[jasonssdev]

Refs #1579

Step 1 of #1579: measure what the selector judge's cut buys before changing it. Measurement only. No default changes.

What

evals/concept_model_screen/cut_precision.py (with --self-test) runs #1557's G1 source-grounded precision (judge gpt-oss:20b, the same prompt and parser) on the objects the judge cuts and on those it keeps, over the cached #1571 replays. No candidate model was called.

• Pre-registration (committed before any model call): CUT_PRECISION_PREREGISTRATION.md.
• Results: CUT_PRECISION_RESULTS.md.
• Run: 216 of 216 window calls, about 36 min of Ollama.

judge arm	kept	cut	kept minus cut [95% CI]
self	153 of 165 (0.927)	128 of 138 (0.928)	-0.0003 [-0.0681, +0.0689]
fixed	147 of 160 (0.919)	134 of 143 (0.937)	-0.0183 [-0.0963, +0.0510]

Verdict: inconclusive under the registered rule

Cut objects are about as well supported as kept ones, so the cut does not buy source-grounded precision at this n.

• "The cut buys precision" fails clearly.
• "The cut is too strict" misses narrowly: the upper bound is +0.051 against the registered +0.05.

Deviation from the pre-registration (disclosed in RESULTS)

After labelling, gemma4:26b-a4b's self arm was changed to read its fixed records, because for that extractor the self judge and the fixed judge are the same model. A sensitivity check without gemma4, from stored labels and with no model call, gives:

• self: -0.0170 [-0.0964, +0.0618], still inconclusive.
• fixed: -0.0395 [-0.1336, +0.0334], which on its own reads "too strict".

The overall verdict stays inconclusive. The fixed arm's label depends on whether gemma4 is pooled, and RESULTS says so.

Checks

ruff check, ruff format --check, mypy, evals/run_self_tests.py (66 of 66), and tests/unit/test_eval_self_test_runner.py.