Repository navigation
Conversation
Implements CL 3b-3 of docs/elaborator-containment-design.md: - Bump verification receipt schema_version from 2 to 3. - Introduce a closed sorted VerificationTaint set (sorry_in_proof, unpinned_lean, unsandboxed_elaboration) governed by the Uniform Flag-Driven Taint Law across passing and all failure receipts. - Wire global --allow-sorry-in-proof into TargetAuditRequest and fail-closed validate_sorry_declarations (cross-checking both audit.sorryDeclarations and audit.axioms for sorryAx). - Surface elaboration_sandbox on FailureReceipt once sandbox selection completes and fail closed on missing or flag-contradicting sandboxes. - Gate --export-lean-lib full proof .olean export strictly behind all post-audit binding, provenance, sandbox, and taint checks. - Update CLI, integration, mvp_e2e, Starnix CI, and docs to Receipt v3. Source: https://github.com/joshlf/anneal/tree/bf7c496a225b19faa57f1bcd4482733f5507e72a Agent-authored-by: AI agent acting on joshlf's behalf
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7341798621
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| Ok(audit) => audit, | ||
| Err(e) => { | ||
| if args.export_lean_lib && e.code == 1 { | ||
| let _ = checker.export_target_oleans_gated(&work.join("lean-lib"), false); |
There was a problem hiding this comment.
Propagate partial-export failures
When --export-lean-lib is requested and the audit rejects the proof with exit code 1, this discards any error from export_target_oleans_gated. An I/O failure or an unexpected symlink/non-regular artifact therefore still produces an ordinary proof_failed receipt and may retain evidence without the promised partial Lean library, masking a potentially hostile elaborator side effect. Preserve the previous behavior of reporting the export failure as tool_failed.
Useful? React with 👍 / 👎.
| `"none"`; omitted on pre-sandbox failure receipts), optional | ||
| `"sorry_declarations"` (present iff `--allow-sorry-in-proof` is active), and |
There was a problem hiding this comment.
Limit the sorry-declarations guarantee to successful receipts
The documented “present iff” contract does not hold for failure receipts: FailureReceipt has no sorry_declarations field, so passing --allow-sorry-in-proof with a malformed request, staging failure, rejected specification, or later tool failure emits the taint atom but omits this key. Receipt consumers following this schema description can therefore reject valid failure receipts or assume missing metadata indicates the flag was inactive; either qualify this as a successful-receipt field or include it consistently where promised.
Useful? React with 👍 / 👎.
Human review surface
Implements CL 3b-3 of docs/elaborator-containment-design.md:
unpinned_lean, unsandboxed_elaboration) governed by the Uniform
Flag-Driven Taint Law across passing and all failure receipts.
fail-closed validate_sorry_declarations (cross-checking both
audit.sorryDeclarations and audit.axioms for sorryAx).
completes and fail closed on missing or flag-contradicting sandboxes.
post-audit binding, provenance, sandbox, and taint checks.
Layer 19 of 19 in the migrated Anneal verification stack under
anneal/.Public Anneal source checkpoint.
Authored by an AI agent acting on joshlf's behalf.