Skip to content

Own native research cohorts with shared libraries and retained evidence - #9

Merged
afflom merged 1 commit into
mainfrom
codex/owned-research-session
Nov 10, 2026
Merged

afflom merged 1 commit into
mainfrom
codex/owned-research-session

Conversation

@afflom

@afflom afflom commented Nov 10, 2026 •

Copy link
Copy Markdown
Owner

Candidate verification depended on an external caller to prepare projects and retain proof evidence, and each attempt required a fresh library build. This change adds native LexLean commands that own an individual attempt or a bounded cohort sharing one prepared library.

ResearchWorkspace, ResearchProject and ResearchSession own pinned checkout/build, exact-goal source construction, snapshot/kernel verification, evidence retention and release. ResearchRecord captures exact inputs and consumed reports. ResearchEvidence verifies the canonical record and Git references retaining both the record and its library commit before returning checked evidence. Interrupted tools and uncertain cleanup retain the workspace.

ResearchCohort runs up to 64 distinct candidate serials sequentially with identical library configuration and tool identities. Each candidate retains its own request, directory and evidence. Tracked sources are rechecked between attempts without rebuilding. Ordinary candidate failure permits continuation; uncertain execution stops the cohort. Only the cohort releases the shared library. A library-preparation timeout marks the result cut, even when no candidate completed.

ResearchCommand.step and ResearchCohortCommand.step are generated native Host executables with bounded, exact JSON admission and structured per-attempt outcomes. A completed cohort can contain failed candidates; exit 0 requires all checked results and no cancellation. Terminal commands cannot restart. The supervisor also reads the monotonic clock before spawning: the Host spawn response contains an OS PID, not a timestamp.

Session operations remain executable internal components, certified through the command roots without duplicate standalone exports. The larger generated certificates require raising the per-file source limit from 8 to 16 MiB and aggregate source limit from 64 to 96 MiB. Removing redundant session exports keeps the batch within that 96 MiB bound. Proof, replay and axiom requirements remain unchanged.

The compiler pin advances to 15c667bd36604c94c28e39502edec3fcad38e0ce on LexLean PR #439, based on the current release/v0.5. This fixes extraction JSON literals being rewritten or rejected as diagnostic paths, using generated JSON escapes without relaxing path checks. The regression, full NE-01 conformance, formatting, Clippy, generated-file consistency and model audits passed upstream; its broader release gate remains outstanding. Locks and PrismPM compiler identities are regenerated together.

All new implementation and behavioral traces are LexLean. The attention catalog registers the modules and their features. No remote model is introduced and the acceptance hook is unchanged.

Validation: the unchanged pre-commit gate passed for 7e99c8418cade772c10d79be2731c193de7cd31a, including all 91 API elaborations and kernel replays, the exact axiom audit, and preservation, Rust-rendering and composed certificates for all 30 production roots. API attestation: c0571492292a055cdfe7fb903010a19eec731b922903f1b7f07e4ea83a8d3754. The accepted cohort program (be4b861f22a7164dbc8d9a2124825577dcf0ade464352d8b55a67ba29b5f91cd) and certified Rust prefix match the native execution evidence exactly. Native generated-Host diagnostics demonstrated failed/checked/checked outcomes with one library build, two retained evidence records and one worktree removal. A signal immediately after evidence-reference publication retained the library and evidence, withheld a checked result, and prevented the next candidate from starting.

This supplies owned execution and bounded library reuse. Autonomous candidate generation, imported-premise inspection, durable restart recovery and claim publication remain unfinished. It does not claim a proof of Rule 30 Problem 1 or replacement of the complete legacy engine.

Final acceptance: CI run 42786571256 passed on the reviewed head. Merged as 05f904dbbcada29adfbd77e7367c47cf30cf3fe3; its tree 36036125b8aaf348f2fd3a02e82dc5734cc0be2e exactly matches the locally accepted tree. The pre-push hook reused that accepted tree with the same proof tools. Maintainer self-review is recorded; no independent review is claimed.

@afflom afflom left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maintainer source review (not an independent approval):

  • Admission checks exact bounded JSON, complete requests, all tool identities, shared workspace configuration and distinct candidate serials before preparation.
  • Candidate request identity is separate from library ownership; source directories and retained records bind the candidate, while only the cohort releases the shared library.
  • Candidate failure permits continuation; interruption/protocol/cleanup uncertainty stops dispatch and retains files. Tracked library sources are rechecked before reuse.
  • Checked status still requires the exact contextual sequent, kernel replay, allowed axioms, candidate identity, canonical evidence readback and retained library reference. Neither exit zero nor a stored record grants proof authority.
  • Timeout during library preparation marks the cohort cut; completed batches can contain failures and exit zero requires all checked outcomes.
  • Normal and signal-after-publication native diagnostics passed. The accepted program and certified Rust prefix match both diagnostics exactly. CI remains required before merge.
  • The unchanged hook validates the full tree. Removing unused standalone session exports avoids duplicate certification; the native command roots retain their behavior. The project source limits increase from 8/64 MiB to 16/96 MiB for generated certificates; no proof or replay requirement is weakened.
  • Candidate generation, imported-premise inspection, durable recovery and claim publication remain explicit future work. No Rule 30 proof or complete replacement of the legacy engine is claimed.

@afflom
afflom merged commit 05f904d into main Nov 10, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant