sessions/session_20260921_225908_522335/runs/c6-225908-522335-photo: package_final batch 1/1

sessions/session_20260921_225908_522335/runs/c6-225908-522335-photo/README.md

@@ -0,0 +1,39 @@
+# c6-225908-522335-photo
+
+C6 photo_production production fit of session 225908_522335; configured CMA and 500 Adam traversals completed, after the 2026-11-03 corpus cutoff.
+
+| Field | Value |
+|---|---|
+| Status | complete |
+| Identity | session_id `session_20260921_225908_522335`, window_id `anchor0328-0560` |
+| Parent | `c6-225908-522335-tracking` adam/c0086 (Exact tracking snapshot selected by the saved photometric scan) |
+| Stages | cma, photometric_adam |
+| Budget (completed/configured) | cma:  (completed_generations 100, issued_asks 800, told_asks 800, physical_failures 0); photometric_adam: traversals 500/500, updates 116000/116000 |
+| Labels | BEST adam/c0280, FINAL adam/c0500 |
+| Code revision(s) | `1341c18fc121f59ec4cd3da6883c5299b4aaae85`; dirty tree by stage: cma reported clean, photometric_adam not recorded |
+| Checkpoint rows | 3048 (`checkpoints.tsv`) |
+| Containers | 9,031 native and evidence files packed in 34 uncompressed tar containers by category (adam_continuation 25, adam_snapshot 1, adam_update_rows 1, cma_candidate 1, cma_generation_state 1, frozen_score 3, logs 1, records 1); `containers/index.json` lists each member's path, offset, size and sha256, and `checkpoints.tsv` still lists every checkpoint file |
+
+## Stages
+
+- **cma** (`native/cma/manifest.json`): "mean of all four views per transition, then mean over K; RP1 active Q5 objective"
+- **photometric_adam** (`native/adam/manifest.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: finite raw log Y moved back onto the physical cap after every update (protocol.projected_log_y true)
+
+## Capabilities
+
+- **portable_use**: native model arrays retained; fresh use (canonical q0/v0, zero updates) from a clean local copy is implemented and validated by PR4 on representative checkpoints; Hub fetch pending; photometric rendering also needs the recorded initial appearance and binding listed in dependencies.json
+- **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
+
+## Limitations
+
+- continuing this fit re-hashes author weights TAPNext++ tapnextpp_512 checkpoint (Google DeepMind TAPNet) (/mnt/data2/magna/belt_perception/results/phystwin_real_rgbd/session_20260921_225630_786148/rt_waft_tapnextpp_refit_01/checkpoints/tapnextpp_512.ckpt); they are not published here (dependencies.json third_party_weights names their source and sha256)
+- historical depth: the run's records name depth_source waft_stereo_dav2l5 but no per-stream digest, so the consumed depth bytes and their publication revision are unresolved; correction ledgers replaced 4 depth paths of this session (c9fa16c0c8ac)
+- photometric_adam: records name 1341c18fc121 but none states whether the working tree was clean when this stage ran
+- Completion records establish completion of the configured work, not owner acceptance of model quality.
+- The accepted input has disclosed appearance limitations.
+- Photometric continuation remains unsupported. The cache bytes are preserved, but its original identity includes source video mtimes; relocated cached replay has not been qualified.
+
+Native records under `native/` are authoritative; `run.json`, `checkpoints.tsv` and `dependencies.json` index them. License: see the repository `LICENSE` and card.

sessions/session_20260921_225908_522335/runs/c6-225908-522335-photo/run.json

@@ -0,0 +1,660 @@
+{
+  "aliases": [
+    {
+      "bytes_identical": true,
+      "duplicate_of": "native/cma/generations/g072/a03/model/model.npz",
+      "path": "native/cma/winner/model/model.npz",
+      "relation": "byte copy of the committed ask (cma._stage_winner)",
+      "role": "CMA winner"
+    },
+    {
+      "bytes_identical": true,
+      "duplicate_of": "native/cma/generations/g072/a03/model/topology.npz",
+      "path": "native/cma/winner/model/topology.npz",
+      "relation": "byte copy of the committed ask (cma._stage_winner)",
+      "role": "CMA winner"
+    }
+  ],
+  "capabilities": {
+    "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",
+    "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",
+    "portable_use": "native model arrays retained; fresh use (canonical q0/v0, zero updates) from a clean local copy is implemented and validated by PR4 on representative checkpoints; Hub fetch pending; photometric rendering also needs the recorded initial appearance and binding listed in dependencies.json"
+  },
+  "checkpoints": {
+    "counts": {
+      "adam_continuation": 1641,
+      "adam_traversal_snapshot": 501,
+      "adam_zero_update_check": 1,
+      "cma_candidate": 800,
+      "cma_generation_state": 100,
+      "cma_to_adam_handoff": 1,
+      "cma_winner": 1,
+      "model_initialization": 2,
+      "photometric_parent_p0": 1
+    },
+    "rows": 3048,
+    "table": "checkpoints.tsv"
+  },
+  "code": {
+    "belt_perception_commits": [
+      "1341c18fc121f59ec4cd3da6883c5299b4aaae85"
+    ],
+    "dirty_reported": null,
+    "limitations": [
+      "photometric_adam: records name 1341c18fc121 but none states whether the working tree was clean when this stage ran"
+    ],
+    "observations": [
+      {
+        "key": "code_revision",
+        "scope": "cma",
+        "source": "native/cma/manifest.json",
+        "value": {
+          "dirty": false,
+          "head": "1341c18fc121f59ec4cd3da6883c5299b4aaae85"
+        }
+      },
+      {
+        "key": "revision",
+        "scope": "run",
+        "source": "native/campaign.json",
+        "value": "1341c18fc121f59ec4cd3da6883c5299b4aaae85"
+      }
+    ],
+    "stages": {
+      "cma": {
+        "belt_perception_commits": [
+          "1341c18fc121f59ec4cd3da6883c5299b4aaae85"
+        ],
+        "dirty_reported": false,
+        "sources": [
+          "native/cma/manifest.json:code_revision",
+          "native/campaign.json:revision"
+        ]
+      },
+      "photometric_adam": {
+        "belt_perception_commits": [
+          "1341c18fc121f59ec4cd3da6883c5299b4aaae85"
+        ],
+        "dirty_reported": null,
+        "sources": [
+          "native/campaign.json:revision"
+        ]
+      }
+    }
+  },
+  "containers": {
+    "categories": {
+      "adam_continuation": {
+        "containers": 25,
+        "member_bytes": 26322951552,
+        "members": 1641
+      },
+      "adam_snapshot": {
+        "containers": 1,
+        "member_bytes": 223439252,
+        "members": 1506
+      },
+      "adam_update_rows": {
+        "containers": 1,
+        "member_bytes": 521877163,
+        "members": 1641
+      },
+      "cma_candidate": {
+        "containers": 1,
+        "member_bytes": 575076759,
+        "members": 3200
+      },
+      "cma_generation_state": {
+        "containers": 1,
+        "member_bytes": 5693030,
+        "members": 500
+      },
+      "frozen_score": {
+        "containers": 3,
+        "member_bytes": 2956343025,
+        "members": 502
+      },
+      "logs": {
+        "containers": 1,
+        "member_bytes": 35996,
+        "members": 8
+      },
+      "records": {
+        "containers": 1,
+        "member_bytes": 83158076,
+        "members": 33
+      }
+    },
+    "containers": 34,
+    "encoded_bytes": 30695720960,
+    "index": "containers/index.json",
+    "index_sha256": "5f42aa6a2a6b0529d13921d3286b8f4ec0047eb29a709643a8600d99c8f02433",
+    "individual_payload_files": 0,
+    "member_bytes": 30688574853,
+    "members": 9031,
+    "package_files": 39,
+    "rule": "Scaling plan \u00a73.2 (owner D30, 2026-11-01): run.json, checkpoints.tsv, dependencies.json, README.md, this index, every BEST/FINAL export component (preflight copies included) and every file whose own container would exceed 1 GiB stay individual files. Other files under native/, evidence/ and inputs/ are packed in path order into uncompressed USTAR containers of at most 1 GiB encoded size per category: the per-step history categories (CMA candidates and generation state, Adam snapshots, continuations, update rows, pass records and scores), 'logs' (.log/.txt/.md/.out/.err) and 'records' (everything else). Member bytes and logical paths are unchanged; each member's data offset is listed, and scripts/phystwin_publish.py fetch restores the native layout.",
+    "schema": "hf_publication_containers_v2"
+  },
+  "description": "C6 photo_production production fit of session 225908_522335; configured CMA and 500 Adam traversals completed, after the 2026-11-03 corpus cutoff.",
+  "environment": {
+    "native/adam/manifest.json:identity.execution": {
+      "belt_perception_source": {
+        "files": 954,
+        "sha256": "31646433c2e7a47c6ee189d96b254430775e0f63c0600cd2c391166ea01ce77a"
+      },
+      "device_name": "NVIDIA GeForce RTX 4090",
+      "python": "3.11.14",
+      "torch_cuda": "12.8",
+      "versions": {
+        "cma": "4.4.4",
+        "numpy": "2.4.4",
+        "open3d": "0.19.0",
+        "torch": "2.11.0+cu128",
+        "warp-lang": "1.16.0"
+      }
+    },
+    "native/campaign.json:device_identities": {
+      "cuda:0": {
+        "name": "NVIDIA GeForce RTX 4090",
+        "total_memory": 25215303680,
+        "uuid": "30be4c6b-0cc3-a19f-14c8-b53e90ace6dc"
+      }
+    },
+    "native/campaign.json:lock_sha256": "1e86a807f5e15de4e9c46e76c4b8c89740dbde8ee572b8d3f288d654416c75ed",
+    "native/cma/manifest.json:execution_identity": {
+      "belt_perception_source": {
+        "files": 954,
+        "sha256": "31646433c2e7a47c6ee189d96b254430775e0f63c0600cd2c391166ea01ce77a"
+      },
+      "device_name": "NVIDIA GeForce RTX 4090",
+      "python": "3.11.14",
+      "torch_cuda": "12.8",
+      "versions": {
+        "cma": "4.4.4",
+        "numpy": "2.4.4",
+        "open3d": "0.19.0",
+        "torch": "2.11.0+cu128",
+        "warp-lang": "1.16.0"
+      }
+    }
+  },
+  "evidence_files": [
+    "evidence/c6_dispatch/task_input.json",
+    "evidence/selected_tracking_trajectory/rollout.npz",
+    "evidence/terminal_progress/photo_production_progress.json",
+    "evidence/tracking_photometric_scan/candidates/c0086/inputs.json",
+    "evidence/tracking_photometric_scan/candidates/c0086/p0_scores.npz",
+    "evidence/tracking_photometric_scan/candidates/c0086/score.json",
+    "evidence/tracking_photometric_scan/contract.json",
+    "evidence/tracking_photometric_scan/selected_parent_p0/inputs.json",
+    "evidence/tracking_photometric_scan/selected_parent_p0/p0_scores.npz",
+    "evidence/tracking_photometric_scan/selected_parent_p0/score.json",
+    "evidence/tracking_photometric_scan/selection.json"
+  ],
+  "family": "real_rgbd_npz",
+  "gaps": [],
+  "identity": {
+    "session_id": "session_20260921_225908_522335",
+    "window_id": "anchor0328-0560"
+  },
+  "inputs": [],
+  "labels": {
+    "BEST": {
+      "key": "best_candidate",
+      "metric": "RT photometric active full-window mean",
+      "score": 0.006685305404187961,
+      "source": "native/adam/best.json",
+      "status": "final",
+      "value": "adam/c0280"
+    },
+    "FINAL": {
+      "key": "final_candidate",
+      "note": "final traversal snapshot at the planned horizon (no FINAL export written)",
+      "source": "native/adam/training_complete.json",
+      "value": "adam/c0500"
+    }
+  },
+  "license": {
+    "attribution": null,
+    "id": "mit",
+    "scope": "MIT (owner decision D9) applies to the project-produced fit packages in this repository: native checkpoints, run records and the generated metadata. It does not relicense the linked DAIRLab/phystwin_dataset recordings, which declare no license, or any third-party code or data."
+  },
+  "limitations": [
+    "continuing this fit re-hashes author weights TAPNext++ tapnextpp_512 checkpoint (Google DeepMind TAPNet) (/mnt/data2/magna/belt_perception/results/phystwin_real_rgbd/session_20260921_225630_786148/rt_waft_tapnextpp_refit_01/checkpoints/tapnextpp_512.ckpt); they are not published here (dependencies.json third_party_weights names their source and sha256)",
+    "historical depth: the run's records name depth_source waft_stereo_dav2l5 but no per-stream digest, so the consumed depth bytes and their publication revision are unresolved; correction ledgers replaced 4 depth paths of this session (c9fa16c0c8ac)",
+    "photometric_adam: records name 1341c18fc121 but none states whether the working tree was clean when this stage ran",
+    "Completion records establish completion of the configured work, not owner acceptance of model quality.",
+    "The accepted input has disclosed appearance limitations.",
+    "Photometric continuation remains unsupported. The cache bytes are preserved, but its original identity includes source video mtimes; relocated cached replay has not been qualified."
+  ],
+  "native_files": {
+    "bytes": 30660533224,
+    "files": 9020
+  },
+  "native_reconciliation": {
+    "included_by_category": {
+      "adam_continuation": {
+        "bytes": 26322951552,
+        "files": 1641
+      },
+      "adam_snapshot": {
+        "bytes": 223439252,
+        "files": 1506
+      },
+      "adam_update_rows": {
+        "bytes": 521877163,
+        "files": 1641
+      },
+      "cma_candidate": {
+        "bytes": 575076759,
+        "files": 3200
+      },
+      "cma_center_winner": {
+        "bytes": 5392995,
+        "files": 4
+      },
+      "cma_generation_state": {
+        "bytes": 5693030,
+        "files": 500
+      },
+      "frozen_score": {
+        "bytes": 2956343025,
+        "files": 502
+      },
+      "handoff": {
+        "bytes": 9458581,
+        "files": 3
+      },
+      "log": {
+        "bytes": 35996,
+        "files": 8
+      },
+      "run_record": {
+        "bytes": 28511168,
+        "files": 5
+      },
+      "stage_record": {
+        "bytes": 11753703,
+        "files": 10
+      }
+    },
+    "included_files": 9020,
+    "not_selected": {
+      "post-hoc frozen-score evaluation trajectory; its scores remain in evaluations/*/result.json and best.json": {
+        "bytes": 3060766629,
+        "examples": [
+          "adam/evaluations/c0000/rollout.npz",
+          "adam/evaluations/c0001/rollout.npz",
+          "adam/evaluations/c0002/rollout.npz"
+        ],
+        "files": 502
+      },
+      "runtime process or lock marker": {
+        "bytes": 0,
+        "examples": [
+          "supervisor.lock"
+        ],
+        "files": 1
+      },
+      "telemetry journal for W&B delivery; no telemetry process owns scientific state": {
+        "bytes": 686959535,
+        "examples": [
+          "telemetry-collector/real_collector.lock",
+          "telemetry-collector/real_collector_cursor.json",
+          "telemetry-collector/real_source_identity.json"
+        ],
+        "files": 8
+      }
+    },
+    "source_files": 9531,
+    "unclassified": [],
+    "unclassified_files": 0
+  },
+  "notes": [
+    "All retained native optimizer state and checkpoints are selected under the established model-package policy; no scientific execution was performed for publication.",
+    "The fit ran in two local segments: the first yielded at a committed boundary at update 12,417 and the second resumed it at one code revision."
+  ],
+  "parent": {
+    "declared": {
+      "checkpoint": "adam/c0086",
+      "relation": "Exact tracking snapshot selected by the saved photometric scan",
+      "run_id": "c6-225908-522335-tracking"
+    },
+    "native": {
+      "checkpoint": "adam/c0086",
+      "expected_sha256": {
+        "adam/model_init.npz": "91c23d44038477ae3541470d4006ef6557e3ee62b0da4e761fcec499f95a99f8",
+        "adam/snapshots/c0086/model/model.npz": "a6e6a4bd63dd4dc25716bc0248d713faa81a8b52a6deda48f549264ada119d72",
+        "adam/snapshots/c0086/model/topology.npz": "90ba248a4ab761ca0d1a6098491443347f184f3dd79f51cc523e00d5157b9d67"
+      },
+      "fit_run": "/mnt/data0/magna/belt_perception_runs/c6_lab_fitting_20261029/225908_522335/tracking",
+      "key": "identity.inputs.parent",
+      "kind": "parent_traversal_snapshot",
+      "selection_role": "scan_selected",
+      "source": "native/adam/manifest.json"
+    },
+    "native_run_id": "c6-225908-522335-tracking",
+    "verified": true,
+    "verified_files": [
+      "adam/model_init.npz",
+      "adam/snapshots/c0086/model/model.npz",
+      "adam/snapshots/c0086/model/topology.npz"
+    ]
+  },
+  "relocation": {
+    "path_in_repo": "sessions/session_20260921_225908_522335/runs/c6-225908-522335-photo/native",
+    "recorded_aliases": [],
+    "recorded_root": "/mnt/data0/magna/belt_perception_runs/c6_lab_fitting_20261029/225908_522335/photo_production",
+    "resolved_root": "/mnt/data0/magna/belt_perception_runs/c6_lab_fitting_20261029/225908_522335/photo_production",
+    "rule": "native/<path relative to the recorded root>; native records keep recorded paths"
+  },
+  "repo_id": "DAIRLab/phystwin_models",
+  "run_id": "c6-225908-522335-photo",
+  "schema": "phystwin_model_run_v1",
+  "stages": [
+    {
+      "active_coordinates": {
+        "key": "protocol.map.coordinates (each maps to a leaf or setting)",
+        "source": "native/cma/manifest.json",
+        "value": [
+          "ground_restitution",
+          "ground_friction",
+          "drag_damping",
+          "dashpot_damping"
+        ]
+      },
+      "budget": {
+        "completed": {
+          "center_evaluations": {
+            "key": "center_evaluations",
+            "source": "native/cma/result.json",
+            "value": null
+          },
+          "completed_generations": {
+            "key": "completed_generations",
+            "source": "native/cma/result.json",
+            "value": 100
+          },
+          "issued_asks": {
+            "key": "issued_asks",
+            "source": "native/cma/result.json",
+            "value": 800
+          },
+          "physical_failures": {
+            "key": "physical_failures",
+            "source": "native/cma/result.json",
+            "value": 0
+          },
+          "status": {
+            "key": "status",
+            "source": "native/cma/result.json",
+            "value": "winner"
+          },
+          "stop_reason": {
+            "key": "stop_reason",
+            "source": "native/cma/result.json",
+            "value": "max_generations"
+          },
+          "told_asks": {
+            "key": "told_asks",
+            "source": "native/cma/result.json",
+            "value": 800
+          }
+        },
+        "configured": {
+          "generations": {
+            "key": "protocol.max_generations",
+            "source": "native/cma/manifest.json",
+            "value": 100
+          }
+        }
+      },
+      "carried_fixed": {
+        "key": "protocol.map.carried_inactive",
+        "source": "native/cma/manifest.json",
+        "value": [
+          "object_restitution",
+          "object_friction",
+          "collision_distance_m"
+        ]
+      },
+      "inactive_coordinates": {
+        "key": "protocol.map.coordinates",
+        "source": "native/cma/manifest.json",
+        "value": []
+      },
+      "kind": "photometric_cma",
+      "manifest": "native/cma/manifest.json",
+      "objective": {
+        "key": "protocol.score",
+        "source": "native/cma/manifest.json",
+        "value": "mean of all four views per transition, then mean over K; RP1 active Q5 objective"
+      },
+      "schema": "real_rgbd_photometric_cma_v1",
+      "searched_coordinates": {
+        "key": "protocol.map.coordinates",
+        "source": "native/cma/manifest.json",
+        "value": [
+          "ground_restitution",
+          "ground_friction",
+          "drag_damping",
+          "dashpot_damping"
+        ]
+      },
+      "settings": {
+        "backend": {
+          "key": "strategy.backend_id",
+          "source": "native/cma/manifest.json",
+          "value": "pycma_4_4_4_serial_ask_tell_v1"
+        },
+        "dimension": {
+          "key": "strategy.dimension",
+          "source": "native/cma/manifest.json",
+          "value": 4
+        },
+        "initial_sigma": {
+          "key": "strategy.initial_sigma",
+          "source": "native/cma/manifest.json",
+          "value": 0.16666666666666666
+        },
+        "population_size": {
+          "key": "strategy.population_size",
+          "source": "native/cma/manifest.json",
+          "value": 8
+        },
+        "seed": {
+          "key": "strategy.seed",
+          "source": "native/cma/manifest.json",
+          "value": 42
+        },
+        "stopping": {
+          "key": "protocol.stopping",
+          "source": "native/cma/manifest.json",
+          "value": "native es.stop() at loop head after first generation; else generation cap"
+        }
+      },
+      "stage": "cma",
+      "supervision": {
+        "key": "supervision",
+        "source": "native/cma/manifest.json",
+        "value": null
+      }
+    },
+    {
+      "best_metric": {
+        "key": "protocol.best_metric",
+        "source": "native/adam/manifest.json",
+        "value": "RT photometric active full-window mean"
+      },
+      "budget": {
+        "completed": {
+          "traversals": {
+            "key": "complete_traversals",
+            "source": "native/adam/training_complete.json",
+            "value": 500
+          },
+          "updates": {
+            "key": "total_updates",
+            "source": "native/adam/training_complete.json",
+            "value": 116000
+          }
+        },
+        "configured": {
+          "traversals": {
+            "key": "protocol.complete_traversals",
+            "source": "native/adam/manifest.json",
+            "value": 500
+          },
+          "updates": {
+            "key": "protocol.complete_traversals * protocol.updates_per_traversal",
+            "source": "native/adam/manifest.json",
+            "value": 116000
+          },
+          "updates_per_traversal": {
+            "key": "protocol.updates_per_traversal",
+            "source": "native/adam/manifest.json",
+            "value": 232
+          }
+        }
+      },
+      "manifest": "native/adam/manifest.json",
+      "objective": {
+        "key": "identity.inputs.objective",
+        "source": "native/adam/manifest.json",
+        "value": {
+          "depth_rule": "q5_valid_indicator_numerator",
+          "depth_target": "stored uint16 / 1000.0 (float64), masked by 1[O > 0.5]",
+          "fused_ssim_binary_sha256": "4b65a2ca679f10420b1c6946e284edf69dceaa6c24c2f1a7290449203b5fc363",
+          "fused_ssim_module": "/mnt/data0/magna/belt_perception_wt_c6_lab_fitting_20261029/.venv/lib/python3.11/site-packages/fused_ssim/__init__.py",
+          "fused_ssim_wrapper_sha256": "c4a9caf98599b378050d1aa258a0b5d8b5817635bc5aefa29ab6893f442490bf",
+          "masking": "rendered x (1 - H); observed RGB x O; observed depth x 1[O > 0.5]",
+          "means": "full image, all channels",
+          "objective_id": "real_rgbd_active_photometric_v1",
+          "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, zero padding)",
+          "source": "U gs_train.py:128-188 + loss_utils.py (PhysTwin 81c718790a37e5e0102eb77af2c6edd34a9db25f)",
+          "view_aggregation": "arithmetic mean over all supervising views",
+          "weights": {
+            "lambda_depth": 0.001,
+            "lambda_dssim": 0.2,
+            "lambda_seg": 1.0
+          }
+        }
+      },
+      "optimizer": {
+        "betas": {
+          "key": "protocol.betas",
+          "source": "native/adam/manifest.json",
+          "value": [
+            0.9,
+            0.99
+          ]
+        },
+        "eps": {
+          "key": "protocol.eps",
+          "source": "native/adam/manifest.json",
+          "value": 1e-08
+        },
+        "lr": {
+          "key": "protocol.lr",
+          "source": "native/adam/manifest.json",
+          "value": 0.001
+        },
+        "optimizer": {
+          "key": "protocol.optimizer",
+          "source": "native/adam/manifest.json",
+          "value": null
+        },
+        "scheduler": {
+          "key": "protocol.scheduler",
+          "source": "native/adam/manifest.json",
+          "value": null
+        },
+        "weight_decay": {
+          "key": "protocol.weight_decay",
+          "source": "native/adam/manifest.json",
+          "value": null
+        }
+      },
+      "parameters": {
+        "leaf_decision": {
+          "key": "protocol.leaf_decision",
+          "source": "native/adam/manifest.json",
+          "value": "Q7 (A) 2026-09-25: the R7/source leaf set (every object and attachment log_spring_y, raw ground restitution and friction; inactive object-contact leaves carried with zero gradient asserted)"
+        },
+        "leaf_set_contract": {
+          "key": "contract.adam_leaf_set",
+          "source": "native/cma/manifest.json",
+          "value": {
+            "decision": "Q7 (A) 2026-09-25: the R7/source leaf set (every object and attachment log_spring_y, raw ground restitution and friction; inactive object-contact leaves carried with zero gradient asserted)",
+            "frozen": [
+              "drag_damping",
+              "dashpot_damping",
+              "collision_distance_m",
+              "topology",
+              "rests",
+              "masses",
+              "attachments",
+              "q0/v0",
+              "controls",
+              "clock"
+            ],
+            "initialization": "exact photometric CMA winner arrays after a passed zero-update handoff",
+            "logged_activity": [
+              "raw_ground_restitution",
+              "raw_ground_friction"
+            ],
+            "optimizer_leaves": [
+              "log_spring_y",
+              "raw_ground_restitution",
+              "raw_ground_friction",
+              "raw_object_restitution",
+              "raw_object_friction"
+            ],
+            "zero_gradient_asserted": [
+              "raw_object_restitution",
+              "raw_object_friction"
+            ]
+          }
+        },
+        "optimizer_leaves": {
+          "key": "protocol.leaves",
+          "source": "native/adam/manifest.json",
+          "value": [
+            "log_spring_y",
+            "raw_ground_restitution",
+            "raw_ground_friction",
+            "raw_object_restitution",
+            "raw_object_friction"
+          ]
+        }
+      },
+      "schema": "real_rgbd_photometric_adam_v1",
+      "seed": {
+        "reason": "the Adam stage records no random seed; its traversal order is chronological",
+        "searched": [
+          "native/adam/manifest.json"
+        ],
+        "value": null
+      },
+      "stage": "photometric_adam",
+      "stiffness_update": {
+        "key": "protocol.projected_log_y",
+        "source": "native/adam/manifest.json",
+        "value": "projected: finite raw log Y moved back onto the physical cap after every update (protocol.projected_log_y true)"
+      }
+    }
+  ],
+  "status": {
+    "completion_evidence": [
+      "native/adam/training_complete.json",
+      "native/adam/best.json",
+      "native/cma/result.json"
+    ],
+    "label": "complete",
+    "note": null
+  },
+  "unknowns": []
+}