delegation-and-review: coverage before clearance — an all-clear binds only what the review actually covered

[F-e-u-e-r]

Summary

One new delegation-and-review §3 rule — coverage before clearance — placed immediately after the packet-errors rule as its declared pair (claims before findings; coverage before clearances), plus its evidence package. One canonical file (+58, pure additions); 6 evidence files. Commit 81945a8 on main-tip 0e92d4a.

Spine: a clean verdict is never evidence about material the reviewer did not receive or otherwise verify.

Why (adjudicated PARTIAL-GAP)

A falsification-first orientation showed every existing rule passes this counterexample: an honest reviewer PROCEEDs over a silently-truncated packet (a pagination cap dropped one requested file's diff) and the orchestrator reads the verdict as full-requested-scope clean — a false clearance with zero rule violations. The dispatch-side scope declaration only declares; the completion-claim audit is dishonesty-bound, execution-deliverable-bound, and claim-side; the packet-errors rule fires only when crediting a finding; the settled-tree protocol guards motion, not assembly gaps.

The rule (owner-locked semantics)

• Invariant: credited clearance scope ⊆ actually covered scope, reconciled over three distinct concepts — REQUIRED scope / AVAILABLE material / COVERED scope. No strict path-set equality; no machine-readable manifest; granularity follows the dispatch's own declaration.
• Missing required material → UNREVIEWED, never clean; the reviewer's silence cannot clear it; sole carve-out = the orchestrator's own independent verified-unchanged ground truth.
• Extra paths → authorized producer output, or a discovered dependency to escalate/re-scope, or outside the clearance — never auto-fraud, never silently absorbed.
• A subset reviewer's PROCEED never inflates beyond its subset.
• Direct-access carve-out: trusted frozen-tree/repo access satisfies availability — availability is not prompt-embedding (this keeps the rule from degrading into attachment-equality).
• Truncation/pagination is a dispatch coverage defect, never reviewer fault.

Review

Design gate, dual-blind, mutually blind, isolated dirs, identity from tool banners: r1 = gpt-5.6-luna (max) PROCEED + gpt-5.6-sol (max) PROCEED — 2/2, zero findings, all ten mandatory axes answered line-anchored by both, and both explicitly confirmed correctness requires no new runtime path-set tooling.

Landing fidelity

The reviewed rule and Provenance blocks landed byte-verbatim (machine-proven: each block exactly once, full diff's line-multiset == exactly the two reviewed blocks, zero deletions). DECLARED-LANDING-ADAPTATIONS: NONE. Neighboring doctrine byte-unchanged (packet-errors, completion-claim audit, settled-tree protocol blob-identical, both recently-landed §3/§4 units untouched). Static controls 7/7 (EXTRA · MISSING-CLAIM · UNCHANGED · AUTHORIZED-DERIVED · DISCOVERED-DEPENDENCY · SUBSET · DIRECT-ACCESS). checks.py green (invisible-Unicode sweep included).

Probe debt

Behavioral transmission/effectiveness remains unprobed. The rule carries exactly one new unprobed marker; provenance class = counterexample-derived / design-reviewed; the future behavioral probe routes to the standing #115 queue.

The sibling queued candidate (evidence re-read-not-regenerate) remains LOCKED — zero of its content here.

🤖 Generated with Claude Code

https://claude.ai/code/session_0132RthrKSsMkywcEwtkXhkx