Published result

A readiness regression claim remains tentative after targeted repeats

A plain-language summary comes first, followed by the full text of the result's reproduction pack.

Result summary and pack text · reproduce/kogen-coherent-worseattr-confirmation-20261103/README.md

In plain terms

A bounded recheck of two coherent component versions keeps earlier readiness-regression claims tentative. Intent validity, generated acceptance, final hidden correctness and environment stops are reported separately. No harmful component commit is established, and useful checks remain intact.

Verdict key: Yes, The result supports the change or claim that was tested. No, The result does not support it, so nothing changes. Unresolved, The evidence is not strong enough either way. Superseded, A later test replaced this result.

Full text of the reproduction pack

The text below is the pack's README as published in the repository (a system folder name may show without its leading slash, so that the site's privacy scan passes).

Verdict: TENTATIVE. The original registered WORSE readings rested on single attempts per decisive task. Bounded exact-version candidate repetitions separate readiness, acceptance quality and final hidden correctness; no harmful component commit is isolated. Useful coverage and binding checks stay intact. No revert is requested.

The full-pipeline candidates use native OpenAI gpt-6.1-terra medium in all roles. The fixed cap is six added attempts: two deployment attempts on the newer tuple, then two deployment and two cleanup attempts on the older tuple. There are no new controls, retries, tasks or judges. Original assignments and every failed attempt remain in per-task means. Matched stored model/effort cohorts stay separate, and historical approval-timeout sensitivity is explicit. Twenty different-model supplementary rows are excluded; pre-inference environment stops provide no valid held-out comparator. The data reports each endpoint separately; acceptance separation is unestablished without an independent replay.

This is selected, small and historically compared evidence, with host differences. It supports no significance, equivalence, isolated component effect or causal rescue claim. A valid Intent can still contain a fixture that later blocks Build; an honest external-environment stop is distinct from a harness void.

Rerun

Patch: none. Every component revision for both exact candidate tuples is linked by public commit URL in source-commits.json; no source or source patch is shipped. Run python3 report.py to reproduce the recorded aggregate results from data.json. A paid replication uses the existing qualified task kits and hidden evaluators, a candidate-only full-pipeline lanes experiment, the linked frozen tuple, the same model and effort, two repetitions per selected task and the unchanged 3600-second per-cell deadline. Raw runtime receipts and task identities remain in the private execution records; no new qualification or fresh control is part of this recheck.

Accounting and disposition

All tokens and summed cell wall include failed attempts. Cached input is part of input, and reasoning is part of output. Dollars are unpriced, not zero; orchestration cost is unmetered. No extra author, helper or judge calls, source builds or timeout increases. Disposition: drop the bounded attribution recheck at its cap; both historical regression claims remain TENTATIVE. No component is qualified.

This research disposition does not hold adoption or replace repository and release tests.