Historical report, 26 September 2026. This reading edition preserves the findings and limits of the original run; it is not current product guidance. Original report, revision 61e75e9aa9a8719dfb480bf1f5226553b3a21d70.

Exploratory smoke: no model result

This two-trial smoke used one repo-derived case (repo-default-okf-version), gpt-6-luna with low reasoning effort, a 120-second wall limit, a 20,000 observed-token limit, and at most two model calls. It ran against a dirty worktree before the corrected implementation was committed. It is explicitly invalid for task 012 acceptance.

The treatment arm invoked the real Go CLI three times (validate, then parse for both concepts). The original treatment row records three tool calls. Both model attempts failed before inference: Codex CLI could not write its state database under the filesystem sandbox and then could not initialize its app-server client. Therefore there are no factual observations, token usage, or accuracy estimates.

plan.original.json and rows.original.jsonl are the exact files written by the harness version used for this smoke. plan.json and rows.jsonl add fields introduced after the smoke (spec_sha256, adapter_sha256, mode, exploratory marker, zero/unknown trial elapsed, zero observed direct CLI calls, and empty tool-output copies) so the current analyzer can recount it. Those added values are post hoc and cannot be treated as preregistered evidence. report.json is the resulting invalid report. corpus.json is the frozen one-case input.

An escalated retry was rejected by automatic approval review because the external model could access repository content and shell and write state outside the workspace. Do not use a different home directory or another provider to bypass that rejection. A live pilot needs explicit user approval for the external model run, then a clean corrected commit and a new plan.