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.
The authorized eight-trial primary pilot ran on clean Git HEAD
f6ebfebac965e681fa43365d5075007767984e05 using
benchmark/agent/corpus/primary_realistic.json. The immutable plan, all eight
raw rows, and the analyzer output are saved beside this file. Four control and
four treatment trials were attempted. All eight failed before a model answer:
codex-cli 0.137.0 could not decode its model catalog (unknown variant max,
expected none|minimal|low|medium|high|xhigh). Each raw row contains the
adapter error. There are no model tool traces or answers; no observed toolkit
calls or token usage. Billing cannot be inferred from the zero usage fields.
The analyzer reports valid=false, eight operational failures, no factual
attempts, and no quality estimate. The sum of trial durations is 43,828 ms,
within the 360-second wall budget. This consumed the registered eight-call
allowance as attempts; it does not justify an unapproved rerun. The treatment
could not demonstrate direct Go CLI use because Codex failed before inference.
Pinned settings: gpt-6-luna, model-version label gpt-6-luna, low reasoning
effort, direct toolkit mode, eight-trial cap, 360-second wall cap, 45-second
per-trial timeout, 12,000 observed-token stop condition, unpriced. The Go CLI
binary SHA-256 was
96c93890c25cba6e9d9ce75779eb4f9f52922f92598b6a55d29a63fcfd955e17;
the adapter SHA-256 was
c405f6a976599db2fd7042660aab822074f64ea975f770e1a41a8f6267dc9f09.
The SPEC revision was 0b87c52c6ef999286c745e19998fdfcd03d5dbee.
Exact command:
GOCACHE=/private/tmp/okf-agent-gocache go run ./benchmark/agent/cmd/okf-agent-eval run \
-corpus benchmark/agent/corpus/primary_realistic.json \
-plan /private/tmp/okf-primary-plan.json \
-rows /private/tmp/okf-primary-rows.jsonl \
-adapter /private/tmp/okf-eval-codex-adapter \
-go-cli /private/tmp/okf-eval-cli \
-model gpt-6-luna -model-version gpt-6-luna \
-settings '{"reasoning_effort":"low"}' \
-commit f6ebfebac965e681fa43365d5075007767984e05 \
-spec 0b87c52c6ef999286c745e19998fdfcd03d5dbee \
-repeats 1 -max-trials 8 -max-seconds 360 -max-tokens 12000 -unpriced \
-toolkit-mode direct \
-model-tool-access 'Codex read-only ephemeral temp with shell; Go CLI binary supplied'
The earlier auto-review refusal, before this more specific authorization, is
recorded separately in ../20260926-primary-blocked/README.md.