Auto-sync: 2026-06-29 12:01:54 (part 4)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_composemasked_grid035040045_noopbonus003",
-  "generated_utc": "2026-06-29T15:56:21+00:00",
+  "generated_utc": "2026-06-29T16:13:16+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.07246376811594196,
     "best_clean_vs_h16": 0.0579710144927536,
@@ -844,6 +844,11 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
       "std_success": 0.01578047256674342
     },
+    "residual_k4_composemasked_compbonus_grid035040045_noopbonus003": {
+      "label": "K4 composed type-consensus tangents, masked, component no-op bonus 0.03",
+      "missing": true,
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+    },
     "residual_k4_composemasked_grid035040045": {
       "ci95_success": 0.03024411731871433,
       "label": "K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45",

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T15:56:21+00:00`
+Generated: `2026-06-29T16:13:16+00:00`
 
 ## Main Seed Statistics
 
@@ -47,6 +47,7 @@ Generated: `2026-06-29T15:56:21+00:00`
 | residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus tangents, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.14% +/- 1.58 | +/- 3.92 | 56.00% | 0.482 | +4.41 pp |
 | residual_k4_composemasked_grid035040045 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45 | 3 | 35.30% +/- 1.22 | +/- 3.02 | 56.91% | 0.410 | +5.57 pp |
 | residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 35.54% +/- 1.02 | +/- 2.53 | 57.02% | 0.411 | +5.80 pp |
+| residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, component no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
 | repair_nearmiss_k4_grid025035050_margin020 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | 3 | 34.32% +/- 1.35 | +/- 3.36 | 55.97% | 0.394 | +4.58 pp |
 | repair_nearmiss_k4_grid035050075_margin020 | K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20 | 3 | 34.38% +/- 1.50 | +/- 3.73 | 56.05% | 0.394 | +4.64 pp |
 | repair_nearmiss_k4_grid025035050_margin010 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10 | 3 | 34.14% +/- 1.48 | +/- 3.67 | 56.01% | 0.393 | +4.41 pp |

results/paper_core_results.md

@@ -5,9 +5,10 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 
 For paired seed deltas, per-task gaps, and selection histograms, regenerate and
 read `paper_analysis.md` with `python3 scripts/build_paper_analysis.py`. Current
-paired analysis: best clean K4 mean-by-type scale-grid consensus with typed
-no-op prior is `+5.68 pp` over canonical h=16, same-state no-expert lattice is
-`+27.25 pp`, and the remaining clean-to-same-state proposal gap is `+21.57 pp`.
+paired analysis: best clean K4 masked composed type-consensus transport with
+typed no-op prior is `+5.80 pp` over canonical h=16, same-state no-expert
+lattice is `+27.25 pp`, and the remaining clean-to-same-state proposal gap is
+`+21.45 pp`.
 
 | Method | Uses same-state proposals | Uses expert proposal | Success | Gain vs policy | Interpretation |
 |---|---:|---:|---:|---:|---|
@@ -41,17 +42,19 @@ no-op prior is `+5.68 pp` over canonical h=16, same-state no-expert lattice is
 | K2 task-relative residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.26% | +4.52 pp | Actor-pose-only retrieval is too lossy; raw full-state similarity is better for residual transfer |
 | K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus alone does not beat raw K2 residuals |
 | K4 mean-by-type residual retrieval + no-op prior 0.03 | No | No | 35.25% | +5.51 pp | Previous fixed-scale clean plateau; 0.025-0.035 nudges high-value no-op residuals without changing the core proposal family |
-| K4 mean-by-type residual retrieval + scale-grid no-op prior 0.03 | No | No | 35.42% | +5.68 pp | Current best clean diagnostic; field-gated tangent length calibration improves the fixed-scale plateau while staying within the same local residual geometry |
+| K4 mean-by-type residual retrieval + scale-grid no-op prior 0.03 | No | No | 35.42% | +5.68 pp | Previous clean best; field-gated tangent length calibration improves the fixed-scale plateau while staying within the same local residual geometry |
 | K4 mean-by-type residual retrieval + source-progress prior 0.03 | No | No | 35.25% | +5.51 pp | Ties the fixed-scale typed prior without a hand typed no-op prior; train-measured source progress can replace but not improve the typed prior |
 | K4 mean-by-type residual retrieval + source-progress prior 0.05 | No | No | 35.13% | +5.39 pp | A stronger measured-progress prior over-selects nonzero residuals and drops below the plateau |
 | K4 mean-by-type residual retrieval + source-score prior 0.015/0.020 | No | No | 35.25% | +5.51 pp | Full train reward score, including terminal success, also replaces the fixed-scale typed prior without improving it |
 | K4 mean-by-type residual retrieval + scale-grid source-score prior 0.02 | No | No | 35.30% | +5.57 pp | Measured train-source prior benefits from tangent length calibration but remains below the typed no-op scale-grid row |
 | K4 mean-by-type residual retrieval + upper scale-grid no-op prior 0.03 | No | No | 35.36% | +5.62 pp | Scales 0.40/0.45/0.50 nearly tie but do not beat the 0.35/0.40/0.45 row |
 | K4 mean-by-type residual retrieval + wide scale-grid no-op prior 0.03 | No | No | 35.13% | +5.39 pp | Including 0.55 over-extends the transported tangent and drops below the best local calibration |
-| K4 mean-by-type residual retrieval + minimum-energy action penalty | No | No | 35.36-35.42% | +5.62-5.68 pp | A tiny action L2 penalty (0.05) ties the best row, while 0.10/0.20 drop slightly; shortest-action regularization does not add the gain |
+| K4 mean-by-type residual retrieval + minimum-energy action penalty | No | No | 35.36-35.42% | +5.62-5.68 pp | A tiny action L2 penalty (0.05) ties the previous scale-grid row, while 0.10/0.20 drop slightly; shortest-action regularization does not add the gain |
 | K4 mean-by-type residual retrieval + source-advantage prior/gate | No | No | 35.13-35.30% | +5.39-5.57 pp | Measuring local train-source utility lift over the anchor does not replace the typed no-op prior; positive-advantage gates over-filter useful residual geometry |
-| K4 mean-by-type residual retrieval + train-family success bonus | No | No | 35.25-35.42% | +5.51-5.68 pp | A continuous train terminal-success prior is below the best by itself and only ties when added to the no-op row; train outcome reliability does not add the gain |
+| K4 mean-by-type residual retrieval + train-family success bonus | No | No | 35.25-35.42% | +5.51-5.68 pp | A continuous train terminal-success prior is below the best by itself and only ties the previous scale-grid row when added to the no-op row; train outcome reliability does not add the gain |
 | K4 mean-by-type residual retrieval + train-neighbor consensus penalty | No | No | 35.19-35.36% | +5.45-5.62 pp | Penalizing high-dispersion local tangent families is coherent but over-abstains by one success; geometric confidence does not improve the sparse no-op scale-grid row |
+| K4 composed type-consensus residual retrieval, masked | No | No | 35.30% | +5.57 pp | Clean anti-goal composite masking removes the confound; tangent composition alone near-ties but does not beat mean-by-type scale-grid transport |
+| K4 composed type-consensus residual retrieval, masked + no-op prior 0.03 | No | No | 35.54% | +5.80 pp | Current best clean diagnostic; masked local tangent composition adds one success over the previous scale-grid no-op row while staying sparse |
 | K4 repair-tangent residual transport | No | No | 34.14-34.43% | +4.41-4.70 pp | Reversing residuals into failure-to-expert repair tangents is a clean negative diagnostic; the current gain is not recovered by transporting near-miss-to-expert vectors |
 | K4 mean-by-type residual retrieval + source-score prior 0.025 | No | No | 35.19% | +5.45 pp | A stronger reward-score prior drops below the plateau |
 | K4 mean-by-type residual retrieval, no-op-only residuals | No | No | 35.19% | +5.45 pp | Removing wrong-gripper residuals loses one success versus the fixed-scale safe-family plateau; the core gain is sparse no-op/tangent repair, with wrong-gripper acting only as a marginal helper |
@@ -98,34 +101,35 @@ Suggested main-table rows:
 12. K4 train-state residual retrieval, mean-by-type tangent consensus
 13. K4 mean-by-type residual retrieval + fixed-scale no-op prior plateau, canonical 0.03
 14. K4 mean-by-type residual retrieval + scale-grid no-op prior 0.03
-15. K4 mean-by-type residual retrieval + upper/wide tangent-length diagnostics
-16. K4 mean-by-type residual retrieval + minimum-energy action penalty diagnostics
-17. K4 mean-by-type residual retrieval + source-progress/source-score/source-advantage prior diagnostics
-18. K4 mean-by-type residual retrieval + train-family success bonus diagnostics
-19. K4 mean-by-type residual retrieval + train-neighbor consensus-confidence diagnostics
-20. K4 repair-tangent residual transport diagnostics
-21. K4 mean-by-type residual retrieval + no-op-only family diagnostic
-22. K4 mean-by-type residual retrieval + abstention margin fine sweep
-23. Source-progress viability gate diagnostics
-24. K2/K4 task-relative retrieval metric diagnostics
-25. K4 kernel-weighted residual consensus + no-op prior diagnostics
-26. K4 field-softmax residual barycenter + margin diagnostics
-27. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
-28. K2 broad tangent ray-search
-29. Residual-tangent distillation policy
-30. Residual+Gaussian hybrid, K32 sigma0.35
-31. Lattice, near-miss only
-32. Lattice, no expert
-33. Lattice, no expert + policy baseline candidate
-34. Lattice, full
-35. Oracle ceiling
+15. K4 masked composed type-consensus residual retrieval + no-op prior 0.03
+16. K4 mean-by-type residual retrieval + upper/wide tangent-length diagnostics
+17. K4 mean-by-type residual retrieval + minimum-energy action penalty diagnostics
+18. K4 mean-by-type residual retrieval + source-progress/source-score/source-advantage prior diagnostics
+19. K4 mean-by-type residual retrieval + train-family success bonus diagnostics
+20. K4 mean-by-type residual retrieval + train-neighbor consensus-confidence diagnostics
+21. K4 repair-tangent residual transport diagnostics
+22. K4 mean-by-type residual retrieval + no-op-only family diagnostic
+23. K4 mean-by-type residual retrieval + abstention margin fine sweep
+24. Source-progress viability gate diagnostics
+25. K2/K4 task-relative retrieval metric diagnostics
+26. K4 kernel-weighted residual consensus + no-op prior diagnostics
+27. K4 field-softmax residual barycenter + margin diagnostics
+28. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+29. K2 broad tangent ray-search
+30. Residual-tangent distillation policy
+31. Residual+Gaussian hybrid, K32 sigma0.35
+32. Lattice, near-miss only
+33. Lattice, no expert
+34. Lattice, no expert + policy baseline candidate
+35. Lattice, full
+36. Oracle ceiling
 
 Suggested claim:
 
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
 > selection rule. Deployment-clean K4 consensus residual transport with advantage
-> abstention, a small typed no-op prior, and field-gated tangent length
-> calibration over a narrow local scale grid gives the strongest clean gain so far;
+> abstention, a small typed no-op prior, field-gated tangent length calibration,
+> and masked local tangent composition gives the strongest clean gain so far;
 > extending the scale grid upward and adding minimum-energy action regularization
 > are near-tie/negative, so the effect is local calibration rather than larger or
 > simply shorter steps. Train-source progress/reward-score priors provide cleaner
@@ -133,8 +137,10 @@ Suggested claim:
 > suggesting transferable residuals need not beat the expert anchor in their source
 > state. Continuous train-family success priors likewise tie or drop rather than
 > explain the top row. A train-neighbor consensus penalty is also negative/near-tie,
-> suggesting the current field already performs most of the useful abstention. Repair-tangent
-> transport is negative, showing that simply reversing train failures into
+> suggesting the current field already performs most of the useful abstention.
+> The clean composition row improves only after anti-goal composite masking, which
+> supports a controlled local tangent-chart story rather than an uncontrolled
+> proposal pile-up. Repair-tangent transport is negative, showing that simply reversing train failures into
 > near-miss-to-expert correction vectors is not the missing deployment proposal.
 > Ungated KNN residual
 > retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score/task-relative retrieval,

results/paper_story_memo.md

@@ -16,7 +16,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Same-state local counterfactual proposals are the mechanism | near-miss-only lattice is 55.94%; removing expert+near_miss drops to 25.57% | Strongly supported |
 | Conservative same-state result is large | no-expert lattice is 56.99% vs 29.74% policy | Main result |
 | Full lattice gives upper result | full lattice is 69.33%, oracle is 86.78% | Strong but label expert proposal clearly |
-| Deployment-clean proposal is currently a bottleneck | best clean K4 scale-grid consensus with a small typed no-op prior is 35.42%, far below 56.99% | Supported |
+| Deployment-clean proposal is currently a bottleneck | best clean K4 masked composed type-consensus transport with a small typed no-op prior is 35.54%, far below 56.99% | Supported |
 | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
 | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
 | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
@@ -26,27 +26,28 @@ when queried on proposal geometry that matches those local counterfactuals.
 | All-split field-teacher distillation does not fix checkpointing/coverage | allmap direct is 28.00%; field-guided best is 26.49% despite 100% target coverage | Negative diagnostic |
 | Residual family consistency improves clean transport | policy/no-op/wrong-gripper typed residuals reach 33.74%, above raw 33.33% | Supported as diagnostic |
 | Counterfactual advantage abstention improves clean transport | requiring field advantage over the zero-residual policy raises typed residual transport to 34.84%, and K2 retrieval reaches 35.01% | Supported as the previous clean best |
-| Clean residual transport behaves like sparse intervention | `paper_analysis.md` shows the best clean row abstains to zero-residual policy on 94.6% of states, while selected nonzero residuals succeed at 50.0% vs 34.58% for abstention | Stronger clean-mechanism framing |
+| Clean residual transport behaves like sparse intervention | the best clean row abstains to zero-residual policy on 93.2% of states, while selected nonzero residuals succeed at 49.6% vs 34.5% for abstention | Stronger clean-mechanism framing |
 | Tangent consensus is close but needs sparse typing | K4 mean-by-type residual consensus reaches 34.96%; a small no-op residual prior plateau at 0.025-0.035 raises fixed-scale transport to 35.25% | Fixed-scale clean diagnostic |
-| Field-gated tangent length calibration improves the clean bridge | K4 mean-by-type scale grid 0.35/0.40/0.45 with no-op bonus 0.03 reaches 35.42%; the source-score version reaches 35.30%. Upper 0.40/0.45/0.50 nearly ties at 35.36%, while wide 0.35/0.45/0.55 drops to 35.13% | Current best clean result; local scale calibration, not a larger-step effect |
-| Minimum-energy residual regularization does not add the gain | action L2 penalty 0.05 ties 35.42%, while 0.10/0.20 reach 35.36% | Negative/tie diagnostic: the clean bridge is not explained by shortest-action bias |
+| Field-gated tangent length calibration improves the clean bridge | K4 mean-by-type scale grid 0.35/0.40/0.45 with no-op bonus 0.03 reaches 35.42%; the source-score version reaches 35.30%. Upper 0.40/0.45/0.50 nearly ties at 35.36%, while wide 0.35/0.45/0.55 drops to 35.13% | Previous clean best; local scale calibration, not a larger-step effect |
+| Masked tangent composition gives a small clean lift | K4 composed type-consensus transport is 35.30% after anti-goal composite masking, and reaches 35.54% with the typed no-op prior; raw selected types contain no random-negative or wrong-direction composites | Current best clean result; controlled local tangent-chart composition, not unmasked proposal accumulation |
+| Minimum-energy residual regularization does not add the gain | action L2 penalty 0.05 ties the previous 35.42% scale-grid row, while 0.10/0.20 reach 35.36% | Negative/tie diagnostic: the clean bridge is not explained by shortest-action bias |
 | Source-advantage priors/gates are too brittle | source-advantage bonuses 0.02/0.05 reach 35.13%; no-op+advantage bonus reaches 35.30%; positive-advantage gates reach 35.13% with or without no-op prior | Negative diagnostic: useful transferable tangents need not beat the expert anchor in their own source state |
-| Continuous train-family success priors do not add the gain | scale-grid family-success bonuses 0.02/0.03/0.05 reach 35.25%; no-op+family-success 0.02 ties the best at 35.42% | Negative/tie diagnostic: train terminal success is not the right confidence signal for transferred tangents |
+| Continuous train-family success priors do not add the gain | scale-grid family-success bonuses 0.02/0.03/0.05 reach 35.25%; no-op+family-success 0.02 ties the previous scale-grid row at 35.42% | Negative/tie diagnostic: train terminal success is not the right confidence signal for transferred tangents |
 | Train-neighbor consensus confidence does not improve the top row | consensus-only 0.05 reaches 35.19%; no-op+consensus penalties 0.02/0.05/0.10 reach 35.36% | Negative/near-tie diagnostic: residual dispersion is a plausible confidence signal, but the field+margin already abstains better |
-| Repair-tangent transport is not the missing clean proposal | reversing residual direction to build near-miss/failure-to-expert tangents reaches only 34.14-34.43%, below the 35.42% scale-grid no-op row | Negative diagnostic: the failure-to-expert vector hypothesis is cleaner than a new prior, but does not explain the gap |
+| Repair-tangent transport is not the missing clean proposal | reversing residual direction to build near-miss/failure-to-expert tangents reaches only 34.14-34.43%, below the previous 35.42% scale-grid no-op row | Negative diagnostic: the failure-to-expert vector hypothesis is cleaner than a new prior, but does not explain the gap |
 | Kernel-weighted tangent interpolation does not beat equal consensus | K4 kernel-weighted residual consensus reaches 34.96%; with no-op prior and scales 0.35/0.40/0.45 it reaches 35.13%/35.19%/35.19%, below the 35.25% mean-consensus plateau | Negative/near-tie diagnostic |
 | Field-conditioned tangent barycenters identify good sparse corrections but do not close the proposal gap | K4 field-softmax transport reaches 34.96%; with no-op prior and margins 0.10/0.05/0.00 it reaches 35.19%/35.07%/34.84%. Selected aggregate residuals are high-value (up to 60.00% success), but selecting more of them degrades the global row | Negative/near-tie diagnostic |
-| Tangent ray-search does not beat the typed-prior clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below the scale-grid mean-consensus row at 35.42% | Near-tie/negative diagnostic |
+| Tangent ray-search does not beat the typed-prior clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below the previous scale-grid mean-consensus row at 35.42% | Near-tie/negative diagnostic |
 | Typed no-op residual prior improves the clean bridge | CPU smoke `14883591` passed; bonuses 0.025/0.03/0.035 tie at 35.25%, while 0.01/0.02/0.05/0.08 are slightly lower | Fixed-scale clean diagnostic |
 | Wrong-gripper typed prior does not add a new clean bridge | wrong-gripper-only reaches 35.19%; no-op+wrong-gripper 0.02 ties 35.25%; no-op+wrong-gripper 0.04 drops to 35.13% | Negative/tie diagnostic |
 | No-op-only residuals nearly preserve the fixed-scale clean bridge | excluding wrong-gripper residuals gives 35.19% with either no-op bonus 0.03 or source-score bonus 0.02, one success below the 35.25% fixed-scale safe-family plateau | Mechanism sharpened: wrong-gripper is marginal, not core |
-| The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.68 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.57 pp clean-to-same-state gap | Core paper tension |
+| The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.80 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.45 pp clean-to-same-state gap | Core paper tension |
 | Policy fallback is not the same-state mechanism | adding a policy baseline candidate to the no-expert same-state lattice drops 56.99% to 40.70% even with margin 0.00 | Negative diagnostic |
 | Z-score retrieval metric does not help | z-score rows reach 32.23-32.81%, below raw retrieval | Negative diagnostic |
 | Task-relative actor-pose retrieval metric does not improve tangent transfer | K2 task-relative residual retrieval reaches 34.26% vs raw K2 35.01%; K4 task-relative mean-by-type + no-op reaches 34.43% vs raw K4 35.25% | Negative diagnostic |
 | Train-source progress viability is too blunt a residual gate | source-progress thresholds 0.25/0.50/0.75 reach 35.19%/34.96%/34.72%, below the unfiltered no-op plateau at 35.25% | Negative/near-tie diagnostic |
 | Continuous train-source progress prior can replace the fixed-scale typed no-op prior but not improve it | source-progress bonus 0.03 ties the 35.25% fixed-scale row exactly; bonus 0.05 drops to 35.13% | Cleaner tie diagnostic |
-| Full train-source reward-score prior also ties fixed-scale but does not improve the clean best | source-score bonuses 0.015/0.020 tie 35.25%; scale-grid source-score reaches 35.30%, still below no-op scale-grid at 35.42% | Cleaner near-tie diagnostic |
+| Full train-source reward-score prior also ties fixed-scale but does not improve the clean best | source-score bonuses 0.015/0.020 tie 35.25%; scale-grid source-score reaches 35.30%, still below the previous no-op scale-grid row at 35.42% | Cleaner near-tie diagnostic |
 | Advantage margin 0.20 is a local optimum for K4 tangent consensus | no-op prior margins 0.15/0.20/0.25 reach 35.07%/35.25%/34.84%; source-score prior margins reach 34.96%/35.25%/34.84% | Abstention plateau sharpened |
 | Train-split residual family reliability does not recover the typed mask | after fixing threshold pass-through, scale-0.35 thresholds 0.10/0.25 reach 33.33%/33.28%, below typed safe residuals | Negative diagnostic |
 | Residual-tangent distillation does not solve clean proposal generation | aligned allmap tangent student reaches 28.87% despite low pseudo-target BC loss | Negative diagnostic |
@@ -74,33 +75,34 @@ clean proposal result, the intended main rows are:
 14. K4 mean-by-type tangent consensus: 34.96%
 15. K4 mean-by-type tangent consensus + typed no-op prior 0.025-0.035: 35.25%
 16. K4 mean-by-type tangent consensus + scale-grid typed no-op prior: 35.42%
-17. K4 mean-by-type tangent consensus + upper/wide scale diagnostics: 35.36% for 0.40/0.45/0.50; 35.13% for 0.35/0.45/0.55
-18. K4 mean-by-type tangent consensus + action L2 penalty: 35.42% at 0.05; 35.36% at 0.10/0.20
-19. K4 mean-by-type tangent consensus + train-source progress prior: 35.25% at bonus 0.03; 35.13% at bonus 0.05
-20. K4 mean-by-type tangent consensus + train-source reward-score prior: 35.25% at bonuses 0.015/0.020; 35.30% with scale grid; 35.19% at 0.025
-21. K4 mean-by-type tangent consensus + train-source advantage prior/gate: 35.13% at bonuses 0.02/0.05; 35.30% with no-op+advantage; 35.13% with positive-advantage gates
-22. K4 mean-by-type tangent consensus + train-family success bonus: 35.25% alone; 35.42% with no-op bonus 0.03
-23. K4 mean-by-type tangent consensus + train-neighbor consensus penalty: 35.19% alone; 35.36% with no-op bonus 0.03
-24. K4 repair-tangent transport: 34.14-34.43%
-25. K4 mean-by-type tangent consensus, no-op-only residuals: 35.19% with either no-op bonus 0.03 or source-score bonus 0.02
-26. K4 mean-by-type abstention margin sweep: 35.07% / 35.25% / 34.84% for typed no-op margins 0.15 / 0.20 / 0.25; 34.96% / 35.25% / 34.84% for source-score margins
-27. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
-28. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
-29. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
-30. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
-31. K2 broad tangent ray-search: 34.96%
-32. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-33. K4 tight tangent ray-search: 34.55%
-34. Residual-tangent distillation policy: 28.87%
-35. Z-score residual retrieval: 32.23-32.81%
-36. Task-relative residual retrieval metric: 34.26-34.43%
-37. Train-family reliability prior: 33.28-33.33%
-38. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-39. Lattice, near-miss only: 55.94%
-40. Lattice, no expert: 56.99%
-41. Lattice, no expert + policy baseline candidate: 40.70%
-42. Lattice, full: 69.33%
-43. Oracle ceiling: 86.78%
+17. K4 masked composed type-consensus tangent transport: 35.30%; with typed no-op prior: 35.54%
+18. K4 mean-by-type tangent consensus + upper/wide scale diagnostics: 35.36% for 0.40/0.45/0.50; 35.13% for 0.35/0.45/0.55
+19. K4 mean-by-type tangent consensus + action L2 penalty: 35.42% at 0.05; 35.36% at 0.10/0.20
+20. K4 mean-by-type tangent consensus + train-source progress prior: 35.25% at bonus 0.03; 35.13% at bonus 0.05
+21. K4 mean-by-type tangent consensus + train-source reward-score prior: 35.25% at bonuses 0.015/0.020; 35.30% with scale grid; 35.19% at 0.025
+22. K4 mean-by-type tangent consensus + train-source advantage prior/gate: 35.13% at bonuses 0.02/0.05; 35.30% with no-op+advantage; 35.13% with positive-advantage gates
+23. K4 mean-by-type tangent consensus + train-family success bonus: 35.25% alone; 35.42% with no-op bonus 0.03
+24. K4 mean-by-type tangent consensus + train-neighbor consensus penalty: 35.19% alone; 35.36% with no-op bonus 0.03
+25. K4 repair-tangent transport: 34.14-34.43%
+26. K4 mean-by-type tangent consensus, no-op-only residuals: 35.19% with either no-op bonus 0.03 or source-score bonus 0.02
+27. K4 mean-by-type abstention margin sweep: 35.07% / 35.25% / 34.84% for typed no-op margins 0.15 / 0.20 / 0.25; 34.96% / 35.25% / 34.84% for source-score margins
+28. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
+29. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
+30. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
+31. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+32. K2 broad tangent ray-search: 34.96%
+33. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+34. K4 tight tangent ray-search: 34.55%
+35. Residual-tangent distillation policy: 28.87%
+36. Z-score residual retrieval: 32.23-32.81%
+37. Task-relative residual retrieval metric: 34.26-34.43%
+38. Train-family reliability prior: 33.28-33.33%
+39. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+40. Lattice, near-miss only: 55.94%
+41. Lattice, no expert: 56.99%
+42. Lattice, no expert + policy baseline candidate: 40.70%
+43. Lattice, full: 69.33%
+44. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -128,13 +130,13 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 12:03 UTC`. The K4 mean-by-type scale-grid sweep
-completed and produced a new clean best, 35.42%, while upper/wide,
-minimum-energy, source-advantage, and train-family success-prior follow-ups
-completed without improving it. Consensus-confidence follow-ups also completed
-below the best row. Repair-tangent follow-ups completed below the best row too,
-with a best repair result of 34.43%. The paper table/paired analysis continue
-to use the scale-grid no-op row as `best_clean_key`.
+Last checked: `2026-06-29 16:12 UTC`. The K4 masked composed type-consensus
+sweep completed and produced a new clean best, 35.54%, while the pure masked
+composition row reached 35.30%. Raw selected candidate types show no
+random-negative or wrong-direction composite leak. The paper table and paired
+analysis now use the masked composed no-op row as `best_clean_key`. A
+component-wise composite prior follow-up has passed CPU smoke and is queued for
+3-seed GPU rollout.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -250,14 +252,14 @@ to use the scale-grid no-op row as `best_clean_key`.
 - `14903128`/`14903130`/`14903132`/`14903134`: completed continuous
   train-family success-prior GPU arrays. Family-success bonuses `0.02`, `0.03`,
   and `0.05` reach 35.25%; adding family-success `0.02` to the no-op `0.03`
-  best row ties 35.42% without adding a new gain. Summary jobs `14903129`/
+  row ties the previous 35.42% scale-grid result without adding a new gain. Summary jobs `14903129`/
   `14903131`/`14903133`/`14903135` and rebuild job `14903136` completed.
 - `14903296`: completed CPU smoke for the train-neighbor consensus-confidence
   penalty path, validating metadata and Slurm/CLI wiring.
 - `14903384`/`14903386`/`14903388`/`14903390`: completed consensus-confidence
   GPU arrays. Consensus-only `0.05` reaches 35.19%; no-op `0.03` plus
   consensus penalties `0.02`, `0.05`, and `0.10` all reach 35.36%, one success
-  below the 35.42% best. Summary jobs `14903385`/`14903387`/`14903389`/
+  below the previous 35.42% scale-grid best. Summary jobs `14903385`/`14903387`/`14903389`/
   `14903391` and rebuild job `14903392` completed.
 - `14904575`: completed CPU smoke for repair-tangent residual direction
   (`anchor_minus_candidate`). The smoke wrote valid metadata and selected the
@@ -267,6 +269,21 @@ to use the scale-grid no-op row as `best_clean_key`.
   repair row reaches 34.43%. Summary jobs `14904738`/`14904741`/`14904743`/
   `14904745` completed, local paper builders updated the artifacts, and the
   queued rebuild job `14904803` was canceled after local rebuilds finished.
+- `14911977`: completed CPU smoke for masked composed type-consensus transport.
+  The smoke selected only `policy_residual` on 8 groups and confirmed composite
+  candidate masking excludes random-negative and wrong-direction parts.
+- `14911979`/`14911980`: completed K4 masked composed type-consensus GPU arrays.
+  Pure masked composition reaches 35.30%; adding the typed no-op prior reaches
+  the new clean best, 35.54%. Raw selected candidate types contain no
+  random-negative or wrong-direction composites. Summary jobs `14911982`/
+  `14911983` and rebuild job `14911984` completed.
+- `14912552`: completed CPU smoke for component-wise candidate-type bonuses on
+  masked composed type-consensus transport. Metadata records
+  `candidate_type_bonus_components=True`, selected candidate types have no
+  random-negative or wrong-direction leak, and the smoke selected
+  `policy_residual` on 8/8 groups.
+- `14912561`/`14912562`/`14912563`: queued component-wise composite-prior GPU
+  array, summary, and paper-artifact rebuild behind the passed smoke.
 - `14894281`: completed the Apptainer unit smoke for the train-source
   progress-viability gate, including the variable residual-count padding check
   (`source_progress_lengths == [3, 3]`).
@@ -284,14 +301,14 @@ to use the scale-grid no-op row as `best_clean_key`.
   train-source progress bonus path. The unit smoke validated bonus padding
   (`source_progress_bonuses == [[0, 0.08, 0], [0, 0.08, 0.08]]`).
 - `14894674`/`14894675`: completed source-progress bonus arrays with no fixed
-  no-op prior. Bonus `0.03` ties the current best at 35.25%; bonus `0.05`
+  no-op prior. Bonus `0.03` ties the fixed-scale plateau at 35.25%; bonus `0.05`
   reaches 35.13%. Summary jobs `14894676`/`14894677` completed; rebuild job
   `14894678` was queued after them.
 - `14897121`/`14897122`: completed unit and CPU rollout smokes for the
   train-source reward-score bonus path. The unit smoke validates that terminal
   success contributes to the candidate prior.
 - `14897123`/`14897124`/`14897125`: completed source-score bonus arrays.
-  Bonuses `0.015` and `0.020` tie the current best at 35.25%; bonus `0.025`
+  Bonuses `0.015` and `0.020` tie the fixed-scale plateau at 35.25%; bonus `0.025`
   reaches 35.19%. Summary jobs `14897126`/`14897127`/`14897128` and rebuild job
   `14897129` completed.
 - `14897548`/`14897549`: completed no-op-only CPU rollout smokes after excluding
@@ -308,15 +325,15 @@ to use the scale-grid no-op row as `best_clean_key`.
   `14897845`-`14897848` and rebuild job `14897849` completed.
 - `14897988`/`14897989`: completed K4 mean-by-type scale-grid sweeps using
   scales `0.35/0.40/0.45`, margin `0.20`, and safe residual families. The typed
-  no-op prior row reaches a new clean best, 35.42%; the source-score prior row
+  no-op prior row reaches a then-new clean best, 35.42%; the source-score prior row
   reaches 35.30%. Summary jobs `14897990`/`14897991` completed; rebuild job
   `14897992` was submitted, and local rebuilds updated the paper artifacts.
 - `14898107`/`14898108`/`14898109`: completed upper and wide K4 mean-by-type
   scale-grid follow-ups. The no-op upper grid `0.40/0.45/0.50` reaches 35.36%,
   the source-score upper grid reaches 35.30%, and the no-op wide grid
   `0.35/0.45/0.55` reaches 35.13%. Summary jobs `14898110`/`14898111`/
-  `14898112` and rebuild job `14898113` completed; the best clean row remains
-  the `0.35/0.40/0.45` no-op grid at 35.42%.
+  `14898112` and rebuild job `14898113` completed; at that stage the best clean
+  row remained the `0.35/0.40/0.45` no-op grid at 35.42%.
 - `14898293`: completed the CPU Apptainer smoke for the residual action-L2
   penalty path with the best scale-grid/no-op configuration.
 - `14898327`/`14898329`/`14898331`: completed minimum-energy tangent GPU sweeps
@@ -350,10 +367,13 @@ to use the scale-grid no-op row as `best_clean_key`.
 
 - Promote same-state no-expert lattice (56.99%) as the conservative mechanism
   result.
-- Use K4 mean-by-type residual transport with advantage abstention, a small
-  typed no-op prior, and field-gated tangent length calibration over
-  `0.35/0.40/0.45` as the current best clean deployment diagnostic, 35.42%, not
-  as a SOTA claim. The fixed-scale no-op plateau remains 35.25%; continuous
+- Use K4 masked composed type-consensus residual transport with advantage
+  abstention, a small typed no-op prior, and field-gated tangent length
+  calibration over `0.35/0.40/0.45` as the current best clean deployment
+  diagnostic, 35.54%, not as a SOTA claim. The previous mean-by-type scale-grid
+  no-op row remains 35.42%; pure masked composition is 35.30%, so composition
+  helps only when the sparse typed prior is preserved and anti-goal composite
+  parts are masked. The fixed-scale no-op plateau remains 35.25%; continuous
   train-source progress/reward-score priors tie that fixed-scale row, and
   scale-grid source-score reaches 35.30% but not the new best. Source-advantage
   priors/gates reach at most 35.30%, so local utility lift over the source
@@ -374,6 +394,6 @@ to use the scale-grid no-op row as `best_clean_key`.
   repaired train-family reliability priors, Gaussian hybrids,
   field optimization, field-teacher/tangent distillation, repair-tangent transport, policy-relative anchoring, tangent consensus,
   kernel-weighted tangent interpolation, field-softmax tangent barycenters,
-  wrong-gripper typed priors, and same-state policy-baseline fallback as negative
-  or near-tie diagnostics that sharpen the story around local counterfactual
-  proposal geometry.
+  unmasked or prior-free tangent composition, wrong-gripper typed priors, and
+  same-state policy-baseline fallback as negative or near-tie diagnostics that
+  sharpen the story around local counterfactual proposal geometry.

results/paper_table_status.json

@@ -1296,6 +1296,25 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05797101449275366
     },
+    {
+      "key": "retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
+      "label": "K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "component-wise sparse prior on the masked local tangent composition chart",
+      "fallback_success": null,
+      "pending_job": "14912561/14912562",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
     {
       "key": "retrieval_repair_nearmiss_k4_grid025035050_margin020",
       "label": "K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

results/paper_table_status.md

@@ -71,6 +71,7 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 34.14% | +4.41 pp | yes | no | no | local tangent composition on the current best typed prior |
 | retrieval_residual_k4_composemasked_grid035040045 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20 | complete | 35.30% | +5.57 pp | yes | no | no | local tangent composition with anti-goal composite masks |
 | retrieval_residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 35.54% | +5.80 pp | yes | no | no | local tangent composition with anti-goal composite masks on the current best typed prior |
+| retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03 | pending 14912561/14912562 | pending | pending | yes | no | no | component-wise sparse prior on the masked local tangent composition chart |
 | retrieval_repair_nearmiss_k4_grid025035050_margin020 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | complete | 34.32% | +4.58 pp | yes | no | no | deployment-clean corrective tangent transport from train near-misses back toward expert actions |
 | retrieval_repair_nearmiss_k4_grid035050075_margin020 | K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20 | complete | 34.38% | +4.64 pp | yes | no | no | repair-tangent scale diagnostic for near-miss counterfactual geometry |
 | retrieval_repair_nearmiss_k4_grid025035050_margin010 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10 | complete | 34.14% | +4.41 pp | yes | no | no | repair-tangent abstention diagnostic for near-miss counterfactual geometry |

scripts/build_paper_analysis.py

@@ -353,6 +353,14 @@ METHODS = [
             "k4_composemasked_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
+        label="K4 composed type-consensus tangents, masked, component no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+        ),
+    ),
     MethodSpec(
         key="repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/build_paper_table_status.py

@@ -695,6 +695,16 @@ SPECS = [
         story_role="local tangent composition with anti-goal composite masks on the current best typed prior",
         pending_job="14911980/14911983",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003",
+        label="K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="component-wise sparse prior on the masked local tangent composition chart",
+        pending_job="14912561/14912562",
+    ),
     ResultSpec(
         key="retrieval_repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/eval_maniskill_policy_rollout.py

@@ -239,6 +239,12 @@ def main(argv: list[str] | None = None) -> int:
         help="Comma-separated candidate_type=bonus priors added to field potentials before "
         "selection, e.g. 'residual_no_op=0.05'. Empty preserves previous behavior.",
     )
+    parser.add_argument(
+        "--candidate-type-bonus-components",
+        action="store_true",
+        help="Let composite candidate types inherit the sum of configured component bonuses "
+        "unless an exact composite bonus is configured.",
+    )
     args = parser.parse_args(argv)
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
@@ -309,6 +315,7 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_reduce=args.retrieval_residual_reduce,
         lattice_exclude_types=lattice_exclude_types,
         candidate_type_bonuses=candidate_type_bonuses,
+        candidate_type_bonus_components=args.candidate_type_bonus_components,
     )
     print(json.dumps({key: value for key, value in result.items() if key != "rows"}, indent=2))
     return 0

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -77,6 +77,7 @@ CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES:-}"
 if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
   CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES_COLON//:/,}"
 fi
+CANDIDATE_TYPE_BONUS_COMPONENTS="${CANDIDATE_TYPE_BONUS_COMPONENTS:-0}"
 
 module load StdEnv/2023 apptainer/1.4.5
 cd "$PROJECT_DIR"
@@ -102,6 +103,9 @@ fi
 if [[ "$PREPEND_POLICY_CANDIDATE" == "1" ]]; then
   EXTRA_ARGS+=(--prepend-policy-candidate)
 fi
+if [[ "$CANDIDATE_TYPE_BONUS_COMPONENTS" == "1" ]]; then
+  EXTRA_ARGS+=(--candidate-type-bonus-components)
+fi
 
 apptainer exec --nv \
   --env "LD_LIBRARY_PATH=$CPU_RENDER_LIBS/lib:$NATIVE_LIBS:/.singularity.d/libs,VK_ICD_FILENAMES=$VULKAN_ICD,VK_DRIVER_FILES=$VULKAN_ICD,XDG_RUNTIME_DIR=$RUNTIME_DIR,MESA_SHADER_CACHE_DIR=$CACHE_DIR,LIBGL_ALWAYS_SOFTWARE=1,LP_NUM_THREADS=1,SSL_CERT_FILE=$CA_BUNDLE,REQUESTS_CA_BUNDLE=$CA_BUNDLE,OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1,DOVLA_TORCH_THREADS=1,MPLBACKEND=Agg,PYTHONDONTWRITEBYTECODE=1" \

scripts/slurm/eval_maniskill_policy_rollout_cpu_smoke.sbatch

@@ -76,6 +76,7 @@ CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES:-}"
 if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
   CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES_COLON//:/,}"
 fi
+CANDIDATE_TYPE_BONUS_COMPONENTS="${CANDIDATE_TYPE_BONUS_COMPONENTS:-0}"
 
 module load StdEnv/2023 apptainer/1.4.5
 cd "$PROJECT_DIR"
@@ -98,6 +99,9 @@ fi
 if [[ "$MAX_GROUPS" != "all" ]]; then
   EXTRA_ARGS+=(--max-groups "$MAX_GROUPS")
 fi
+if [[ "$CANDIDATE_TYPE_BONUS_COMPONENTS" == "1" ]]; then
+  EXTRA_ARGS+=(--candidate-type-bonus-components)
+fi
 
 apptainer exec \
   --env "LD_LIBRARY_PATH=$CPU_RENDER_LIBS/lib:$NATIVE_LIBS,VK_ICD_FILENAMES=$VULKAN_ICD,VK_DRIVER_FILES=$VULKAN_ICD,XDG_RUNTIME_DIR=$RUNTIME_DIR,MESA_SHADER_CACHE_DIR=$CACHE_DIR,LIBGL_ALWAYS_SOFTWARE=1,LP_NUM_THREADS=1,SSL_CERT_FILE=$CA_BUNDLE,REQUESTS_CA_BUNDLE=$CA_BUNDLE,OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1,DOVLA_TORCH_THREADS=1,MPLBACKEND=Agg,PYTHONDONTWRITEBYTECODE=1" \

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -99,6 +99,9 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             ),
             "retrieval_residual_reduce": data.get("retrieval_residual_reduce", "none"),
             "candidate_type_bonuses": data.get("candidate_type_bonuses", {}),
+            "candidate_type_bonus_components": data.get(
+                "candidate_type_bonus_components", False
+            ),
             "selected_residual_scale_counts": dict(selected_scale_counts),
             "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
             "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),

tests/test_maniskill_policy_rollout.py

@@ -20,6 +20,7 @@ from dovla_cil.eval.maniskill_policy_rollout import (
     _attach_retrieved_residual_candidates,
     _effective_lattice_candidate_count,
     _lattice_candidate_mask,
+    _lattice_candidate_type_bonus,
     _load_state_archive,
     _numeric_action_values,
     _reduce_residual_candidates_by_type,
@@ -387,6 +388,60 @@ def test_lattice_candidate_mask_excludes_composite_candidate_parts() -> None:
     assert mask.tolist() == [[True, False, True]]
 
 
+def test_lattice_candidate_type_bonus_can_sum_composite_parts() -> None:
+    import torch
+
+    case = _RolloutCase(
+        group_id="g",
+        task_id="PickCube-v1",
+        source_dataset=Path("."),
+        state={},
+        observation={"features": [0.0]},
+        instruction="pick",
+        oracle_score=1.0,
+        oracle_success=True,
+        expert_score=1.0,
+        expert_success=True,
+        best_action_values=[[0.0]],
+        candidate_action_values=[[[0.0]], [[0.2]], [[0.4]]],
+        candidate_types=[
+            "policy_residual",
+            "residual_no_op+residual_wrong_gripper",
+            "residual_no_op+residual_random_negative",
+        ],
+        candidate_score_bonuses=[0.0, 0.01, 0.0],
+    )
+
+    exact_only = _lattice_candidate_type_bonus(
+        [case],
+        torch=torch,
+        device="cpu",
+        candidate_type_bonuses={"residual_no_op": 0.03, "residual_wrong_gripper": 0.02},
+    )
+    component = _lattice_candidate_type_bonus(
+        [case],
+        torch=torch,
+        device="cpu",
+        candidate_type_bonuses={"residual_no_op": 0.03, "residual_wrong_gripper": 0.02},
+        use_components=True,
+    )
+    exact_override = _lattice_candidate_type_bonus(
+        [case],
+        torch=torch,
+        device="cpu",
+        candidate_type_bonuses={
+            "residual_no_op": 0.03,
+            "residual_wrong_gripper": 0.02,
+            "residual_no_op+residual_wrong_gripper": 0.07,
+        },
+        use_components=True,
+    )
+
+    assert torch.allclose(exact_only, torch.tensor([[0.0, 0.01, 0.0]]))
+    assert torch.allclose(component, torch.tensor([[0.0, 0.06, 0.03]]))
+    assert torch.allclose(exact_override, torch.tensor([[0.0, 0.08, 0.03]]))
+
+
 def test_lattice_mode_can_prepend_policy_baseline_for_margin_abstention() -> None:
     import torch