Auto-sync: 2026-06-29 12:50:13

results/README.md

@@ -66,6 +66,9 @@ Use these summaries for the current paper tables:
 	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.md`,
 	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_summary.md`,
 	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_noopbonus0p03_summary.md`,
+	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.md`,
+	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.md`,
+	  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.md`
@@ -116,13 +119,15 @@ Completed deployment-clean diagnostics:
 
 Completed residual transport jobs:
 
-Last checked: `2026-06-29 16:27 UTC`. The K4 masked composed type-consensus
+Last checked: `2026-06-29 16:46 UTC`. The K4 masked composed type-consensus
 sweep completed and produced a new clean best, 35.54%, while the pure masked
 composition row reached 35.30%. Raw selected candidate types show no
 random-negative or wrong-direction composite leak. The paper table/paired
 analysis now use the masked composed no-op row as `best_clean_key`. The
 component-wise composite-prior follow-up completed at 35.36%, below the exact
-no-op prior row.
+no-op prior row. Composite-only L2 trust penalties completed: penalty `0.02`
+ties the 35.54% best with lower action MSE, while penalty `0.05` reaches
+35.48%, so this is explanatory regularization rather than a new top-line gain.
 
 - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
   completed 32.12% result is unchanged.
@@ -258,6 +263,15 @@ no-op prior row.
   34.78%/34.61%/36.70%, below the 35.54% exact-prior row. Local paper builders
   updated artifacts; redundant rebuild job `14912563` was canceled while still
   pending.
+- Job `14913943` completed the CPU smoke for composite-only L2 trust penalties
+  on masked composed type-consensus transport. Metadata records
+  `retrieval_residual_composite_l2_penalty_scale=0.02`, selected candidate types
+  have no random-negative or wrong-direction leak, and the smoke selected
+  `policy_residual` on 8/8 groups. Jobs `14913944`/`14913951` completed the
+  GPU arrays. Penalty `0.02` ties the 35.54% exact-prior row with seeds
+  35.30%/34.61%/36.70%, slightly higher progress and lower action MSE. Penalty
+  `0.05` reaches 35.48%. Local summaries and builders updated artifacts;
+  redundant summary/rebuild jobs `14913955`/`14913956`/`14913958` were canceled.
 - Job `14894281` completed the unit smoke for the source-progress viability gate,
   including the variable residual-count padding check.
 - Job `14894282` completed the ManiSkill CPU rollout smoke for K4 mean-by-type