Auto-sync: 2026-06-28 02:25:24

results/README.md

@@ -24,7 +24,7 @@ Use these summaries for the current paper tables:
 - `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_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 summaries, currently active:
+- residual-family mask and retrieval diagnostic summaries:
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_no_random_wrongdir_summary.md`,
   `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_zscore_no_random_wrongdir_summary.md`,
@@ -82,7 +82,7 @@ Completed deployment-clean diagnostics:
 
 Completed residual transport jobs:
 
-Last checked: `2026-06-28 05:47 UTC`.
+Last checked: `2026-06-28 06:24 UTC`. No DoVLA jobs are currently queued.
 
 - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
   completed 32.12% result is unchanged.
@@ -90,42 +90,23 @@ Last checked: `2026-06-28 05:47 UTC`.
   selection after smoke jobs `14858889`/`14858894` caught two wiring bugs before
   rollout jobs started.
 - Jobs `14858875`-`14858883` completed residual scales `0.25`, `0.50`, `0.75`,
-  and `1.25`. Scale `0.50` is the current best clean deployment bridge at
-  33.33%; scale `0.25` ties the previous clean best at 32.93%.
+  and `1.25`. Scale `0.50` reaches 33.33%; scale `0.25` ties the previous clean
+  best at 32.93%.
 - Job `14859041` completed the CPU Apptainer unit smoke for hybrid
   residual+Gaussian selection.
 - Jobs `14859042`-`14859046` completed residual+Gaussian hybrids. K32/sigma0.35
   reaches 31.30%; K64/sigma0.50 reaches 30.90%, so Gaussian hybridization is a
   negative diagnostic.
 
-Active residual-family mask jobs:
-
-- Jobs `14859188`/`14859189` evaluate and summarize residual scale `0.50` while
-  excluding `residual_random_negative`.
-- Jobs `14859191`/`14859192` evaluate and summarize residual scale `0.50` while
-  excluding `residual_random_negative` and `residual_wrong_direction`.
-- Jobs `14859193`/`14859194` evaluate and summarize residual scale `0.25` with
-  the same random/wrong-direction mask.
-- Jobs `14859195`/`14859196` evaluate and summarize residual scale `0.50` with
-  the policy/no-op/wrong-gripper residual families.
+- Jobs `14859188`-`14859203` completed masked and z-score residual diagnostics.
+  The best clean deployment diagnostic is scale `0.50` with policy/no-op/
+  wrong-gripper residual families at 33.68%. Z-score retrieval is negative
+  (32.23-32.81%).
 - Job `14859165` completed the Apptainer unit smoke for z-score retrieval
   metric selection.
-- Jobs `14859197`/`14859198` evaluate and summarize z-score retrieval at
-  residual scale `0.50`.
-- Jobs `14859199`/`14859200` evaluate and summarize z-score retrieval at
-  residual scale `0.50` with the random/wrong-direction mask.
-- Jobs `14859201`/`14859202` evaluate and summarize z-score retrieval at
-  residual scale `0.25` with the random/wrong-direction mask.
-- Job `14859203` rebuilds `paper_table_status.*` after all masked and z-score
-  summaries.
-- Jobs `14859293`/`14859294` evaluate and summarize a train-family reliability
-  prior at residual scale `0.50` with minimum train success `0.10`.
-- Jobs `14859295`/`14859296` evaluate and summarize the same prior at residual
-  scale `0.50` with minimum train success `0.25`.
-- Jobs `14859297`/`14859298` evaluate and summarize the same prior at residual
-  scale `0.25` with minimum train success `0.25`.
-- Job `14859299` rebuilds `paper_table_status.*` after the reliability-prior
-  summaries.
+- Jobs `14859293`-`14859402` completed train-family reliability-prior
+  diagnostics. Thresholds through `0.75` do not filter enough; scale `0.50`
+  stays at 33.33%, and scale `0.25`/threshold `0.25` is 32.93%.
 
 Paper table automation:
 

results/dovla_cil_run_report_2026-06-27.md

@@ -29,7 +29,11 @@ The deployed h=16 behavior-cloning policy remains weak, but the interventional f
 | Field-selected no-expert policy + field, aligned allmap | 26.49% | -3.25 pp | Aligned field-teacher student remains below baseline |
 | 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 | Current best deployment-clean bridge |
+| 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 | Current best deployment-clean diagnostic |
+| 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, 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 |
@@ -55,6 +59,9 @@ Canonical summaries:
 - `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.md`
 - `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.md`
 - `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.md`
+- `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.md`
+- `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_zscore_summary.md`
+- `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_type_success0p75_summary.md`
 - `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.md`
 - `results/h16_policy_ckpt_nonexpert_policy_bc5_summary.md`
 - `results/h16_policy_ckpt_nonexpert_policy_bc5_bestpt_field_sweep_summary.md`
@@ -95,34 +102,50 @@ Completed since the first report:
   diagnostic: direct student 28.00%, field-guided best 26.49%, despite 100%
   target-map coverage.
 - Residual scale sweep completed. Scale `0.50` is now the strongest
-  deployment-clean result at 33.33%; scale `0.25` ties the previous clean best
+  residual-only result at 33.33%; scale `0.25` ties the previous clean best
   at 32.93%; scales `0.75` and `1.25` are weaker at 32.70% and 32.52%.
 - Residual+Gaussian hybrids completed as negative diagnostics: K32/sigma0.35 is
   31.30%, K64/sigma0.50 is 30.90%.
-
-Active residual-family mask jobs:
-
-- `14859188`/`14859189`: scale `0.50`, excluding `residual_random_negative`.
+- Masked residual diagnostics completed. The best current clean deployment
+  diagnostic is scale `0.50` with policy/no-op/wrong-gripper residual families
+  at 33.68%; excluding random/wrong-direction residuals reaches 33.57%.
+- 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`.
+
+Completed residual-family and retrieval-metric jobs:
+
+- `14859188`/`14859189`: scale `0.50`, excluding `residual_random_negative`:
+  33.45%.
 - `14859191`/`14859192`: scale `0.50`, excluding `residual_random_negative` and
-  `residual_wrong_direction`.
-- `14859193`/`14859194`: scale `0.25`, excluding the same anti-goal residuals.
+  `residual_wrong_direction`: 33.57%.
+- `14859193`/`14859194`: scale `0.25`, excluding the same anti-goal residuals:
+  33.45%.
 - `14859195`/`14859196`: scale `0.50`, keeping policy/no-op/wrong-gripper
-  residual families.
+  residual families: 33.68%.
 - `14859165`: completed Apptainer unit smoke for the z-score retrieval metric.
-- `14859197`/`14859198`: z-score retrieval at scale `0.50`.
+- `14859197`/`14859198`: z-score retrieval at scale `0.50`: 32.23%.
 - `14859199`/`14859200`: z-score retrieval at scale `0.50`, excluding
-  `residual_random_negative` and `residual_wrong_direction`.
+  `residual_random_negative` and `residual_wrong_direction`: 32.75%.
 - `14859201`/`14859202`: z-score retrieval at scale `0.25`, excluding the same
-  anti-goal residuals.
-- `14859203`: rebuilds `paper_table_status.*` after all masked and z-score
+  anti-goal residuals: 32.81%.
+- `14859203`: rebuilt `paper_table_status.*` after all masked and z-score
   summaries.
 - `14859293`/`14859294`: train-family reliability prior at scale `0.50`, minimum
-  train success `0.10`.
+  train success `0.10`: 33.33%.
 - `14859295`/`14859296`: train-family reliability prior at scale `0.50`, minimum
-  train success `0.25`.
+  train success `0.25`: 33.33%.
 - `14859297`/`14859298`: train-family reliability prior at scale `0.25`, minimum
-  train success `0.25`.
-- `14859299`: rebuilds `paper_table_status.*` after reliability-prior summaries.
+  train success `0.25`: 32.93%.
+- `14859299`: rebuilt `paper_table_status.*` after reliability-prior summaries.
+- `14859398`/`14859399`: train-family reliability prior at scale `0.50`, minimum
+  train success `0.50`: 33.33%.
+- `14859400`/`14859401`: train-family reliability prior at scale `0.50`, minimum
+  train success `0.75`: 33.33%.
+- `14859402`: rebuilt `paper_table_status.*` after high-threshold reliability
+  summaries.
 
 The motivation for this batch is the observed selected-candidate distribution
 in the 33.33% scale-0.50 run: `residual_random_negative` and
@@ -133,9 +156,9 @@ proposal family is consistent with the counterfactual field's local geometry.
 The z-score branch tests the other half of the same bottleneck: whether nearest
 train-state tangent transport was limited by raw feature scale rather than by
 the residual idea itself.
-The train-family reliability branch is the paper-safe version of typed residual
-masking: it filters residual families using only train-split terminal success
-rates, never validation candidate rewards.
+The train-family reliability branch was intended as the paper-safe version of
+typed residual masking, but terminal-success thresholds did not filter the
+low-performing families, so it should be framed as a negative diagnostic.
 
 Use `results/paper_table_status.md` as the canonical current table. The job
 table below is historical and includes statuses that were true at earlier