[Dong1017]

One bounded CPU KV editing comparison

Refs #248; activates the reviewed launcher from #276 for the owner-authorized A2/A3 pair, not a new experiment framework or a result claim. The exact public freeze and rate/accounting evidence govern this configuration.

• Bind the existing credentials only in the trusted execute step after secret-free preparation. Candidates, packaging and final grading receive no key. Keep manual, owner-only, main-only, first-attempt execution and same-host preflight.
• Freeze identical A2/A3 common inputs; A3 alone receives the complete nominated K. Preserve the qualified visible workbench and independent 77-trace final correctness/8-pair timing gate. This is exploratory synthetic near-transfer; timing is not held out and no service or K-efficacy claim is made.
• Charge only PR275: at most 16 requests and $162.72 full-pair reservation; preserve its $200 lifetime cap, the $1,000 Shanghai-day cap including $491.35 retained obligations, and the $1,000 pilot ceiling. Unknown usage or abnormal completion stops further requests without refund/relaunch. Retain first-run bounded evidence, including failures and partial admission.
• Reuse the already published original hosted actor archive and receipt. Register prospective inputs in the existing evidence inventory; there are no live model outputs in this PR.

Execution has not occurred. This freeze expires at 03:00 UTC on 31 October; its required 22-minute runway makes 02:38 the latest admission. Review/CI does not extend that window. Dispatch, live price/ledger/current-main checks and same-host preflight remain action-time requirements; a stale plan fails closed.

Source binding and validation boundary

Local activation commit ccadb62797c278ae62cc018e1f90dba51e0dfcee has tree bf3f8051e3ed3a51cc75e87f3bd49359d0539f1e: 23 files versus merged #276/main, with no controller, provider, sandbox, admission, qualifier or grader source changes. Published head 55481bb6af3abc4c86b5f5f7807fa7f2668f3d79 preserves this exact tree on the current main parent. The locator binds the verified public comment, not a supplied dispatch input.

• Activation changes: 100 launcher tests and 17 inventory tests passed; after the final locator binding, 47 policy/inventory tests passed. Whole-tree Ruff and diff checks passed.
• Actual public-comment parsing, qualification/source matching, runway and data-only staging passed locally. These checks did not execute a candidate, read a key or mutate a ledger; they are not same-host execution preflight.
• The unchanged merged software baseline previously passed 3,169 tests, 12 skipped and 284 subtests. That aggregate is reused only for unchanged source; it is not labelled a new full aggregate of this configuration.
• Hosted qualification 41154288295 is reused with its exact source/protocol bindings. Prior raw audits cover that unchanged closure, not a fresh recomputation of this prospective comparison.
• Independent source/configuration review and final real-public-comment/locator binding found no open findings. Current-head Ready run 41158813655 passed both Python 3.10/3.11 full checks, both installed-package checks and model-free task/actor qualification. Its test-merge tree matches the reviewed activation tree above; draft lint and declaration policy also passed. The opened-event Ready run was superseded/cancelled on marking Ready, not counted as PASS.
• Native code review completed on the current head with no findings or unresolved findings; native security review was not run. This is software/configuration validation, not live same-host preflight or an efficacy result.

Authorship

Actual implementation and independent review: Codex / gpt-6-astra / xhigh. Publication preserves the local tested tree through an API-created wrapper; authorship, source review, CI and operational acceptance remain distinct.

{"schema_version":1,"head_sha":"55481bb6af3abc4c86b5f5f7807fa7f2668f3d79","human_only":false,"material_ai_authors":[{"tool":"Codex (bounded KV activation configuration)","model_id":"gpt-6-astra","effort":"xhigh","role":"author"}]}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-31T02:12:32.100899Z	55481bb	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.