Skip to content

0018: Proposed admission terms and review binding

Date: 2026-10-20. Status: engineering proposal awaiting owner/relying-party review. Scope: W13. No scientific protocol is adopted and no prior experiment is reused.

Subsequent scope approval and delegated design are recorded in decision 0019. The statements below describe this decision's original proposal stage; its binding semantics remain unchanged.

The concrete proposal names one assay, seven comparisons and explicit holds. Its proposal schema is distinct from AdmissionProtocol; incomplete or null terms cannot pass as an admitted protocol.

Separate assay and deployed policy identities

An admission protocol must exist before training. The runtime Policy requires final classifier identities, so it cannot also name an untrained assay without a dependency cycle. StudyPolicy uses hallmark/study-policy/1 in the existing policy hash domain and binds inj.direct/1, English prompt scope, rubric byte hash and a positive byte cap no greater than 32,768. The schema separates its identity from the runtime Policy. Runtime enrollment remains a later, distinct review.

Acyclic review

  1. Freeze all protocol terms, including nested source and budget approval hashes.
  2. Compute H("protocol", ["hallmark/admission-terms/1", terms]), excluding only the top-level approval_ref.
  3. Prepare AdmissionReview naming those terms, owner, relying-party role, both approval documents and an exact bounded document inventory.
  4. Set the final protocol's approval_ref to that review's identity.
  5. The caller authenticates the reviewers and independently enrolls that identity.

Changing a term requires a new review; changing the top review reference changes the final protocol identity but not the terms hash. Copying a pin from a supplied dossier provides no authority. The verifier checks names, identities and bytes, not signatures, review content, legal rights or statistical adequacy.

Document semantics and bounds

SourceApproval.revision is the SHA-256 of a revision-manifest document, not a Git commit ID. The document records the exact source revision and declared inventory. license_text, source/budget approval references, rubric, split manifest, evaluator and candidate code/config fields are SHA-256 hashes of exact document bytes. Code/config documents must identify the complete implementation and dependencies; a filename alone is insufficient review evidence. The final protocol and study/review links use typed artifact identities.

Candidate base files are declared, sorted unique FileRefs. This check does not fetch or inspect asset bytes. Import/runtime adapters must verify approved bytes separately. Review evidence is sorted, unique and exact: at most 256 documents, 1 MiB each and 8 MiB total. All required references/hashes must be represented. Private operational dossiers must be supplied through an authorized private channel; fixture documents are harmless public markers.

Decimal strings are canonical rationals in [0,1], with at most 18 fractional digits, no redundant zeros or exponent. FPR/alpha are strictly between 0 and 1, TPR is positive, loss is below 1. Three distinct uint64 seeds and the seven-item ordered roster are mandatory. Resource caps are positive except an explicitly zero cost, stages fit within total wall time and verification is capped at 60 s. Declarations do not allocate or enforce resources.

Validation and migration

Fixed CBOR/JSON/schema vectors and admission-negative tests cover missing pins, changed terms, role mismatches, incomplete evidence, invalid budgets, ambiguous numbers and unsafe constructed records. make check-admission and both package gates exercise the no-download marker fixture. The actual proposal is rejected as an admitted protocol. The old conformance corpus and prior results are unchanged; new admission vectors are separate component evidence.

This API does not authorize the legacy synthetic runner or implement data import, training, hardware access or proof execution. A future runner must bind its actual inputs to independently admitted terms and enforce the full resource envelope. Owner review is required before this proposed contract is used for real research.