Published result

Screen of shaping evidence and probe handling

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

Result summary and pack text · reproduce/lab-shaping-evidence-probe-screen-20261108/README.md

In plain terms

A combined evidence and probe screen reached 15/20 ready versus 15/20 stored control, with 1 gains and 1 losses; it was dropped.

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 Superseded, A later test replaced this result.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).

This candidate-only screen compared two repetitions of the same ten small requests against stored current-release cells. Candidate readiness was 15/20 (8/10 and 7/10), versus 15/20 (8/10 and 7/10) for the stored control, with 1 paired gains and 1 paired losses. The pre-registered keep rule required at least 18/20 and no paired regressions; the disposition is drop. All 20 cells per arm produced a valid Intent. Acceptance that separates correct from incorrect implementations, downstream landing, and hidden-grader performance were not measured; per-task fields are marked unmeasured. Wall times include failed attempts.

The scratch change combined conservative path-and-line citation normalization before the existing tracked-file/read checks with passing successful no-injected-file probe text to the decision checker, requiring probe quotes to appear in the observed output and in the decision statement alongside the command. The base was Kogen 47236bbb, using shaping 0484edf8. Scratch diff SHA-256 was 674b0d534ab6946f410d28ee9000cac7818199f9254074e710438742e54e1639; Darwin archive SHA-256 was 21555044db3b6e792c71fef8a5a6408de30ae50c0ef922333d7f45e607e26bcc. The candidate was scratch-only. Both arms used Sol low and no answerer; tasks ran on the corresponding control host where available; any cross-host comparisons are identified in the data. The shared credit balance did not move from 53,624; exact attributable credit spend is unknown. Small Go and Node workspaces are reconstructions whose original byte identity is unverified. Held-out tasks were not run.

The checker saw probe information in candidate cell logs, but the events do not retain probe-quote fields, so acceptance of a particular probe quote is not independently observable. Citation normalization firing is not established without a trace showing the before-and-after parse. Do not infer causation from paired changes. Discordance attribution from candidate cell logs:

  • c05 rep 2: gain; control ended at step self-check after decision refusals, while candidate recovered after decision-check refusals; shaping stage, class e (causation unproven).
  • c02 rep 2: loss; control reached readiness, while candidate ended after Intent and decision-check refusals; shaping stage, class e (causation unproven).

Rerun

Run python3 report.py to recompute the paired counts from data.json. The stored control describes the requests and baseline. To repeat the model screen, implement the described change in scratch on the pinned source, verify the diff and release hashes, then run the same requests with Sol low, no answerer and two repetitions per arm; record terminal readiness, refusals, questions, wall and credit readings. The published data and report reproduce the reported counts.