feat: add tune-only Qwen3.8 E02 candidate (#185)

docs/QWEN38_27B_RESEARCH_ROADMAP.md

@@ -2,15 +2,18 @@
 
 ## Decision summary
 
-The current Qwen3.8-27B result is a useful capability-preserving control, not a
-successful refusal-removal result:
+The promotion-grade E01 run is a useful causal control, not a successful
+refusal-removal result:
 
-- refusal rate: 90.9% (30/33 still refusing);
-- coherence: 80% (8/10);
-- perplexity: 3.11;
-- first-token KL divergence: 0.2405;
-- 39 middle-depth layers and 312 matrices modified;
-- 13 layers partially reverted by the current correction pass.
+- archive run: `run-4787f0f2c2ee4a6b8ca8840684eff99b`;
+- refusal rate: 92% (184/200 still refusing);
+- coherence: 70% (7/10);
+- capability: 83% (5/6 checks);
+- perplexity: 3.06;
+- sequence-token KL divergence: 0.0513;
+- first-token KL divergence: 0.1235;
+- 62 layers and 62 attention-output matrices modified;
+- checkpoint: 53,812,173,294 bytes, 35 files, all SHA-256 inventoried.
 
 Do not increase surgery strength or enable the full advanced toggle set yet.
 The next engineering release should make the optimizer and evaluation reliable,
@@ -211,7 +214,7 @@ 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 | How much data is needed? | 33/64/128/256/400 | fixed DIM/SVD setup | direction and outcome stability plateau |
+| 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 |
 | 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 |
@@ -253,12 +256,11 @@ from 60% refusal in S2 to 100% in several other strata.
 
 ## Recommended next operational run
 
-Do not treat another 54 GB permanent checkpoint as the next experiment. After
-P0 and hook-mode P1 exist, run E00--E03 first. If an immediate pre-fix diagnostic
-is unavoidable, use the last stable settings but increase direction-training
-volume to at least 256 pairs, keep all destructive optional techniques off, and
-do not promote or publish the output because the current KL and spectral gates
-are not trustworthy enough for release qualification.
+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.
 
 ## Evidence base
 

docs/qwen38-experiment-matrix-v1.json

@@ -42,8 +42,20 @@
     },
     {
       "id": "E02",
-      "purpose": "direction training-volume ablation",
-      "depends_on": "E01"
+      "purpose": "tune-only multi-direction residual-writer candidate",
+      "depends_on": "E01",
+      "method": "qwen38_e02",
+      "evaluation_split": "optimizer_tune",
+      "direction_method": "svd",
+      "directions": 4,
+      "layer_selection": "middle60",
+      "projection_target": "output",
+      "regularization": 0.0,
+      "norm_preserve": true,
+      "refinement_passes": 1,
+      "rdo_refinement": true,
+      "winsorize_activations": true,
+      "kl_optimization": true
     },
     {
       "id": "E03",