Repository navigation
Own local candidate snapshots and proof verification in LexLean - #6
Conversation
afflom
left a comment
There was a problem hiding this comment.
Maintainer self-review of the candidate ownership change (not independent approval).
Reviewed the actual snapshot envelope and full linked-IR reconstruction, compiler/semantic pin derivation, source/declaration/report binding, snapshot-to-verifier transfer, cancellation during spawn/read, stale dependencies, overlap refusal, and protocol/cleanup failure ownership. Checked that external reviewed events still refuse checked proof answers and that the export request selects the certified runner roots. No remote model calls or new non-LexLean runtime authoring are introduced.
Validation passed at 89ff2db6a9e6d68c7c2c63b5b84a3f01a693fbff: the unchanged pre-commit hook ran 243 tooling tests (4 expected skips), retained/bridge audits, all 68 API elaborations and kernel replays, exact axiom policies, all 27 production roots' preservation certificates, and generated publication checks. Actual generated-Rust runner integration accepted a contextual proof only after owned snapshot/verify execution and ordered report reads. A second case changed valid source between snapshot and verify: both compiler processes exited successfully and semantic identity stayed unchanged, but the changed source identity was rejected before the attestation read. Both cases left no child process. The final head's runner program digest matches the tested program.
Pull-request CI is pending.
The isolated project and pinned imported library remain caller-owned boundaries. Candidate construction, inspection, strategy planning, native research CLI integration and durable publication are not completed by this change. No Rule 30 result is asserted.
Research attempts need to obtain compiler identities and candidate IR from their own local execution before verifying a proof. The runner now accepts a candidate context and owns the complete snapshot-to-verification lifecycle in LexLean.
CompilerSnapshotreconstructs full linked IR from the public compiler snapshot and preserves the exact lexicon closure. It checks the expected compiler and recomputes the semantic digest.ResearchCandidateruns the fixed local snapshot command, reads its output, derives the verification pins, checks the contextual goal, and entersResearchProofin the same isolated project. Snapshot metadata never constitutes proof evidence. Cancellation drains pending replies and cannot cross into verification; overlapping work is refused and failed cleanup retains ownership.Unexpected replies while reading verification reports now halt the runner instead of becoming a candidate failure that permits redispatch. The attention catalog covers the four new modules and their modeled behaviors. The production request exports the runner begin/step component libraries alongside the controller and catalog roots. The compiler pin follows the v0.5-based snapshot CLI in UOR-Foundation/lexlean#439. The API file limit is 8 MiB because the expanded runner's measured Rust certificate B is 6,681,897 bytes. The acceptance hook and proof/axiom policies are unchanged. No remote models are used.
Validation passed at
89ff2db6a9e6d68c7c2c63b5b84a3f01a693fbff: the unchanged pre-commit hook ran 243 tooling tests (4 expected skips), retained/bridge audits, all 68 API elaborations and kernel replays, exact axiom policies, all 27 production roots' preservation certificates, and generated publication checks. Actual generated-Rust runner integration accepted a contextual proof only after owned snapshot/verify execution and ordered report reads. A second case changed valid source between snapshot and verify: both compiler processes exited successfully and semantic identity stayed unchanged, but the changed source identity was rejected before the attestation read. Both cases left no child process. The final head's runner program digest matches the tested program.Pull-request CI passed the complete acceptance gate and artifact upload.
Candidate construction, source inspection and imported-library ownership, strategy planning, native research command integration and durable publication remain implementation work. This change does not claim a proof of Rule 30 Problem 1 or completion of the research engine.