The published standard defining the mandatory boundary between LLM output and deterministic decision execution. DSK V1.0's Deterministic Conformer Layer, sefaIntake endpoint, and QuarantineRecord entity are the canonical reference implementation. Third-party LLM decision systems conforming to this protocol may submit to the conformance harness for public verification.

1. Mandatory Deterministic Canonicalization

LLM output must never be passed directly to a decision engine. A deterministic conformer layer must canonicalize LLM-structured input into a typed, validated canonical form before any decision logic executes. The conformer coerces free-text context fields to strict per-domain types, clamps advisory fields to their range, and passes safety-critical fields through to the engine's hard-rule gates. The LLM cannot influence the decision logic, the score, or the ruleset — it only structures the input envelope.

2. Typed Rejection Code Set

The conformer must reject non-conformant input with a typed error code from the closed set below. Rejections are never silent and never passthrough — a rejected input produces no decision. The code set is exhaustive and versioned.

3. Mandatory Quarantine Records

Every AI-sourced decision must persist a quarantine record capturing the pre-conformer raw payload, the post-conformer canonical payload, the conformer version, the conformer hash, the ruleset version, the ruleset hash, the input hash, and the output hash. The frozen conformed payload is the replay anchor — the LLM is never re-invoked on replay. Quarantine records are audit-only and never re-run through the kernel except for deterministic verification.

4. Signed Receipt Bundles

Every AI-sourced decision must carry an HMAC-signed receipt bundle bound to both the ruleset version and the conformer version. The signature attests the receipt is real and reproducible. A receipt without a valid signature is not conformant.

5. Replay-Verifiability Across Ruleset Evolution

A historical quarantine record re-executed against its frozen ruleset version must reproduce the stored output hash byte-for-byte. The conformer version, ruleset version, and ruleset hash are pinned to every receipt so determinism is durable across ruleset and conformer evolution.

6. Hard-Coded Domain Governance

The AI intake path must enforce a hard-coded domain check — unrestricted AI intake without an explicit, valid domain is rejected as a governance violation (DC_AMBIGUOUS_VALUE). The domain routes the payload to the sealed domain lane; the LLM cannot select or invent domains.

7. Reference Implementation

DSK V1.0 is the canonical reference implementation of this protocol: the Deterministic Conformer Layer (DC-V1.0), the sefaIntake endpoint, and the QuarantineRecord entity implement every section above. Third-party systems conforming to this protocol may submit to the conformance harness for public verification.

The conformer must reject non-conformant LLM output with a typed code from this closed, versioned set. Rejections are never silent and never passthrough.