weight-steering/docs/review/v7_evidence_review.md

v7 evidence cold-review

Reviewer: cold-eyes subagent (read evidence files only, no source). File generated from subagent output; numeric verification reproduced from v7_summary_pct.tsv.

Verdicts

Finding	Tag	Detail
A. Read vs write distinguishable in TSV	[OK]+[CONCERN]	axis_kind column tags every row. Conclusion claim "separate sub-table" is true only in the figure, not the TSV (single flat table). Mild over-claim.
B. dW_left_basis_ceiling = 100% on all three	[OK]	Row 1 of v7_summary_pct.tsv: pct_w_oracle_combined=100.0, oproj=100.0, downproj=100.0. No per-tensor normalization bug.
C. Top-5 intersection = {chars_clusters, attn_min_taskdiff}	[OK]	Reproduces from table.
D1. Single-seed overfitting	[CONCERN]	attn_min_taskdiff R_w lead is 1.30 vs cluster at 1.12-1.15, no error bars. Cannot rule out seed luck.
D2. Write-family R_w < null (z = -1 to -2)	[CONCERN, important]	write 0.954, mlp_write 0.950, global_write 0.946, attn_write 0.917, mlp_roundtrip_write 0.935 -- all BELOW random-rotation null. Two readings: (a) null mis-specified (top-SVD-of-weights bases live in a low-rank atypical region of the d_model sphere relative to LoRA's task-specific delta); (b) substantive: LoRA picks task-specific write directions, NOT generic high-singular-value ones. Either way, this is a real signal, not noise. v7 conclusion ignores it.
D3. dW_left ceiling R_w=16.39 not 128	[OK]	Consistent with LoRA rank 32 (not d_model). Frobenius-balanced concentration of a rank-32 delta projected onto its own top-PCS=8 basis is bounded by ~rank/PCS, not d_model/PCS.
D4. chars_clusters rank-collapse inflates R_act ~14%	[CONCERN]	Acknowledged inline but TSV shows mean_rank=8 (one direction is noise). After ~14% deflation chars_clusters R_act ~ 10.4, still top-5 but fragile. attn_min_taskdiff R_w lead would not survive this artifact.
D5. attn_min_taskdiff topic-set leakage	[CONCERN, low]	LoRA training set and probe set share SYCOPHANCY_TOPICS. Basis-construction signal (base-model attention on persona-prompted probes) differs from training signal (chat completions), but worth checking whether splitting topics across train/probe changes the lead.
E. Conclusion calibration	[CLAIM]	Overclaims "best primitive = chars_clusters" (knife-edge: R_w_combined 1.155 vs TaskDiff_contrast 1.152). Underclaims by not surfacing D2.

BLUF

v7 evidence checks out mechanically: dW oracle row hits 100% on all three pct_w columns, top-5 intersection {chars_clusters, attn_min_taskdiff} reproduces, and the R_w_combined=16.39 ceiling is consistent with the LoRA delta being rank-32 (not a normalization bug). Headline "best primitive = chars_clusters" is on a knife-edge: R_w lead (1.155 vs 1.152 for TaskDiff_contrast) is statistically tied and R_act is inflated by the acknowledged k-1 rank collapse to 7. The most interesting unflagged signal is that write-family bases (write, mlp_write, global_write, attn_write) all sit at R_w_combined ~ 0.92-0.96 with z ~ -1 to -2 -- below random-rotation null on the very axis they were designed for. That's either a null-spec issue or a substantive finding that LoRA's delta lives in task-specific write directions, not top-SVD write directions. Conclusion should say which.

Recommended next steps

1. Add a sentence in v7_conclusion.md flagging D2 as a real result needing interpretation.
2. Re-baseline against "random rank-8 basis sampled from the rank-32 LoRA delta column space" (i.e. null inside the LoRA-relevant slice). If write-family climbs back above 1 and chars_clusters/attn_min_taskdiff lose their lead, D2(a) wins. If write-family stays below, D2(b) wins.
3. v7b: multi-LoRA-seed, with stability filter on top-PCS principal angles.
4. Topic-split holdout for attn_min_taskdiff to rule out D5.