Datasets:

---

alexshpunt

/

explicit-edit-benchmark

Like 5

Intentional duplicate for variance attribution (second observation of an existing configuration).

This run deliberately repeats PR #164 (explicit-edit-2026-10-08T07-06-25.448Z, score 0.956858 = 213/226 first-exact, 226/226 final) with the identical configuration: harness pi-better-edit@2.9.0 (pi-better-edit/npm), model opencode-go/muse-spark-1.3-contributor, thinking low, --concurrency 10 --timeout-seconds 120, oracle recoveries 5. The lane pre-flight rule that refuses same-harness+version+model+thinking duplicates was overridden this once because the point of this observation is variance attribution: PR #164 scored 0.956858 (213/226 first-exact) while an earlier build on this route scored 1.000000, and the question is whether the gap is run variance or the build.

Deviations: none from the prior run's config shape (new config file only; /scratch/config-290.json untouched). Smoke on replace-all-10-plain passed. No normalized rows edited; failures, timeouts, and recoveries preserved.

Maintainer note: this repeat is confounded by sandbox resource exhaustion and should NOT be merged as a model result. 327 attempts failed with exitCode 1 and zero tool calls (bwrap: Can't fork for pid 1: Resource temporarily unavailable); 71 trials exhausted. Recomputed score 0.4934 (110/226 first-exact, 116/226 final) is infra-driven, not model behavior. A clean re-run will follow in a separate PR; the variance question vs PR #164 remains open.

Closing this candidate as withdrawn at the contributor's request above. The contributor reports sandbox startup resource exhaustion (bwrap: Can't fork for pid 1: Resource temporarily unavailable), so this repeat does not resolve the variance question against #164. This is not a rejection based on score. No observation is accepted by this action, and no Dataset evidence is changed. A clean rerun can be submitted in a separate PR.