feat(qwen38): add E03 coherence recovery candidate

docs/QWEN38_27B_RESEARCH_ROADMAP.md

@@ -215,7 +215,7 @@ method needs an isolated ablation and an interaction test before composition.
 | E00 | Is the pristine evaluator stable? | final-test only, 3 seeds | none | metric variance and baselines recorded |
 | E01 | Does the Arditi control work? | 400/100 | DIM, one layer/direction sweep | lower refusal with <=1.25x PPL |
 | E02 | Does a low-rank writer intervention improve E01? | 500/142; final test prohibited | SVD-4 + RDO, middle-60% residual writers, KL rollback | tune refusal <30% and coherence >=80% |
-| E03 | Which architecture component carries refusal? | 400/100 | DeltaNet vs full-attn vs MLP groups | best causal effect per KL unit |
+| E03 | Can slight attenuation recover E02 coherence? | 500/142; final test prohibited | E02 kernel with 0.10 regularization | tune refusal <30% and coherence >=80% |
 | E04 | Does SVD rank help? | 400/100 | k=1..7, joint held-out selection | nondominated gain over E01 |
 | E05 | Does RDO improve targeting? | 400/100 | one and multiple RDO directions | held-out gain over E01/E04 |
 | E06 | Do concept-cone combinations help? | large corpus + 128--512 trials | SOM/independent combinations | reproducible gain over E05 |
@@ -256,11 +256,15 @@ from 60% refusal in S2 to 100% in several other strata.
 
 ## Recommended next operational run
 
-Run the pre-registered E02 candidate on the 142-pair optimizer-tune partition.
-E02 must not read the 200-pair final partition. Retain its complete archive even
-if it fails. If it misses either target, use only tune evidence to define the
-next isolated ablation. If it passes both targets, reload that saved checkpoint
-and run the final partition exactly once for release qualification.
+E02 completed on the optimizer-tune partition as run
+`run-efb2f334197e48fb82587f39a84fe6c9`: refusal fell to 1% (1/142), but
+coherence was 70% and 6/142 harmful responses were degenerate. Its checkpoint
+and full archive are retained. E03 changes only regularization from 0.00 to
+0.10, retaining 10% of the measured refusal subspace to recover coherence while
+using E02's 29-point refusal margin. E03 must not read the 200-pair final
+partition. Promote only if tune refusal remains below 30% and coherence reaches
+at least 80%; then evaluate the final partition exactly once and independently
+reload the saved checkpoint for release qualification.
 
 ## Evidence base
 

docs/qwen38-experiment-matrix-v1.json

@@ -59,8 +59,20 @@
     },
     {
       "id": "E03",
-      "purpose": "DeltaNet/full-attention/MLP writer ablation",
-      "depends_on": "E01"
+      "purpose": "tune-only coherence recovery by attenuating E02",
+      "depends_on": "E02",
+      "method": "qwen38_e03",
+      "evaluation_split": "optimizer_tune",
+      "direction_method": "svd",
+      "directions": 4,
+      "layer_selection": "middle60",
+      "projection_target": "output",
+      "regularization": 0.1,
+      "norm_preserve": true,
+      "refinement_passes": 1,
+      "rdo_refinement": true,
+      "winsorize_activations": true,
+      "kl_optimization": true
     }
   ]
 }