Skip to content

Specialist material admission: protected shadow controller

Control: GCL-AGENT-ADMISSION-SPECIALIST-001
Work package: MATH-PROGRAMME #1251
Current authority: OBSERVE_ONLY_NOT_REQUIRED.

The shadow controller reads the live pull request or GitHub-native merge group, classifies actual changed bytes/paths against the protected bounded-routine registry, and writes a structured nonauthoritative artifact. It never posts a GitHub approval, publishes an admission status, certifies a mathematical statement, modifies a protected branch, or claims that external specialist evidence exists.

On pull_request_target, it reads the current PR file manifest, pins its exact head/base GitHub objects and routes the full file bytes through the protected routine classifier. Files/bytes/metadata outside the admitted scope become SPECIALIST_REVIEW_PENDING or RESERVED_OR_CONTROL_PLANE_PENDING, not ROUTINE_CANDIDATE_ONLY.

On merge_group: checks_requested, it requires the GitHub-owned queue ref and head SHA to agree with the event, checks the current protected main ancestry, obtains the queue ref without checking out or executing candidate code, inspects the effective diff, and fails closed if protected main moves during a positive routine decision.

ROUTINE_CANDIDATE_ONLY is an observation about the changed material, not permission to merge or proof of scientific correctness. It cannot substitute for a GitHub App review, actual specialist evidence, a required status, or a protected merge queue's own decision.

Protected-domain review evidence (candidate read-only adapter)

ci/specialist_receipt_adapter.py now implements a fail-closed readback for a future protected specialist-review route. It recognizes only a receipt already present on a domain-authorized protected main: MATHFORGE for source-semantic review, MATHSOLVE for execution/replay/adjudication integrity, MATHCERT for independent mathematical proof review, and INTELLECT for protection/authority review. This four-domain registry lives at governance/specialist_admission_domains.json and is checked against a bounded protected-controller code contract. A change to the registry alone cannot expand App review authority. The deterministic receipt path is

governance/material_admission_receipts/MATH-PROGRAMME-<exact-subject-head>.json.

A valid envelope must assert the precise subject Git commit and canonical complete changed-file fingerprint, positive protected-admission disposition, named review scope, and exact Git blob of a separate domain-owned source document. Both receipt and underlying evidence are fetched from pinned protected repository commits; the protected source must remain unchanged throughout the operation. The underlying evidence must additionally anchor an actual merged PR in the domain repository. The adapter reads the source PR, verifies the exact reviewed head, the precise underlying evidence blob in that PR, a real non-author GitHub APPROVED review with the declared review ID, and that no later reviewer disposition supersedes the approval. The role and independence assertions are checked against this attributable GitHub review and exact protected bytes; GitHub approval alone still cannot certify a theorem. The read token has domain-only contents and pull-request read permissions. The adapter never reads a candidate-supplied GCL-REVIEW/1 comment as review authority. GitHub review statuses alone are not proof of the domain record.

The positive unit fixtures exercise the adapter's identity checks, not an actual protected mathematical certificate. There are currently no new protected domain receipts granted by this code. Even when a valid receipt is observed, the shadow mode returns an explicitly nonauthoritative SPECIALIST_EVIDENCE_RECOGNIZED_SHADOW finding. Domain-issued admission depends upon the original domain's protected certification/review process.

This model is intentionally exact-head rather than general evidence-closure equivalence. Accepting an exact material closure across independent base commits will require a separately reviewed normalization/identity migration; it must never be inferred from a SHA coincidence or PR comment.

MATHSOLVE — solution and execution integrity only

MATHSOLVE is now an explicitly recognized specialist domain: SOLUTION_INTEGRITY, protected owner grandchallenge/MATHSOLVE, and evidence confined to contributions/ on the owning repository's protected main. Its role is to test whether a provisional returned result was captured faithfully, reproduced, and adjudicated on the recorded evidence. It does not verify mathematical truth by itself.

The new Solve-scoped review record type is GCL_DOMAIN_ROLE_SCOPED_REVIEW_V1, with positive disposition ROLE_SCOPED_REVIEW_ACCEPTED and ROLE_SCOPED_NON_AUTHOR_SPECIALIST independence declaration. This is not relabelled as a MATHCERT-independent theorem review. An authenticated non-author GitHub review of the exact evidence PR remains necessary for the specialist evidence packet, but not for ordinary MATHSOLVE maintenance.

The Solve proof record and admission receipt must refer to the same solve_execution source-locked record. That record binds dispatch ID, result reference and SHA-256, plus exact blobs of three distinct, already protected JSON artifacts:

  1. GCL_SOLVE_CAPTURE_RECEIPT_V1 — captured result digest.
  2. GCL_SOLVE_REPLAY_RECEIPT_V1 — passed replay of that captured digest.
  3. GCL_SOLVE_ADJUDICATION_RECEIPT_V1 — adjudicated replay, with exact replay-blob and dispatch/result lineage.

Every artifact is fetched from protected MATHSOLVE and checked against its Git blob SHA and the other links in the chain. Missing, ambiguous, negative, stale, or cross-campaign evidence fails closed. The following claim effects are all false in both the Solve binding and adjudication: mathematical claim promotion, certification, publication, source-semantic adjudication, and security authority expansion.

For now the protected Programme routing surface for MATHSOLVE is deliberately narrow: governance/mathsolve_*, governance/solve_execution_*, and governance/solution_integrity_*. A source theorem in a Lean file still routes to MATHCERT; workflow, admission-controller, and protection changes still route to INTELLECT. Mixed or unknown scopes remain unadmitted. This registration does not install a required material-admission status or bypass existing protection. A real domain-produced positive receipt is necessary only when an explicitly reserved claim or authority transition demands Solve execution-integrity evidence, never for ordinary research, maintenance or already-delegated execution.

Selective verification; universal gate cancelled

The Human Steward explicitly rejected requiring specialist verification for every relevant or substantive change. The proposed repository-wide mandatory material-admission check is cancelled, not pending. See the corrected protected-issue programme mandate, MATH-PROGRAMME #1251.

Research exploration, provisional findings, routine operations and authorized agent-executed work continue under their existing CI and protected GitHub merge requirements without a new specialist signoff.

Protected domain-owned evidence can be required by the corresponding specific authority transaction: MATHCERT certification, MATHFORGE authoritative source adjudication, INTELLECT reserved governance/security authority, or genuine external claim promotion. MATHSOLVE capture/replay/ adjudication integrity is available as a narrowly scoped evidence capability when such an effect explicitly calls for it; it neither certifies mathematical results nor creates a routine Solve approval gate.

Any future scope-specific guard must be independently governed, tested with authentic evidence, and demonstrably nonblocking for ordinary work. No blanket required context is to be added to Programme rulesets. Existing GitHub review, protected branch and GH-OS checks remain operative.

The source-of-truth work index remains the issue and protected repository. MkDocs is this explanatory read model, not an independent source of claim authority.