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.

Manually prepared OKF bundle → independent consumer — 2026-09-26

The frozen protocol ran from clean commit dd8989ba36b8e27ac5762e2e0984cac7f88e8738 with gpt-6-luna at low reasoning effort. The one synthetic queue-mode reversal in writer_consumer_baseline.json (SHA-256 3fd56d975fa6570df6db81a939eea72d51178034a559882a52b1bb753baf522a) compares raw records against a human-prepared OKF bundle containing the same answer and evidence ID. The offline Go validation test for that bundle passed. A fresh consumer invocation received each arm; no AI writer, Go backfill, or upkeep checker participated. The blinded adapter (SHA-256 8f132bc3d12aeb1ea78d74642156427c1921a35b28d60f12b1296354a0b44ea9) sent the model the question and arm artifacts without the harness labels or gold answer.

The saved plan, raw rows, and report are byte-identical to the original run files in /private/tmp. Reanalysis reproduces the report byte-for-byte:

GOCACHE=/private/tmp/okf-agent-gocache GOPROXY=off GOSUMDB=off GOTOOLCHAIN=local go run \
  ./benchmark/agent/cmd/okf-agent-eval analyze \
  -corpus benchmark/agent/corpus/writer_consumer_baseline.json \
  -plan benchmark/agent/runs/20260926-manual-writer-consumer-v3/plan.json \
  -rows benchmark/agent/runs/20260926-manual-writer-consumer-v3/rows.jsonl

Both planned calls completed and the report is valid=true: control 1/1 correct and treatment 1/1 correct. Both consumers answered exactly B and cited decision. Neither was stale, wrong, ungradable, or a refusal; there were no operational failures or model tool calls. Thus this case shows a consumer can answer from the manually prepared bundle, with no measured quality difference against the raw-record control.

Observed usage was 30,291 input + 52 output = 30,343 tokens across two calls; 27,136 input tokens were reported as cached and are already included in input usage. Summed model elapsed time was 11.326 seconds. USD cost was unavailable, so the report’s zero cost_usd with zero cost_known_trials does not mean the calls were free.

This one-case, one-repeat synthetic baseline tests independent consumer reading of a valid, human-prepared bundle. It does not measure AI writer quality, the task-010 backfill or task-013 checker, or a general benefit from OKF. Earlier v1/v2 protocol drafts had no model observations and are outside these denominators.