anhtld commited on
Commit
9b5c4e3
·
verified ·
1 Parent(s): 811688f

Auto-sync: 2026-06-28 22:04:20 (part 2)

Browse files
Files changed (1) hide show
  1. results/README.md +16 -3
results/README.md CHANGED
@@ -41,6 +41,8 @@ Use these summaries for the current paper tables:
41
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_summary.md`,
42
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn2_scale0p40_safe_types_margin0p20_summary.md`,
43
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_summary.md`,
 
 
44
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.md`,
45
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.md`,
46
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.md`
@@ -91,9 +93,9 @@ Completed deployment-clean diagnostics:
91
 
92
  Completed residual transport jobs:
93
 
94
- Last checked: `2026-06-29 00:41 UTC`. The counterfactual tangent ray-search batch
95
- completed; the typed no-op residual-prior clean sweep also completed after a
96
- passing CPU smoke.
97
 
98
  - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
99
  completed 32.12% result is unchanged.
@@ -177,6 +179,17 @@ passing CPU smoke.
177
  35.19%, and 35.19%. These are near-tie/negative diagnostics below the current
178
  35.25% equal mean-consensus no-op plateau. Summary jobs `14891083`/
179
  `14891085`/`14891087`/`14891088` and rebuild job `14891089` completed.
 
 
 
 
 
 
 
 
 
 
 
180
  - Job `14869627` completed the CPU Apptainer smoke for the residual scale-grid
181
  selector: selected index `3` and returned action `0.20` on the toy ray-search
182
  case, validating candidate expansion before rollout.
 
41
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_summary.md`,
42
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn2_scale0p40_safe_types_margin0p20_summary.md`,
43
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_summary.md`,
44
+ `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p20_summary.md`,
45
+ `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.md`,
46
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.md`,
47
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.md`,
48
  `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.md`
 
93
 
94
  Completed residual transport jobs:
95
 
96
+ Last checked: `2026-06-29 02:00 UTC`. The field-conditioned tangent-barycenter
97
+ batch completed after passing CPU smokes, and the paper table rebuild now includes
98
+ the field-softmax rows.
99
 
100
  - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
101
  completed 32.12% result is unchanged.
 
179
  35.19%, and 35.19%. These are near-tie/negative diagnostics below the current
180
  35.25% equal mean-consensus no-op plateau. Summary jobs `14891083`/
181
  `14891085`/`14891087`/`14891088` and rebuild job `14891089` completed.
182
+ - Jobs `14891889`/`14891902`/`14891923` completed the K4/K8 field-softmax
183
+ residual-barycenter sweeps after CPU smokes `14891870` and `14892092`. K4
184
+ field-softmax reaches 34.96% with or without no-op 0.03 at margin `0.20`; K8
185
+ with no-op 0.03 reaches 34.84%. Summary jobs `14891934`/`14891946`/`14891960`
186
+ completed.
187
+ - Jobs `14892958`/`14892975`/`14892990` completed the K4 field-softmax no-op
188
+ margin sweep. Margins `0.10`, `0.05`, and `0.00` reach 35.19%, 35.07%, and
189
+ 34.84%. Lower margins select many more field-softmax aggregates, but the
190
+ overall success drops, so this is a useful negative/near-tie diagnostic rather
191
+ than a new clean best. Summary jobs `14893002`/`14893016`/`14893028` and
192
+ rebuild job `14893069` completed.
193
  - Job `14869627` completed the CPU Apptainer smoke for the residual scale-grid
194
  selector: selected index `3` and returned action `0.20` on the toy ray-search
195
  case, validating candidate expansion before rollout.