refactor: name things after the method, not the falsified conclusion (#46)

[xiaohan2012]

Closes #46.

#45 falsified self-repair in TFMs, so every self_repair name asserted a disproven hypothesis. Rule applied throughout: name the method/mechanism, never the conclusion; a paper-specific experiment carries the paper's name.

Library

from	to	why
evaluation/self_repair.py	evaluation/balef_exp6.py	the per-depth-tuned-decoder design is balef2026's, not a generic ablation
evaluation/direct_effect.py	evaluation/path_patching.py	the module's identity is the method; it returns both DE and TE — old name dropped half
direct_total_effect()	layer_effects()	reads as path_patching.layer_effects(...)
core/resample_ablation.py	core/donor_delta.py	it only builds δ; core/interventions.py executes the ablation
—	new evaluation/native_readout.py	native_final_logits/auc extracted

The extraction is a prerequisite, not a bonus: both path_patching and balef_exp6 use those two functions. Leaving them inside the exp6 module would make its new name a lie.

Scripts

from	to
run_self_repair_sweep.py	run_balef_exp6_sweep.py
plot_self_repair.py	plot_balef_exp6_trajectory.py
plot_cross_dataset.py	plot_balef_exp6_cross_dataset.py
run_direct_effect_sweep.py	run_path_patching_sweep.py
plot_ce_de_law.py	plot_compensation_fit.py — our result is that there is no law
validate_resample_self_repair.py	compare_ablation_stability.py
validate_resample_donor_health.py	check_donor_delta_health.py

compare_ablation_stability evaluates the perturbation method (std(imm) across layers + overshoot count), not Exp6's conclusion — and it explicitly refuses to treat zero as ground truth, so "validate … vs …" was the wrong verb.

Content, not just filenames

• compare_ablation_stability computed SR(m) = imm(m) − TE(m) and called it self-repair. That is exactly the non-identifiable quantity: imm is a mid-depth tuned-probe reading, not the frozen-downstream DE. Renamed to recovery, caveat stated inline, "Claim 1 — self-repair holds" framing dropped.
• Neutralized docstrings/comments that asserted self-repair. Kept the term where it names the hypothesis under test (e.g. "below-diagonal = self-repair" in the DE–TE scatter) — that usage is correct.
• evaluation/preprocess.py docstring claimed LimiX-only; the module holds all four backbones.
• Default artifact paths follow: out/self_repair*.json → out/balef_exp6*.json.

Tests

test_self_repair → test_balef_exp6 · test_direct_effect → test_path_patching · test_resample_ablation → test_donor_delta.

91 passed, 1 skipped (-m "not real_model", 3.5s). No behavior change — renames + docstrings only.

Left alone

• DESIGN.md / docs/reference-pipeline.md reference experiments/exp6_self_repair.py, a historical layout that never existed in this tree.
• Superseded local scripts (plot_hydra_layer.py, plot_ce_vs_de.py, plot_layer_ce.py, draw_ce_methods.py) are untracked — author's call to delete.

🤖 Generated with Claude Code