What SDL publishes

Bounded claims, and artefacts a reader can check.

Two things, for two different problems. SIR is a rule-driven gate for requests to a model. Decision Record is a format for checking recorded decisions. The gate can emit records; the format does not need the gate.

SIR

Signal Integrity Resolver

If you need a rule-driven gate in front of a model, and a signed record of what it blocked. No model sits in the decision path, so the same input under the same policy produces the same decision, and every run leaves a certificate bound to a hash-chained ledger.

What you can check

A published archive, verified from a clean clone in under a minute: signing trust, the signature, the ledger binding, the counters recomputed from the rows, and custody of every file the signed manifest names.

What it will not tell you

Whether the model is safe. SIR's claim is about rule coverage, and the rows it does not cover are published by identifier and left unpatched.

Overview Evidence Limits

Decision Record

If you are being asked to rely on a record of decisions somebody else produced. A format for the record any deterministic procedure leaves behind, and a verifier that checks a record written in that format without trusting whoever produced it. A specification first, with a reference emitter and verifier beside it.

What you can check

Fourteen records built to be refused, each run through the verifier with its output captured rather than written, plus nine worked records across both profiles, streams and receipts.

What it will not tell you

That the record is the whole story. Completeness is never established from inside a self-signed document, and a record can pass every property over a call the system never handled. That record is published too.

Overview Evidence Limits

Bounded claims

Both claims are deliberately narrow.

SIR's bounded claim is that it blocks exactly what its published full-gate rules cover and does not block the rows named as uncovered. For the EU AI Act pack, the published coverage JSON names 26 uncovered rows, and gated runs leaked exactly those rows across 7 SIR gate versions. Read that as a consistency check rather than an independent confirmation: the uncovered list is generated by calling the gate's own rule function, so the prediction and the result are one function evaluated twice. What is not tautological is that those rows are named publicly, by identifier, and left unpatched. Check whether they are still named a year from now. Review the Evidence page for the signed run artefacts.

SIR evidence

Decision Record's bounded claim is that every property resolves to one of four states, and that unknown is never counted as a pass. A check that could not run is not a check that succeeded, so a verdict is the worst state present rather than a tally of the good ones. Three limits are structural and stay that way: completeness cannot be established from inside a self-signed record, the retained inputs are the producer's word, and a party offered as independent has to hold a key the producer does not. Each is exhibited as a record you can run, not described.

Record format limits