diff --git a/docs/spec/20260630_authority_steer_pipeline.md b/docs/spec/20260630_authority_steer_pipeline.md index a2c9a05..fd0441a 100644 --- a/docs/spec/20260630_authority_steer_pipeline.md +++ b/docs/spec/20260630_authority_steer_pipeline.md @@ -174,3 +174,4 @@ Out: - 2026-06-30: Steering-lite commit `b35e269` adds `data/persona_library_selections/pure_authority_qwen3_14b_strict25.jsonl`, matching summary, and `scripts/run_pure_authority_strict25_showcase.sh`. This is the active pure-Authority selection. - 2026-06-30: Queued pure-Authority strict25 method comparison on MFV/MFQ-2 with `c-grid=0.5,1`, `admin-n-samples=8`: pueue 406 `mean_diff`, 407 `pca`, 408 `sspace`, 409 `directional_ablation`, 410 `linear_act`. Pass criterion in labels: small-c Authority sign correct, MFV Social Norms not larger than Authority, coherence ok. - 2026-06-30: Pueue 406 `mean_diff` completed at `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T123208Z_pure_authority_strict25_mean_diff_mfv_mfq2_n8`. Verifier verdict: fail. MFQ-2 small-c direction passed (`authority` C base `30.180`, `+0.5` delta `+5.254`, `-0.5` delta `-4.825`) and coherence passed (`pmass=1.000`). MFV failed direction and selectivity: Authority dlogit deltas were `-0.065` at `+0.5` and `-0.045` at `-0.5`; Social Norms moved more (`-0.632`, `+1.001`). MFV coherence was clean (`pmass=1.000`, `frac_unscorable=0.000`), so this is not answer collapse. +- 2026-06-30: Pueue 407 `pca` completed at `/media/wassname/SGIronWolf/projects5/2026/lite/steering-lite/outputs/20260630T123208Z_pure_authority_strict25_pca_mfv_mfq2_n8`. Verifier verdict: fail. MFV small-c direction barely passed (`Authority` dlogit `+0.024` at `+0.5`, `-0.069` at `-0.5`) but selectivity failed because Social Norms moved more (`-0.005`, `+0.427`). MFQ-2 direction failed: `authority` C base `30.836`, `+0.5` delta `-1.018`, `-0.5` delta `-4.221`; high-c rows also lowered Authority. Coherence was clean for both instruments.