Auto-sync: 2026-06-28 08:27:39

results/README.md

@@ -22,6 +22,7 @@ Use these summaries for the current paper tables:
 - `h16_policy_ckpt_near_miss_policy_bc5_bestpt_field_sweep_summary.md`
 - `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.md`
 - `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.md`
+- `h16_policy_ckpt_residual_tangent_bc5_allmap_v2_best_policy_summary.md`
 - residual scale-sweep summaries:
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale{0p25,0p50,0p75,1p25}_summary.md`
 - residual-family mask and retrieval diagnostic summaries:
@@ -34,7 +35,11 @@ Use these summaries for the current paper tables:
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_no_random_wrongdir_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_no_random_wrongdir_summary.md`,
-  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.md`
+  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.md`,
+  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_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`
 - hybrid summaries:
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k{32,64}_sigma*.md`
 - `h16_policy_ckpt_nonexpert_policy_bc5_summary.md`
@@ -82,7 +87,7 @@ Completed deployment-clean diagnostics:
 
 Completed residual transport jobs:
 
-Last checked: `2026-06-28 06:24 UTC`. No DoVLA jobs are currently queued.
+Last checked: `2026-06-28 12:25 UTC`. No DoVLA jobs are currently queued.
 
 - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
   completed 32.12% result is unchanged.
@@ -110,6 +115,16 @@ Last checked: `2026-06-28 06:24 UTC`. No DoVLA jobs are currently queued.
 - Jobs `14859503`-`14859597` completed typed-safe residual scale fine/zoom
   sweeps. Scales `0.325`, `0.35`, and `0.40` tie at the current clean best,
   33.74%; scale `0.35` has the lowest seed std among the tied rows.
+- Jobs `14862455`-`14862460` completed residual-tangent target export,
+  distillation, and direct/best-rank rollouts. The tangent student is a negative
+  diagnostic: 28.87% for best-policy rollout.
+- Jobs `14862605`-`14862612` completed policy-relative residual-anchor and
+  repaired train-family reliability-prior diagnostics. Policy anchoring ties the
+  old 33.74% row; repaired reliability thresholds at scale `0.35` reach
+  33.33%/33.28%.
+- Jobs `14862635`-`14862828` completed counterfactual-advantage margin sweeps.
+  The current clean best is scale `0.35`, safe residual families, margin `0.20`
+  or `0.22`: 34.84%.
 
 Paper table automation:
 

results/dovla_cil_run_report_2026-06-27.md

@@ -27,14 +27,17 @@ The deployed h=16 behavior-cloning policy remains weak, but the interventional f
 | Field-selected no-expert policy + field, seed-0 train map | 27.65% | -2.09 pp | Field scoring around that student remains below baseline |
 | Field-selected no-expert policy, aligned allmap | 28.00% | -1.74 pp | 100% target-map coverage still fails |
 | Field-selected no-expert policy + field, aligned allmap | 26.49% | -3.25 pp | Aligned field-teacher student remains below baseline |
+| Residual-tangent distillation policy, aligned allmap | 28.87% | -0.87 pp | Tangent pseudo-target imitation does not transfer to rollout |
 | Train-state residual retrieval | 32.12% | +2.38 pp | Positive clean bridge |
 | Train-state residual retrieval, scale 0.25 | 32.93% | +3.19 pp | Smaller tangent step ties prior clean best |
 | Train-state residual retrieval, scale 0.50 | 33.33% | +3.59 pp | Calibrated residual transport |
 | Train-state residual retrieval, no random/wrong-direction | 33.57% | +3.83 pp | Anti-goal family mask improves clean bridge |
 | Train-state residual retrieval, policy/no-op/wrong-gripper | 33.68% | +3.94 pp | Typed family mask improves clean bridge |
-| Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | 33.74% | +4.00 pp | Current best deployment-clean diagnostic |
+| Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | 33.74% | +4.00 pp | Typed tangent transport before abstention |
+| Train-state residual retrieval, safe residuals + advantage margin 0.20 | 34.84% | +5.10 pp | Current best deployment-clean diagnostic |
+| Policy-relative residual anchor, safe residuals | 33.74% | +4.00 pp | Anchor change ties but does not improve expert-relative residuals |
 | Train-state residual retrieval, z-score metric | 32.23% | +2.49 pp | State normalization hurts retrieval here |
-| Train-state residual retrieval, train-family reliability prior | 33.33% | +3.59 pp | Train success thresholds did not filter enough |
+| Train-state residual retrieval, repaired train-family reliability prior | 33.28-33.33% | +3.54-3.59 pp | Train success thresholds do not recover the typed safe mask |
 | Train-state residual retrieval, scale 0.75 | 32.70% | +2.96 pp | Larger tangent step is weaker |
 | Train-state residual retrieval, scale 1.25 | 32.52% | +2.78 pp | Further scale increase does not help |
 | Residual+Gaussian hybrid, K32 sigma0.35 | 31.30% | +1.57 pp | Hybrid proposals dilute residual transport |
@@ -113,11 +116,20 @@ Completed since the first report:
 - Typed-safe scale fine/zoom sweeps completed. Scales `0.325`, `0.35`, and
   `0.40` tie at 33.74%, with scale `0.35` having the lowest seed variance among
   the tied best rows.
+- Residual-tangent distillation completed as a negative diagnostic: the aligned
+  tangent student reaches 28.87% despite low pseudo-target BC loss.
+- Policy-relative residual anchoring completed as a negative/neutral diagnostic:
+  it ties the old 33.74% safe row rather than improving it.
+- Repaired train-family reliability priors at scale `0.35` completed as
+  negative diagnostics: thresholds `0.10` and `0.25` reach 33.33% and 33.28%.
+- Counterfactual-advantage abstention completed as the new clean best: typed-safe
+  residual transport at scale `0.35` with margin `0.20`/`0.22` reaches 34.84%.
 - Z-score retrieval completed as a negative diagnostic: 32.23-32.81%, below raw
   residual retrieval.
 - Train-family reliability priors completed as neutral/negative diagnostics:
-  thresholds through `0.75` do not filter the bad residual families and remain
-  at 33.33% for scale `0.50`.
+  the original scale-`0.50` thresholds stayed at 33.33%; after repairing
+  threshold pass-through, scale-`0.35` thresholds still remain below the typed
+  safe residual result.
 
 Completed residual-family and retrieval-metric jobs:
 
@@ -159,6 +171,14 @@ Completed residual-family and retrieval-metric jobs:
 - `14859593`/`14859594`: typed-safe scale `0.375`: 33.51%.
 - `14859595`/`14859596`: typed-safe scale `0.40`: 33.74%.
 - `14859597`: rebuilt `paper_table_status.*` after typed-safe zoom summaries.
+- `14862455`-`14862460`: residual-tangent distillation export/train/eval:
+  best-policy 28.87%, best-rank 27.48%.
+- `14862605`/`14862606`: policy-relative residual anchor, scale `0.35`,
+  safe non-expert residuals: 33.74%.
+- `14862609`-`14862612`: repaired scale-`0.35` train-family reliability
+  thresholds `0.10`/`0.25`: 33.33%/33.28%.
+- `14862635`-`14862828`: counterfactual-advantage margin sweeps. Best clean
+  row is scale `0.35`, safe residuals, margin `0.20` or `0.22`: 34.84%.
 
 The motivation for this batch is the observed selected-candidate distribution
 in the 33.33% scale-0.50 run: `residual_random_negative` and