Skip to content

Bind research verification to exact candidate identities and types - #5

Merged
afflom merged 1 commit into
mainfrom
codex/research-binding
Nov 4, 2026
Merged

afflom merged 1 commit into
mainfrom
codex/research-binding

Conversation

@afflom

@afflom afflom commented Nov 4, 2026 •

Copy link
Copy Markdown
Owner

A successful verification report must prove the research request's actual contextual goal. Previously, the runner could accept an externally supplied checked answer after a successful job without comparing that goal with the audited declaration's type.

This change authors the binding and its lifecycle in LexLean. It recomputes the candidate semantic identity from canonical linked IR and lexicon data, selects a unique proof-bearing native declaration, reconstructs the selected source premise under its preceding binders, and compares exact core types. Refutation negates the premise within that context. Source origins match the request's recorded source names; they are not mistaken for artifact hashes.

ResearchProof owns verification and report reads before promotion. ResearchRunner routes proof jobs through that owner, refuses externally submitted checked answers, cancels stale work, and retains ownership after failed cleanup. The attention catalog includes the new modules and behavioral features. Canonical encoding rejects duplicate keys, malformed UTF-8, out-of-domain numbers, exhausted work and output limits; oversized keys and numbers are refused before costly comparison or slicing.

The compiler pin follows the v0.5-based native Host driver in UOR-Foundation/lexlean#439. Compiler/target metadata and official locks are refreshed. The API certificate-part limit is 6 MiB, based on the measured 4,541,007-byte part; proof policy and the acceptance hook are unchanged. The only legacy Python edits update two target-digest constants. CI and runtime use no remote models.

Validation completed on 4230698:

  • The unchanged pre-commit gate passed: 243 tooling tests with 4 expected skips, retained-library/bridge checks, 69 reusable research replay receipts, all 64 API elaborations and kernel replays, exact axiom policies, production-preservation certificates, and verified PDF/mirror checks. The push hook reused the accepted tree.
  • The native R30 entry was rendered through the official package renderer from the gate-accepted R9 program and compiled. Help/version and error paths passed; full and open-only frontier output matched the existing CLI byte for byte. Missing and malformed registers returned their modeled errors.
  • Behavioral evidence covers compiler-produced native types, contextual closure and negation, substituted declarations and targets, named-source selection, bounded canonical encoding, and proof lifecycle ownership.

PR CI passed: Main Branch Validation run 41910715189. Maintainer self-review is recorded on the unchanged head.

This completes candidate/type binding within the owned verifier. Source inspection must still establish provenance, and native strategy construction, 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 native research engine.

@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 self-review of 4230698 (not an independent approval): checked candidate selection, semantic/build identity binding, contextual premise closure and refutation, report ownership, stale/cancel/cleanup handling, and the generated native entry smoke evidence. The unchanged pre-commit gate accepted all 64 API modules, kernel replay, exact axiom policies and production certificates; the push hook reused that accepted tree. No merge-blocking finding remains for this change's stated scope. Source provenance, strategy construction, native research wiring and durable publication remain explicitly unfinished; this PR does not establish Problem 1. Merge remains conditional on the PR CI result.

@afflom
afflom merged commit 51c8ce1 into main Nov 4, 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