Correct #825 promoted body: 0.1087 cited as control R² is fraction_of_fulln_ceiling (true R²=0.0732)
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Goal
Correct a wrong-metric figure in task #825's promoted clean-result body: the separator/punctuation-control read cited as "R² ≈ 0.1087" is NOT a held-out R² — it is fraction_of_fulln_ceiling (0.10868122095179335) from eval_results/issue_825/base-separator-control/base_sep_to_chat.json → decision_support.instruct_reference; the actual transfer R² in the same artifact is sep_to_chat_fulln_r2 = 0.0731533298226521, and the committed within-WikiText L19 ridge cell (eval_results/issue_931/cells_armC_sep.json → r2_per_layer_obs[19]) is −3.17. Apply a NON-Takeaway-flip prose correction via task.py set-body where possible; anything touching a bolded Takeaway stays in this task's scope as a body-prose correction with the classification untouched (user-only contract unchanged).
Evidence
- Refuting artifacts:
eval_results/issue_825/base-separator-control/base_sep_to_chat.json(both fields present, verified 2026-08-22, task #1901 round generic-boundary-token-control plan v10 §2/§4/M2);eval_results/issue_931/cells_armC_sep.json. - The #1901 round has already retired the downstream consumer:
docs/posters/mats_2026/make_plot1_scaling.pyno longer draws the 0.1087 hline (commits 291da66d/7475bd58/789e64d2), and a MEASURED single-type generic-boundary-token control curve now exists (eval_results/issue_1901/paper_densify/boundary_token_scaling_L19.json, branch issue-1901-btokctl) — cite it as the replacement reference where #825's body needs one. - Residual also worth fixing in the same pass:
scripts/issue1901_body_figures.pyrendererboundary_hlinekwarg docstring still carries "(instruct R^2 0.1087...)" (pre-existing at 291da66d; flagged r1 g4 Minor).
Provenance
Filed by the #1901 same-issue follow-up round (generic-boundary-token-control) per plan v10 M2 result-time duty: a round that refutes a claim in a promoted body files the correction task in the same turn as the result summary (CLAUDE.md § inline estimator-validity + record-integrity duties, item 3).