# A correction can damage what already worked

**19 October 2026 · additive, post-hoc finding for essai 1.** The original experiment, protocol, data, freeze and saved model results remain unchanged. This diagnostic asks a question the first analysis omitted: after learning the stricter use, does the learner retain its performance in the original use?

The original `after_correction` condition scored only decision use. We replayed the already saved corrected trees on the already held-out exploratory cases, without fitting, changing feedback, changing splits or selecting a different curriculum.

| Held-out use | Initial tree | Corrected tree | Change |
|---|---:|---:|---:|
| Exploratory | 96/96 | 92/96 | −4 |
| Decision | 86/96 | 88/96 | +2 |
| Equally weighted two-use test | 182/192 | 180/192 | −2 |

The combined row is a stated equal weighting of this authored test world. It is not a deployment-mixture estimate or a population finding. All four additional exploratory errors falsely keep briefs with absent evidence: `B12-e-pl`, `B12-e-po`, `B13-e-pl`, `B13-e-po`. These are two substantive bundles, each with two presentation variants. Adequate exploratory cases retained remain 16/16; the loss is four unsupported accepts, not newly blocked good cases.

In fold 1 (the second fold), the corrected tree routes clear → complete → current → keep. On that branch it no longer checks evidence or purpose. Six labelled training records support the leaf, but their local agreement does not establish that absent evidence is harmless. The old training records were retained: the change comes from refitting the partition, not deleting the original history.

This strengthens the practical hypothesis behind the next trial: a useful correction should identify what changes, where it applies and what remains valid. More feedback can otherwise replace a previously useful distinction with an accidental correlation. The replay demonstrates that mechanism in this small tree; it does not establish a defect in language-model feedback learning.

Run `python3 regression.py verify` from any directory. The script verifies the original frozen input hashes, replays the original prediction function, checks the three previously reported scores exactly, and compares every saved diagnostic field and prediction. `results.json` records source hashes, all 192 before/after predictions, paths and confusion matrices. Its `write` command refuses to replace an existing finding.

Continue to [essai 2: lessons with a scope](../scoped-feedback/) or return to the [original review desk](../).
