Insurability requires controls and evidence
Insurability requires controls and evidence, not assurances.
Core publication
Insurable AI
An argument about the preconditions for underwriting AI, not a claim that SDL satisfies them. It sets out what an insurer would need before an AI system is underwritable, and the gap between that and what SIR produces today is stated further down this page rather than left for a reader to find.
Written September 2026. Nothing in it is regenerated, so read it as dated.
Abstract
Core claim
Insurable AI sets out the conditions required for AI systems to be underwritable in the real world. Insurance does not price ideals. It prices controls, evidence, and repeatable performance under stress. This publication describes five criteria that move AI from a trust based posture to an audit based posture, including governance embedded design, enforceable control points, and verifiable decision records. It connects these criteria to practical deployment questions that insurers, auditors, and risk owners care about. What happens when the model is pressured, when policies conflict, when users attempt bypass, and when failure must be proven rather than asserted. The central claim is that insurability requires deterministic infrastructure around the model, including pre inference enforcement and hash-linked audit records. With the right architecture, systems can produce independently verifiable artefacts, demonstrate consistent behaviour, and support independent verification. Without that, safety remains probabilistic and accountability remains unclear. The outcome is a pragmatic framework for building AI systems that can be governed, audited, and insured.
Key findings
Main points
Governance must be embedded in architecture
Governance must be embedded in architecture, not added as a wrapper.
Deterministic enforcement improves predictability
Deterministic enforcement reduces variance and improves predictability.
Audit grade artefacts must be independently verifiable
Audit grade artefacts must be independently verifiable.
Risk evaluation improves when failure modes are testable
Risk evaluation improves when failure modes are measurable and testable.
What SDL does not satisfy
Where the argument and the product part company
SIR cannot attribute anything to anyone
An insurance claim is about who did what and what followed from it. SIR has no actor, user, session or principal anywhere in it, and it is input-only and pre-inference, so it observes neither the model's output nor any downstream action. It can establish that a request matching a published rule was refused before a provider was called. It cannot say by whom, or what happened next. That is architectural rather than unfinished.
The attribution hook is in the record format, not the gate
A decision record can carry a receipt held by the subject of a decision, and a receipt naming a decision the record does not contain refutes the record. That is the one mechanism here that runs against the producer's interest, which is what a claims process needs: the party relying on it holds their own copy and the producer cannot withdraw it. See what a record does not establish.
No production decision has ever been recorded
Every published archive is a benchmark run against a corpus written by the same person who wrote the rules. Nothing here has yet decided anything for anybody else, which is the first question an underwriter would ask and the reason the pilot exists.
Canonical file
The PDF version is the canonical downloadable file for archival and printing.