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.
upkeep-live-20260926-v1)The preregistered analysis failed: the committed upkeepstudy analyze command exited 1 because the checker-on consumer for upkeep-truncated cited handler: The captured diff is partial, so rollout status cannot be established. instead of the artifact ID handler. Its verbatim error is in original-analysis-error.txt. Therefore this run has no valid primary result under the frozen scoring contract. report-posthoc.json is a corrected, post-hoc sensitivity analysis; its valid: true describes that corrected analysis, not the preregistered v1 result. Do not relabel v1 valid or use it to claim a checker benefit. Task 013 still needs a freshly preregistered live rerun with an explicitly specified citation policy and grader.
The model run itself completed: plan.json SHA-256 27cc232a283d0b9993108684a942421a1906fb30e3215c2e974f63388db41327; rows.json SHA-256 5be692ba2f2a5e54a703586b55208d7bae9643fb3edeace09e764144f16f6dec; audit.json SHA-256 403a447e386271b0a69b1ddc3832626be0e6e1c40fbebcd07522af80ecf7c6e4. The completion marker contains the latter two hashes, which match the saved files. The audit pins runner a0565afc973e6c1ef3d0a14ca732069d9a09416b6503e0d7f595688143a44b45 and adapter 3470b13c8e8d31fee44e2de6241a1af917a75bc02b5cb0f4c9f87c8cdbc8e9ea. It records 10 model calls, 155,378 input plus output tokens, zero unknown-usage calls, 10 distinct session IDs, both expected cases and arms, and a Go CLI validation event for every final bundle. These operational checks establish complete artifacts, but cannot repair the original scoring failure. USD cost is unavailable.
For diagnosis only, the post-hoc grader reports:
| Outcome | Checker off | Checker on |
|---|---|---|
| Concept document updated | 2/2 | 1/2 |
| Independent consumer correct by the strict grader | 1/2 | 0/2 |
| Wrong answer | 1/2 | 1/2 |
| Ungradable answer with unsupported citation | 0/2 | 1/2 |
The upkeep-reversal checker-on writer left decision.md saying Mode A after changing the code to Mode B. The final checker status remained needs_review; the independent consumer answered Mode A. Checker off updated that document to Mode B and its consumer answered correctly. For upkeep-truncated, both consumers answered “Insufficient evidence.” The checker-on answer nevertheless failed strict citation grading because its evidence field contained a prose citation unsupported by the registered artifact-ID contract. Its final checker status was reviewed_updated, which records a changed concept file, not factual correctness. The post-hoc summary therefore keeps it ungradable rather than silently counting it correct.
This is a two-case descriptive observation, and the checker-on arm combines a reminder, checker feedback, and an extra revision opportunity. It cannot isolate a checker effect. The observed pattern supplies a regression case for the grader and for writer behavior; it does not satisfy task 013’s live acceptance criterion.