gnomonlearn
Docs/RFC0015 CLAIM REVIEW

RFC0015 CLAIM REVIEW

Versioned draft snapshot

RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.

Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.

This is a trace and gap record from the pinned draft specification. It does not establish proof or implementation conformance.

Status: 62 of 62 claims reviewed on 2026-10-05. Owners: Runtime, AI. Final totals: 38 source matches and 24 whole-node mismatches. No normative facts changed.

Configuration, diagnostics, and atomicity

The reviewed group covers environment-independent agents, binding configuration, M01500/M01513–M01517/M01540–M01541, write atomicity, invocation audits, and model digest pinning. Missing subject/object owners include agent.write, agent.audit, atomicity-of-writes, and audit-trail. The Lean catalog instead names agent-write-atomicity and agent-audit-trail invariants.

The allowlist warning claim omits §13.3’s “where one exists” condition. Platforms without an allowlist need defined behavior before this becomes a universal validation promise.

Failure-path audit persistence

§6.2 requires an audit record after a permission violation rolls back the invocation. §11 requires every invocation’s audit in the same transaction as the object-graph write. An audit written in the transaction being rolled back does not survive that rollback. Successful co-commit is coherent, but durable failed-attempt records need an explicit separate failure transaction or another specified persistence boundary. This cannot be established by source matching or by a catalog-level atomicity assertion.

Acceptance obligations

  • Own or repair each missing structured identity after consumer review.
  • Specify success and failure audit persistence separately, including failures before any graph transaction opens and failures during commit.
  • Verify exact diagnostics and rollback with actual runtime integration tests.
  • Define allowlist absence behavior without weakening configured enforcement.
  • Verify model digest changes require reactivation and circuit-open dispatches cannot reach the inference endpoint.

All claims now have source-bound reviews. No runtime execution or transaction behavior was verified by this review.

Security and declaration constraints

The next fifteen claims have ten matches and five whole-node mismatches. The no-escalation and closed-set permission rules use absent invariant aliases; dispatch and write rules use absent effect identities. Conditional allowlist validation, default-deny prompt scope, direct cross-agent prohibition, draft guards, model roles, and binding properties agree with their authored sources.

The health requirement needs reconciliation: §3.4 and the RDF claim forbid dispatch without a successful probe, while §7 permits omission of healthCheck and dispatch without a probe. Lean b17 makes the GET conditional. The first release must state whether probes are mandatory or conditional before runtime tests can establish conformance.

Default-deny prompt scope remains an enforcement obligation. An implementation must authorize prompt inputs before serialization or network dispatch, including context-derived instance data, rather than checking only returned writes.

Binding selection and dispatch

The next fifteen claims have nine matches and six whole-node mismatches. Selection and ordered dispatch claims target the absent agent.dispatch effect; three write-protection/event-boundary claims target absent agent.write. The no-retry invariant also has no authored domain identity.

The source contract still needs concrete execution details:

  • §5’s fail-fast sequence allows no subsequent step after failure, but Step 4 permits transient failures to restart at Step 2. State that retry exception explicitly and define whether prompt context is rebuilt for each attempt.
  • §4.1’s grammar fixes property order, while its prose declares all entries unordered. The canonical examples put envTag before healthCheck, unlike the grammar. Integrated parser fixtures must test both required properties and permitted ordering.
  • The probe is HTTP GET even for an ollama:// environment. Specify conversion to a concrete HTTP endpoint and path joining; a literal non-HTTP URI cannot serve as an HTTP request URL. Reserved gRPC behavior must remain distinct.

No actual inference call, parser acceptance, or environment selection was verified by this review.

Prompt construction, validation, and completion

The final seventeen claims have nine matches and eight whole-node mismatches. Missing effect and invariant owners recur in response validation, semantic validation, escalation, permission closure, rollback, auditing, and model roles. The semantic-validation reference to RFC-0009 §5 is valid: that section owns Evidence Obligations. It must not be treated as a missing domain identity.

An absent health probe must not skip prompt construction. §7 and the RDF claim say to proceed directly to Step 4; §5 requires Step 3 to construct and authorize the prompt before dispatch. Repair that control-flow edge explicitly and test that a binding without healthCheck still rejects out-of-scope prompt data.

Matching timeout, health-status, retry, and draft-emission claims are source evidence. Actual compiler fixtures and runtime failure tests remain required.