official/main/single_lift_cloth_1/runs/c5-single-lift-cloth-1-photo: package_final batch 1/1

official/main/single_lift_cloth_1/runs/c5-single-lift-cloth-1-photo/README.md

@@ -0,0 +1,43 @@
+# c5-single-lift-cloth-1-photo
+
+Official PhysTwin photometric refinement of the authors' supervised BEST for single_lift_cloth_1 (main): reduced CMA and projected Adam, with every saved checkpoint and retained version.
+
+| Field | Value |
+|---|---|
+| Status | complete — budget complete: 100 CMA generations, 800 asks, 500 Adam traversals; 5 nonfinite optimizer groups disclosed |
+| Identity | source_group `main`, case `single_lift_cloth_1` |
+| Parent | authors' supervised BEST `Jianghanxiao/PhysTwin@05665bcd0eda:experiments.zip:experiments/single_lift_cloth_1/train/best_0.pth` |
+| Stages | photometric_cma, photometric_adam |
+| Budget (completed/configured) | photometric_cma: generations 100/100 (evaluations 800); photometric_adam: traversals 500/500, updates 38000/38000 |
+| Labels | BEST adam/pass_0473, FINAL adam/final |
+| Code revision(s) | `d917d092df1b800c7dbdd212168d9d91d94ab3d5`, `019f3dc1821ad8287be6a26f463865525025f627`; dirty tree by stage: photometric_cma not recorded, photometric_adam not recorded |
+| Checkpoint rows | 1673 (`checkpoints.tsv`) |
+| Containers | 2,987 native and evidence files packed in 10 uncompressed tar containers by category (adam_continuation 1, adam_update_rows 1, cma_candidate 1, frozen_score 5, records 1, retained_version 1); `containers/index.json` lists each member's path, offset, size and sha256, and `checkpoints.tsv` still lists every checkpoint file |
+| Actual supervision | photometric images of the supplied views over the train prefix: RGB, silhouette and object depth (no track, chamfer or acceleration term); see the objective record; the authors' supervised parent is lineage, not this objective |
+| Clock and consumed damping | interval 0.033350000000000005 s (dt 5e-05 x 667 substeps); drag 8, dashpot 150 (int-consumed parent values; CMA and Adam stage candidates are in their score records) |
+| Retained versions | 366 older files (504,051,869 B) under `evidence/retained_versions/`, 1 of them from preserved delta bundles; `run.json` `retained_versions` lists each file's source and aliases |
+| Execution | admitted at `d917d092df1b800c7dbdd212168d9d91d94ab3d5`; segment records name `019f3dc1821a`, `d917d092df1b` |
+
+## Stages
+
+- **photometric_cma** (`native/run_config.json`): {"objective_id": "real_rgbd_active_photometric_v1", "source": "U gs_train.py:128-188 + loss_utils.py (PhysTwin 81c718790a37e5e0102eb77af2c6edd34a9db25f)", "weights": {"lambda_dssim": 0.2, "lambda_seg": 1.0, "lambda_depth": 0.001}, "rgb_ssim": "fused_ssim (U gaussian_splatting/submodules/fused-ssim, padding same, train True)", "seg_ssim": "reference ssim (U loss_utils.ssim, 11x11 window sigma 1.5, 
+- **photometric_adam** (`native/run_config.json`): {"objective_id": "real_rgbd_active_photometric_v1", "source": "U gs_train.py:128-188 + loss_utils.py (PhysTwin 81c718790a37e5e0102eb77af2c6edd34a9db25f)", "weights": {"lambda_dssim": 0.2, "lambda_seg": 1.0, "lambda_depth": 0.001}, "rgb_ssim": "fused_ssim (U gaussian_splatting/submodules/fused-ssim, padding same, train True)", "seg_ssim": "reference ssim (U loss_utils.ssim, 11x11 window sigma 1.5, 
+  - stiffness update: projected log-Y update after every Adam step (cap 11.512925148010254)
+
+## Capabilities
+
+- **portable_use**: native arrays and the run-bound appearance binding retained; published-model use of this family is not implemented yet (plan P3), so no fresh use is claimed
+- **interrupted_run_recovery**: unsupported for this publication (owner D21): every photometric checkpoint and its continuation state (optimizer moments, cursor, q/v and appearance carry, CMA strategy and asks) is retained, and its resume inputs are published or referenced (D15-D20), but no photometric restore is implemented or claimed
+- **extend_completed_budget**: not supported: the native protocols refuse to extend a finished run's horizon
+- **full_reproduction**: not claimed (D3): dirty-state and environment limits are listed under code; the authors' inputs must be fetched at their pinned members
+
+## Limitations
+
+- Saved analysis reports 5 × nonfinite optimizer exp_avg_sq (evidence/analysis/report.json integrity_issues): the budget completed, but the optimizer state is not finite. The original bytes are kept unchanged; completing the budget does not certify optimizer integrity, fit quality or online readiness.
+- Segment records show the refinement ran from 2 checkouts (/mnt/kostas-graid/sw/envs/leonmkim/worktrees/belt_perception_wt_c5-official-sm75-grasp-exec-20261029/scripts/phystwin_official.py, /vast/projects/posa/contact-rich-robot/worktrees/belt_perception_wt_c5_parcc_resume_exec_20261030/scripts/phystwin_official.py). A checkout's commit is known only where execution_provenance names it; run_config keeps the admitted revision.
+- Native score payloads stay whole: each adam/pass_*/score.npz, p0 and c0 combines trajectory, final state, components and leaves, and BEST is the pass that selection.json names rather than a separate export.
+- photometric_cma: records name d917d092df1b, 019f3dc1821a but none states whether the working tree was clean when this stage ran
+- photometric_adam: records name d917d092df1b, 019f3dc1821a but none states whether the working tree was clean when this stage ran
+- records name more than one revision (resumed or continued segments); see stages and observations
+
+Native records under `native/` are authoritative; `run.json`, `checkpoints.tsv` and `dependencies.json` index them. License: see the repository `LICENSE` and card.