do.
One action at the boundary
The engine measures the action against both an approved reference and a prohibited one in a single decision — the arrangement the ablation on our evidence page exists to justify.
Candidate action
claim.deny.execute · $340,000 · protected-class proxy in decision basis
The action read as compliant against the approved reference alone. Scored against the prohibited reference in the same decision, the margin collapsed — the case sat as close to a known impermissible basis as to a permitted one. A single-reference gate passes this class of case; on our held-out set it passed four in ten.
The same measurements, whatever the sector
These are figures from the internal technical evidence report, not projections modelled for this scenario. They describe one pipeline, so they do not change when the mandate does.
0 / 18
Attack episodes passed
Stage 0 SHADOW evaluation of action traces authored by real generative planners. None of nine held-out benign episodes was blocked.
0 / 10
Prohibited cases passed, dual reference
Controlled ablation on the same held-out set. A single-reference gate passed four of ten under identical calibration.
0.47 ms
Mean decision latency
Across 5,000 measured decisions; 1.15 ms at p99. The gate itself completes in 258 nanoseconds.
1,000 / 1,000
Identical digests on replay
Same input, pinned environment. Ten thousand ledger records re-verified in 0.195 seconds.
What this scenario is not
Read this before you quote it
- An illustrative scenario, not a customer engagement. No client is named because none is being described.
- No deployment in this sector is claimed, and no regulator has reviewed or endorsed this material.
- The figures are prototype measurements on one commodity workstation, not production or distributed results.
- Nothing here has been independently reproduced by a third party.
- The authorities cited describe the obligation you carry — not a determination that we satisfy them.
- Operational accuracy on a validated sector corpus remains outstanding; a pilot requires your own reference data.
Check the numbers before you trust the scenario
Every figure above is drawn from the technical evidence page, where the same measurements appear with their tail distribution and their stated limits.