Repository navigation
docs: proposal for verification results reused across runs (UOR-native, piece 3) - #508
humuhumu33 wants to merge 1 commit into
Conversation
A design for owner acceptance: a content-addressed store of verifier process results, keyed by the address of everything a process reads and runs, written only by release-line pushes and read by pull-request runs, with the SPEC.md §9.4 wording it needs. Adds this one file only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
Bounded proposal review at2d6b52c. Exact process-input reuse, retaining current judgment instead of cached verdicts, and explicit estimated savings are useful directions. D1 remains an owner policy decision; this review does not admit cross-run proof reuse or change the existing full final-head/same-run requirements.
Three clarifications are needed before that decision is concrete:
- Lines14/87/93 promise fresh proof on every release-line push and release. Current SPEC §9.4 (lines692/696 at26e5) deliberately skips the whole gate on a same-tree merge after a passed PR gate, and release.yml consumes that evidence without another gate. If the PR gate reused historical records, the current path can release those records without the promised fresh run. Coordinate the existing plan/release evidence eligibility with section6: specify when no-reuse evidence is required, or accurately describe the historical evidence policy. This is not a request to add a redundant release build/gate.
- D6 excludes lexlean bytes so the current judge can rejudge old records, while lines81/93 require those records to name current executables, including lexlean through #498. Define original child-process producer identity separately from current judge/publication identity, and reconcile key/unchanged-record wording. Do not silently replace the old producer identity. #498's selected-compiler binding findings remain unresolved.
- Lines63/82–86/93 need an explicit trust boundary. Hashes and repository/run/commit strings inside records do not authenticate a consistently forged entry. Bind provider-verified event/ref/conclusion/provenance to the exact immutable artifact/index through the existing custody mechanism, and state that trusted release producers remain an authority assumption. One selected replay is useful diagnostic sampling; it cannot guarantee rejection of an unsampled consistent forgery. No duplicate replay of every record is requested.
Please clarify these in the proposal before requesting D1 approval. #498/#506 remain draft dependencies with their existing findings; neither their framing nor same-run partition equivalence proves cross-run hermeticity. No implementation, writer, gate replay or trust change was performed. The proposal stays draft and unaccepted.
|
Docs-only and clean, but kept draft: D1 (SPEC §9.4 wording) needs the owner's decision first, and marking ready would start a ~55-job run for a one-file proposal (see the consolidation request on #492). Will land with the next documentation carrier. Generated by Claude Code |
…, record D1 One reviewed documentation carrier, as requested on #492. D1 (SPEC.md 9.4 reuse of verification results) is decided: not adopted now. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
|
D1 decided from SPEC §9.4 and the #106 plan: not adopted now; §9.4's no-reuse gate stands, since this changes what a PR run verifies and the savings are unmeasured. The file is folded into the consolidated docs carrier #492 (with the decision recorded in it) per afflom's one-carrier request. Closing as superseded by #492. Generated by Claude Code |
Proposal only: this adds one file,
docs/proposals/verification-memo.md. It is the design for piece 3 of the UOR-native work: verification results reused across runs by exact address.Ilya decides D1, the SPEC.md §9.4 wording in section 6 of the file. It changes what the gate verifies. On a pull request, a reused verifier process would have been proved by an earlier
release/*push on byte-identical inputs, not by that pull request's run. Every push tomainorrelease/*, and every release, still re-proves everything.The other decisions, D2–D8, with a recommendation on each:
lexleanexecutable is in the key;What the file covers: the key, exactly; why reuse is sound; what it refuses (a stale store, a changed toolchain, a corrupted entry, a consistent forgery); and estimated savings from full gate 43046966603. Measured numbers come with the implementation.
It builds on: #498 (toolchain addresses) and #506 (
sha256:address scheme, AR-15).validate-modelpasses with the file.🤖 Generated with Claude Code