Skip to content

Use supplied environmental context to select one research attempt - #16

Merged
afflom merged 1 commit into
mainfrom
codex/environment-research
Nov 15, 2026
Merged

afflom merged 1 commit into
mainfrom
codex/environment-research

Conversation

@afflom

@afflom afflom commented Nov 15, 2026 •

Copy link
Copy Markdown
Owner

Scoped release disposition (2026-11-15)

The implementation owner and independent source/acceptance reviewer recommend integrating #14–16 for a bounded native CLI release. This explicitly supersedes the earlier draft hold below requiring full upstream compiler acceptance, and the earlier no-merge/no-release request, for this scope.

The combined source is commit ef0f7a9ae5ef405e0b7dce37ee5d242d17e16b52, accepted tree ae2a9621b1a5765c157ca4cf201ead021918656e, using exact native compiler candidate 9fd01c31e40bb29577833825f902fafcceedb49a and Lean 4.32.2. Rule30's full consumer qualification and exact-head hosted gates passed (#14, #15, #16). Full upstream compiler acceptance remains incomplete; this decision does not declare it complete.

All documented custody, interruption and archival limitations remain in force. In particular, archival case 73116 remains boundary_credited=false; only the supported interrupted-open observation is credited. There is no claim of durable archival, exactly-once/restart recovery, universal descendant cleanup, scheduled Engine integration, synthesis, crawling or an inspector replacement. #13 and the optional query-to-research coordinator are separate work, not prerequisites for this scoped CLI release.

Native asset publication is still pending exact-byte packaging and custody verification. The retained tested binaries are x86_64 GNU/Linux debug builds, dynamically linked and unstripped, with GLIBC symbol requirements through 2.39. Other platforms and older ABI baselines are untested. Each distributed command must retain its own qualified production/build provenance; a GitHub tag by itself does not deliver these native assets.

The validation record below is preserved as history, including the superseded draft wording.

Supplied environmental context can now choose one real research attempt from 1–16 caller-supplied candidates. The new native command validates the complete query plan, context and candidate pool before effects, scores bounded cues deterministically, and passes exactly one unchanged winner to the existing ResearchSession lifecycle. A failed selected proof does not try another candidate or refute the task.

This draft is stacked on #15 (96c1c3a9a12130da34458eea8972b04486417e8c). It preserves existing memory interfaces, controllers and all plan obligations. It depends on native LexLean candidate 9fd01c31e40bb29577833825f902fafcceedb49a and Lean 4.32.2; full upstream compiler acceptance remains outstanding. It does not include #13’s inspection/publication work.

Admission compares every task field, including status, age and weight, and requires consistent workspace, tool, snapshot, serial and execution bounds. Observations must be a unique ordered subsequence of expected source keys, match the query, and pass content/digest/provenance checks. Every candidate receives pure construction preflight, including unselected candidates. Strict target validation checks the entire nodes/root container and every node while preserving legitimate unused nodes and nonzero roots. Only supplied candidates for closed named goals are admitted.

Validation binds commit ef0f7a9ae5ef405e0b7dce37ee5d242d17e16b52, accepted tree ae2a9621b1a5765c157ca4cf201ead021918656e, and reviewed input tree a6289a4415fae857b71b6fdb51d9b78f522a8bb1. Only normal gate-generated PDFs and the PDF manifest differ from the reviewed tree; authored source/configuration/lock bytes are unchanged.

  • Full-profile production passed in 6,120.271 seconds. Bundle 6412781236111b460fbf4e66963a57d1584e26f1f2d620596771acbfb8dab4bf passed artifact verification and offline native build. Executable SHA-256: d8f1861a61c0d6f4d20b9ea17c7f497f56475f2fd5ba3d401cb2b455c091547a.
  • The separate unchanged normal commit gate passed: 243 tooling tests (four skipped), core/bridge audits, exact-input library/paper reuse, full API verification, and PDF checks. API verification took 6,220.563 seconds: 164 elaborations, 164 kernel replays, 164 audit processes, extraction, and 117 Lean/99 checker preservation-certificate processes. Its attestation is 22e24ad7baf4a6de5aabc3b66b2236fc1c07774ebe12d1b4571474e499c1c813.
  • Declaration policies: 5,786 checks comprising 3,288 none, 2,490 exact and eight allow, including the new consumer claims. Generated preservation certificates retain the compiler’s existing policy, including Classical.choice; this is not a choice-free certificate claim. Neither successful production nor the gate increased OOM counters. Limits and pins are unchanged.
  • The exact native executable passed 91 admission cases / 378 assertions, including complete valid-envelope mutations, malformed unselected nodes, strict target validation, bounds and full-task mismatches.
  • Actual Add atomic multi-source environmental memory plans #15-produced contexts selected different candidates with an otherwise identical task/pool. One selected proof genuinely verified, retained matching evidence, invoked one verifier and removed its workspace. The alternate context selected a structurally valid invalid proof, which failed verification with no second attempt, no refutation claim and no retained proof evidence.
  • Completed faults cover combined tool/cleanup-tool unavailability, evidence-write refusal, the first hash-output readback refusal, owned-job deadline, preparation interruption, active-verifier interruption and cleanup interruption. Durable receipts distinguish retained workspaces, absent processes and zombies.
  • Archival case 73116 remains boundary_credited=false, with instrumentation error and harness exit 130. Independent retained-trace review establishes only that the owned consumer was terminated after its archival-open attempt was observed and that the open was interrupted without successful completion or a checked result. The original five-check case is not claimed passed; partial/completed writing, durable archival and the stronger frozen-kernel predicate remain unverified.

Failures remain retained: the first production run had no terminal result and was backed up before one reviewed fresh invocation; external admission/oracle errors were corrected under independent review without changing product source or repeating the successful proof. The first commit-gate attempt failed the orphan-child tooling check under a non-reaping PID 1. The unchanged test passed under the existing Tini reaper, followed by the successful unchanged full gate. No hook was bypassed; pre-push reused the exact accepted-tree receipt.

Context is supplied, self-consistent and unauthenticated; observations are not proof authority. Cue/source bytes never become commands, goals or proof terms. This command does not acquire data or access the memory store. The successful-proof trace showed no access to the two specified private stores under no-alias/no-inherited-store-descriptor assumptions; its syscall filter does not prove global lock absence or exclude arbitrary descriptor transfers/untraced escapes. Internal session cancellation is distinct from native OS interruption. Caller deadlines do not establish global descendant termination; crash-after-archival outcomes remain uncertain and must not trigger automatic retries.

Keep this draft pending review, exact-head CI and upstream compiler acceptance. This is one supplied-context-driven attempt, not synthesis, crawling, scheduled Engine learning, an inspector replacement, merge, release or deployment.

@afflom
afflom changed the base branch from codex/environment-plan to main November 15, 2026 20:18
@afflom
afflom marked this pull request as ready for review November 15, 2026 20:18
@afflom
afflom merged commit 7c76a54 into main Nov 15, 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