Repository navigation
Retain the 2026-09-26 sweep's 32 layer inputs (SHA256SUMS) for the verdict wave's GPT-6 re-votes - #363
Merged
seathatflowsinourveins merged 4 commits intoSep 26, 2026
Conversation
The verdict wave re-runs 27 missing GPT-6 fit votes and needs the run's own inputs, including lane limit 2's truncated slugs, not rebuilt ones. Digests equal the list shared with the coordinator (SHA256SUMS sha256 c026bdf6...d251). code-navigation.json names Sourcegraph and holds catalog pin commit ids, so the anchored pin-line gitleaks exemption now covers that exact file (test_d5: other fields and sibling files stay detected). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
Trading-lane ack (native-agent-stack-f5) at head 3e5a9e6.
|
seathatflowsinourveins
enabled auto-merge (squash)
September 26, 2026 19:11
Owner
Author
|
Handoff for A6 (landscape-sweep-20260926 manifest rebuild): GPT-6 fit re-votes for the 27 proposals refuted only by a missing vote
|
…grew it)
This PR's validate run 36265075411 was cancelled by the job limit ("The job has exceeded
the maximum execution time of 10m0s") after its 6257-test suite had printed OK in 534s; no
evidence or manifest check failed. The limit is exhausted repo-wide: main's validate job
took 598s in run 36263674462, and runs 36265091632 and 36269293122 hit the same limit on
2026-09-26. 20 minutes restores headroom; test_workflow_hardening still requires an
explicit timeout-minutes on every job.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Sep 26, 2026
…grew it)
This PR's validate run 36270619115 was cancelled by the job limit ("The job has exceeded
the maximum execution time of 10m0s") after its 6295-test suite had printed OK in 529s.
The same hunk as #363's commit 676734a (byte-identical, so either PR merges cleanly after
the other): main's validate job took 598s in run 36263674462, and runs 36265075411,
36265091632 and 36269293122 hit the same limit on 2026-09-26.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
seathatflowsinourveins
deleted the
claude/landscape-sweep-20260926-inputs
branch
September 26, 2026 21:08
seathatflowsinourveins
added a commit
that referenced
this pull request
Sep 26, 2026
…m) on the frozen 8-K extraction task (#359) * Preregister GPT-6 family tiering for mechanical 8-K extraction (not run) Adds a frozen, unrun routing test: which GPT-6 family member and reasoning effort matches gpt-6-astra at max effort on a deterministically scored extraction stage at lower token cost. Arms A0 (astra/max, control), A1 (astra/medium), S0/S1 (sol max/medium) and L0/L1 (luna max/medium) run the frozen codex exec command on li26's 360 filings in 25 greedy batches (at most 15 filings, prompts under 110,000 bytes; layout hash frozen). run_arm.py shares the host's three Codex slot locks, retries a failed batch once, stops on Codex's own usage-limit events (exit 3, LIMIT note) and resumes after the reset. analyze.py reuses li26's scoring and paired bootstrap and applies the frozen rule: routable when the 95% lower bound of the micro-F1 difference is at least -0.02, JSON-valid is at least 0.98 and every batch completed; the cheapest routable arm by billed tokens per filing (uncached input plus output, which already includes reasoning) is recommended. No arm ran and no provider call was made. Offline tests use a fake codex on PATH. The new files still need registration in manifests/evidence.json. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Amend the GPT-6 tiering preregistration after review, before any call One bounded review round (claude-opus and gpt-6-astra reviewers) on fbcca8f9. No arm ran and no provider was called; the arms, task, labels, batching, margin, bootstrap and billed-token definition are unchanged. - analyze.py: a candidate is recommended only when its cost is strictly below A0's. Unknown A0 usage makes A0's cost a lower bound (a candidate below it still routes, otherwise inconclusive_control_usage_incomplete) instead of silently dropping A0 from the ranking. Which usage-less calls make usage incomplete is stated once in the plan. analyze.py now verifies every frozen input before reading and records the map it verified. - run_arm.py: a call without a completed turn (outage, sign-in, rate limit, launch failure) or with a missing or unknown session record is unavailable: it uses no retry, publishes UNAVAILABLE and exits 4. unbounded_connection_retries is off, so an outage fails a call instead of hanging it. A finished arm with a failed batch exits 6. Abandoned calls are neutral. - Frozen isolation overrides, measured by isolation-probe/ (codex-cli 0.157.1 against a loopback fake provider in a network namespace, no credential): web search, the shell, sub-agents (agents.enabled=false; features.multi_agent=false alone leaves them), sleep, image viewing, goals, hooks, plugins and apps are off. Tool items in a call's session record or --json stream fail the call, so no tool-assisted reply scores. - Each call has its own CODEX_HOME in its private attempt directory with a link to the native auth.json (never a copy), a scratch HOME, TMPDIR and working directory outside the state tree, and an allowlisted environment; a run refuses while an /etc/codex configuration exists. - A state-wide launch gate: LIMIT is published while the limited call still runs and before its slot is released, and a dead runner's unpublished limit is published before any call. Tests: 39 offline tests with a fake codex (8 targeted mutations killed); scripts/validate.py and validate_convergence.py pass. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Register the GPT-6 family routing preregistration (hot-file commit) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Ignore the environment keys macOS adds to the fake codex itself (validate-macos fix) validate-macos failed test_codex_command_line_homes_environment_and_private_files: the fake codex (a Python child) reported __CF_USER_TEXT_ENCODING in its own environment. run_arm.py's call_env passes only CHILD_ENV_ALLOWLIST plus CODEX_HOME, HOME and TMPDIR; macOS CoreFoundation adds __CF_* (and VERSIONER_*) keys to a process's environment even when exec received an explicit one. The comparison now drops those names on darwin only, as CPython's own Lib/test/test_subprocess.py test_empty_env does (is_env_var_to_ignore); other platforms keep the exact set. run_arm.py and the other frozen inputs are unchanged. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Re-register the macOS env fix's test file (hot-file commit) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Raise the validate job's timeout from 10 to 20 minutes (the suite outgrew it) This PR's validate run 36270619115 was cancelled by the job limit ("The job has exceeded the maximum execution time of 10m0s") after its 6295-test suite had printed OK in 529s. The same hunk as #363's commit 676734a (byte-identical, so either PR merges cleanly after the other): main's validate job took 598s in run 36263674462, and runs 36265075411, 36265091632 and 36269293122 hit the same limit on 2026-09-26. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Re-register validate.yml after the timeout change (hot-file commit) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Scout <scout@local> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds
evidence/artifacts/landscape-sweep-20260926/inputs/: the 32 layer-input files exactly as the 2026-09-26 sweep read them. They were written 2026-09-25 21:30–22:44 local, before the run.inputs/SHA256SUMS(sha256c026bdf6…d251) lists names relative to its own directory, socd inputs && sha256sum -c SHA256SUMSworks.The verdict wave re-runs 27 missing GPT-6 fit votes, and it cites these originals rather than rebuilt inputs, which would differ in lane limit 2's truncated slugs.
code-navigation.jsonnames Sourcegraph (the SCIP index) and holds catalog pin commit ids. The existing anchored"pin"-line gitleaks exemption now also covers that one exact file.test_d5checks that other 40-hex fields in it, and the same pin lines in sibling files, are still detected.Checks:
validate.pypassed, andtests.test_gitleaks_configran 28, OK. gitleaks finds no leaks ininputs/.CI timeout fix (2026-09-26, workstation coordinator). The first
validaterun (36265075411) failed because the job limit cancelled it ("The job has exceeded the maximum execution time of 10m0s"). Its 6257-test suite had already printed OK in 534s, and every evidence and manifest step had passed, so the PR's content was not the cause. The limit is exhausted repo-wide: main's validate job took 598s in run 36263674462, and runs 36265091632 and 36269293122 hit the same limit..github/workflows/validate.ymlnow gives thevalidatejobtimeout-minutes: 20, and its registration is refreshed in the hot-file commit. Local checks at08926673all exited 0:validate.py,evidence_manifest.py --check,component_matrix.py --check,new_host_grand_list.py --check,build_ecosystem.py --check, zizmor 1.30.1--offline(no findings), and main'sverdict_review_gate.pyagainst this head. Eleven related test modules (workflow hardening/security, verdict gate and vendoring, gitleaks config, release pins, freshness pins, saturation ledger, promotion gate) ran 395 tests, OK.Lane consolidation: on 2026-09-26 the user shut down the other lane sessions and assigned this work to the workstation coordinator ("no other peers live, i shut them with the token save repos install into our native workflow and github workflow , harness update and sota foundation of our session as main task"); this quote stands in for the trading lane's ack.
SOTA sources
The files are the run's own inputs, built by the prototype
build_inputs.pyfrom the catalog layer files. There are no new external sources.jobs.<job_id>.timeout-minutes: "The maximum number of minutes to let a job run before GitHub automatically cancels it. Default: 360". The job keeps an explicit limit (tests/test_workflow_hardening.py requires one on every job), now set above the measured run time.🤖 Generated with Claude Code