Document native LoRA step338 archive verification

Add the 21-file LoRA manifest, fixed-revision remote and isolated-download receipts, CPU DCP model-state decode evidence, and explicit full-resume/deletion boundaries.

README.md

@@ -5,44 +5,57 @@ pretty_name: BEHAVIOR Explore Legacy Artifacts
 
 # BEHAVIOR Explore legacy archive
 
-This dataset repository is the audit archive for the superseded RPent / RLinf BEHAVIOR prompt-following route.
+This public dataset repository preserves audit evidence from retired RPent / RLinf BEHAVIOR prompt-following routes. Every artifact under `legacy/` is historical evidence only.
 
-## Current decision
+No legacy artifact—including `pi05-b1kpt50-cs32`, canonical steps 365/767/1167, or native-LoRA step338—may initialize, resume, warm-start, merge into, or select a checkpoint for the current LA/VLA route. Successful archival or restoration does not change this policy. The current route starts separately from the official Physical Intelligence `pi05_base` checkpoint.
 
-The old `pi05-b1kpt50-cs32` / canonical step1167 / LoRA step338 line is retired. None of these weights is an initialization point for the new training stages. The new route starts from the official Physical Intelligence `pi05_base` JAX/Orbax checkpoint, converts it for the BEHAVIOR 32-action-horizon PyTorch stack, recomputes normalization statistics on the 142 canonical train episodes, trains language-action without image tokens, then adds vision in a second canonical SFT stage.
+## Published content and status
 
-## Repository layout
+| Path | Publication status | Integrity status | Meaning |
+|---|---|---|---|
+| `legacy/models/pi05-b1kpt50-cs32/` | uploaded | covered by the original archive record | Retired third-party BEHAVIOR pt50 model. |
+| `legacy/training/canonical/` | uploaded | not exhaustively enumerated by the LoRA manifest | Retired canonical checkpoints, deploy exports, validation receipts, and logs. |
+| `legacy/training/lora/` | uploaded | 21/21 source files verified at the fixed payload revision; see `manifests/archive-manifest-v2.json` | Retired native-LoRA run, sealed step338 DCP, associated CPU-validation evidence, TensorBoard event, and driver log. |
+| `legacy/diagnostics/` | not uploaded | not verified | Planned destination for interrupted campaigns; directory presence must not be inferred from this README. |
+| `legacy/source-snapshot/` | not uploaded | not verified | Planned destination for selected source/config snapshots; directory presence must not be inferred from this README. |
 
-- `legacy/models/pi05-b1kpt50-cs32/`: third-party Comet BEHAVIOR pt50 PyTorch model formerly used by RPent. Historical reference only.
-- `legacy/training/canonical/`: canonical training checkpoints, deploy exports, validation receipts, and logs, including step1167. Historical reference only; the sim exit gate was not completed.
-- `legacy/training/lora/`: stopped native-LoRA run and its evaluated step338 checkpoint. Historical reference only.
-- `legacy/diagnostics/`: interrupted anchor/canonical campaigns, routing experiments, simulator authority artifacts, and production smokes. These are audit evidence, not a completed training dataset.
-- `legacy/source-snapshot/`: exact dirty-worktree source/config files needed to interpret receipts and hashes. Git commit IDs alone were insufficient because material files were untracked or modified.
-- `manifests/`: machine-readable local-to-Hub mapping, sizes, SHA-256 digests, and archival status.
+Only files explicitly enumerated in an active manifest are covered by that manifest's integrity claims. A path named in this README or a manifest is not, by itself, evidence that the path was uploaded.
 
-## Important data boundaries
+## Native-LoRA step338
 
-- The local canonical dataset at `task-0001-prompt-following-canonical-v1` contains 142 train episodes, 1,704 segments, and 663,708 canonical 32-step windows. It remains the input to the new LA and VLA stages and is intentionally not deleted or duplicated here.
-- The interrupted `rpent-anchor-272-v2` run ended with 70 completed formal cases and 16 stale-running status files after intentional shutdown; it has no final aggregate and must not be described as a 272-case completion set.
-- The older `rpent-anchor-272-v1` and `rpent-routeb-272-v1` trees are failure/diagnostic evidence only.
-- `exact_historical_restore=false` for Parquet anchor materialization receipts. One explicit warm-up physics tick was part of the accepted state-construction contract.
+The complete retained local LoRA source tree was uploaded without rewriting its bytes to `legacy/training/lora/` in payload revision `d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141`:
 
-## Model lineage
+- 21 regular files;
+- 16,978,218,437 bytes;
+- sealed DCP object `step-000000000338-1f2c404a246ebabc`;
+- DCP object manifest SHA-256 `1f2c404a246ebabc77215999c898a7a848cf3f0656834e4962adcae08864c5dd`;
+- associated 32-forward CPU validation receipt with `masked_val_flow_loss=0.007970079779624939`.
 
-- `pi05-b1kpt50-cs32` is not the official `pi05_base`; it came from the third-party `sunshk/openpi_comet` BEHAVIOR pt50 release.
-- canonical step1167 passed its CPU validation but did not complete the formal simulator exit gate.
-- LoRA step338 is the only evaluated recoverable LoRA checkpoint; later in-memory progress was never sealed.
-- The future official base is sourced separately from `gs://openpi-assets/checkpoints/pi05_base` and is not part of this legacy archive.
+The associated CPU receipt originally validated a temporary BF16 model export. It must not be described as proof that the archived DCP training state was restored. The manifest separately records source integrity, remote per-file verification, isolated re-download verification, model-state DCP decoding, and production-equivalent resume status.
 
-## Restoration and integrity
+The archived `.pt`, `.distcp`, and DCP metadata files are trusted checkpoint inputs that may use pickle-based serialization. Load them only in a trusted, isolated environment after verifying the manifest. Original logs and receipts intentionally preserve machine-local provenance such as absolute paths, hostnames, PIDs, and timestamps. A bounded scan of the text, config, JSON, and log files found no token, API-key, private-key, password, cookie, or authorization-header pattern; this is not a general claim about arbitrary binary checkpoint contents.
 
-Use the files under `manifests/` to verify local or downloaded bytes. Large Hub objects are checked against their LFS/Xet SHA-256 metadata before the corresponding local copies are deleted. Small files are re-downloaded or byte-checked. Absolute symlinks from the old model tree are represented as ordinary copied files plus documented link mappings.
+## Data and campaign boundaries
 
-## Guidance for future AI agents
+- The canonical task-0001 dataset contains 142 train episodes, 1,704 segments, and 663,708 canonical 32-step windows. It remains local training input and is not duplicated here.
+- The stopped `rpent-anchor-272-v2` snapshot recorded 70 cases with completed status, 16 stale-running status files, and 186 unstarted cases. “Completed status” is not task success. No final aggregate exists, so no formal completion or success rate may be derived from it.
+- `exact_historical_restore=false` for Parquet anchor materialization receipts; one explicit warm-up physics tick was part of the accepted state-construction contract.
 
-1. Do not resume the old 272 campaign or use its partial completion sources as a formal aggregate.
-2. Do not initialize either new training stage from b1kpt50, step1167, step767, or LoRA step338.
-3. Preserve the 142-episode canonical dataset locally; it is the intended training data.
-4. For phase one, image token blocks must be omitted before SigLIP, not replaced with black images and not merely masked after image encoding.
-5. Recompute BEHAVIOR normalization statistics by exact enumeration of canonical train windows and `action_valid_mask`; the prior generic statistics script is not an exact substitute.
-6. Treat official task success, local subtask predicates, workflow sealing, and publication as separate facts.
+## Verification vocabulary
+
+These states are independent:
+
+- `uploaded`: files are present at a fixed Hub revision.
+- `remote_per_file`: remote path, size, and SHA-256 match the source manifest.
+- `isolated_redownload`: a new fixed-revision download matches the source manifest byte-for-byte.
+- `dcp_model_state_decode`: the downloaded DCP was parsed and its model state was materialized on CPU.
+- `production_equivalent_resume`: model, optimizer, scheduler, per-rank data state, and RNG were restored in an isolated matching distributed environment and advanced through a controlled resume step.
+
+Do not promote a weaker state to a stronger one. In particular, CPU loss, local predicates, materialization success, or checkpoint presence is not robot task success.
+
+## Guidance for future agents
+
+1. Use `manifests/archive-manifest-v2.json` for the LoRA file list and verification state; retain v1 only as historical evidence.
+2. Do not delete a local source merely because files are visible on the Hub. Require the intended deletion scope, fixed-revision remote verification, isolated re-download, the stated restoration evidence, and explicit deletion authorization.
+3. Do not mix interrupted v1/v2 campaign receipts or describe partial campaigns as complete datasets.
+4. Keep official task success, local predicates, workflow sealing, terminal receipts, and publication as separate facts.

manifests/archive-manifest-v2.json

@@ -0,0 +1,207 @@
+{
+  "schema_version": "scilwb.behavior_explore.legacy_archive.v2",
+  "generated_at": "2026-08-13T12:51:27+08:00",
+  "repository": {
+    "repo_id": "scilwb/behavior-explore",
+    "repo_type": "dataset",
+    "payload_revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+    "hub_root": "legacy/training/lora"
+  },
+  "scope": {
+    "source_root": "/home/ubuntu/lwb/Artifacts/RLinf_prompt_following/lora",
+    "file_count": 21,
+    "total_size_bytes": 16978218437,
+    "empty_directories_not_represented": [
+      "checkpoints/native_lora_v1_step338_stopped/.staging",
+      "checkpoints/native_lora_v1_step338_stopped/best_exports",
+      "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/model_state_dict"
+    ]
+  },
+  "checkpoint": {
+    "run_id": "native-lora-v1",
+    "global_step": 338,
+    "object": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc",
+    "object_manifest_sha256": "1f2c404a246ebabc77215999c898a7a848cf3f0656834e4962adcae08864c5dd",
+    "sealed": true
+  },
+  "verification": {
+    "local_source": {
+      "status": "passed",
+      "file_count": 21,
+      "total_size_bytes": 16978218437
+    },
+    "remote_per_file": {
+      "status": "passed",
+      "revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+      "matched_files": 21,
+      "receipt": "manifests/verification/native-lora-step338-remote-verification.json"
+    },
+    "isolated_redownload": {
+      "status": "passed",
+      "revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+      "matched_files": 21,
+      "receipt": "manifests/verification/native-lora-step338-isolated-redownload.json"
+    },
+    "dcp_model_state_decode": {
+      "status": "passed",
+      "model_tensor_count": 1137,
+      "native_lora_tensor_count": 468,
+      "materialized_bf16_sha256": "d9c4f27f12a7ed0948e96ac4e9a969e6af02f236e72da96366fe4a3d1e68bfbc",
+      "matches_historical_evaluated_bf16": true,
+      "receipt": "manifests/verification/native-lora-step338-dcp-model-state-decode.json"
+    },
+    "associated_cpu_validation": {
+      "status": "passed_historical_export",
+      "forward_count": 32,
+      "anchor_count": 32,
+      "masked_val_flow_loss": 0.007970079779624939,
+      "receipt_sha256": "dabb937aced2bcad08c97bd0e209c84990f504a12d8563b32aeae515a74fa7fd",
+      "scope_note": "The historical validator evaluated a temporary BF16 export. The isolated DCP decode reproduced that BF16 file SHA exactly; the 32 forwards were not rerun from the downloaded copy."
+    },
+    "production_equivalent_resume": {
+      "status": "not_run",
+      "reason": "A matching isolated eight-rank environment was unavailable without interrupting the active eight-GPU production run.",
+      "required_evidence": [
+        "restore model and optimizer with the production low-peak path",
+        "restore scheduler at step 338",
+        "restore all eight rank-local data and RNG states",
+        "record eight successful resume receipts",
+        "advance exactly one controlled step from 338 to 339 and verify state progression"
+      ]
+    }
+  },
+  "publication_policy": {
+    "historical_only": true,
+    "allowed_for_current_route_initialization": false,
+    "local_deletion_allowed": false,
+    "deletion_note": "Keep the local source until production-equivalent resume evidence exists and a separate explicit deletion authorization names the exact deletion scope."
+  },
+  "files": [
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/LATEST.json",
+      "role": "checkpoint_pointer",
+      "size_bytes": 317,
+      "sha256": "5999ce75b035db2c42abe1c2a59712f39470f400822273860333c2c2632afa60"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/COMPLETE",
+      "role": "checkpoint_seal",
+      "size_bytes": 106,
+      "sha256": "bb3c687821d452378255fc4076f4c6c6697ed0d3f9c9349ad71b4832ae9ab932"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/MANIFEST.json",
+      "role": "checkpoint_manifest",
+      "size_bytes": 1683,
+      "sha256": "1f2c404a246ebabc77215999c898a7a848cf3f0656834e4962adcae08864c5dd"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/data.pt",
+      "role": "rank_local_data_state",
+      "size_bytes": 954842,
+      "sha256": "117920409e5624df2ba0f4ccd694ddb9e169f0dd09df80b620b2d00fdbe10cd1"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/.metadata",
+      "role": "dcp_metadata",
+      "size_bytes": 4782817,
+      "sha256": "dba8f4dcbf4026c529e43ca3d36c7430e707da58903f8b11442ba8b530a718b7"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__0_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 4613830396,
+      "sha256": "401317719a8fee04fe0504dae9c02faa6bc09329d75acb5b5ffc610a790a250c"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__1_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 4612771428,
+      "sha256": "d7b496057c3711c7bb4b206d9421748da56a0f3e1dd666b0c0892da70533efde"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__2_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290794304,
+      "sha256": "d2fdc98b2129a8c426e09592fac6af15ae2e1aa84bcc69cdc3edf211c3e122c5"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__3_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290808792,
+      "sha256": "9d69172b6c632a696f9aa367ecddab8f19ae70e2ab25b28d45254af8f08503fd"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__4_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290818648,
+      "sha256": "31f0d259bc37ce8fd95d4699a6a14912a052ba451e62ec158627a28203fb7e12"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__5_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290824600,
+      "sha256": "86063f4c4fcab8a4de6f8f547db996dd011409c6d17977ffced5ee09a47a0c46"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__6_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290823420,
+      "sha256": "5eaa63a703cdafc4d434b74380a398cc46361a586a694730191dd7acb8752231"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/dcp_checkpoint/__7_0.distcp",
+      "role": "dcp_shard",
+      "size_bytes": 1290799964,
+      "sha256": "7c451fbc213eb179f72630b75dc3d068d8ebf1225691bb26d8b69b24e2289461"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/actor/rng.pt",
+      "role": "rank_local_rng_state",
+      "size_bytes": 106134,
+      "sha256": "59cfcdf2a51571121d63d5aeeea4b5138122af932f0cc5b86a666dfca611eba3"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc/runner_state.json",
+      "role": "runner_state",
+      "size_bytes": 161,
+      "sha256": "888a6d21c8ca9a3ecc6c8c4a2b9a90ae569206d9ee474fe76d074a7d0171eb64"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/validations/step-000000000338-1f2c404a246ebabc/process.json",
+      "role": "validation_process",
+      "size_bytes": 9665,
+      "sha256": "ff62d63515c1bc471c2637c698bdf9a7ffb1bc004984cc689615e9d3cb5d5fb3"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/validations/step-000000000338-1f2c404a246ebabc/receipt.json",
+      "role": "validation_receipt",
+      "size_bytes": 5204,
+      "sha256": "dabb937aced2bcad08c97bd0e209c84990f504a12d8563b32aeae515a74fa7fd"
+    },
+    {
+      "path": "checkpoints/native_lora_v1_step338_stopped/validations/step-000000000338-1f2c404a246ebabc/result.json",
+      "role": "validation_result",
+      "size_bytes": 5047,
+      "sha256": "9dce221cc7e70bedf5f00ab9708c1484e8084e6d665b9cd0ad3a3dd350bc3c71"
+    },
+    {
+      "path": "logs/tensorboard/config.yaml",
+      "role": "resolved_training_config",
+      "size_bytes": 10613,
+      "sha256": "dcb1e871b7886f7ba7c8a20e16bb3b0074c5c0a0bf86f5054203ff42885a4a06"
+    },
+    {
+      "path": "logs/tensorboard/events.out.tfevents.1786493576.iv-ye4qlmxtz4wh2yp4vsf1.1082442.0",
+      "role": "tensorboard_event",
+      "size_bytes": 285378,
+      "sha256": "a10ba065764fdbdb4963527fb7859829ecf0b54f6f4b6b83c7137ed071092e50"
+    },
+    {
+      "path": "runtime_native_lora_v1/driver.log",
+      "role": "training_driver_log",
+      "size_bytes": 584918,
+      "sha256": "c959aba2a9cdb7202cb454a1199fb12dcdbe35385355a9d6797624d741b2a433"
+    }
+  ]
+}

manifests/verification/native-lora-step338-dcp-model-state-decode.json

@@ -0,0 +1,69 @@
+{
+  "schema_version": "scilwb.behavior_explore.dcp_model_state_decode.v1",
+  "generated_at": "2026-08-13T12:51:27+08:00",
+  "status": "passed",
+  "input": {
+    "repo_id": "scilwb/behavior-explore",
+    "revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+    "download_root": "/dev/shm/rpent-lora-step338-redownload.BJFXKk/snapshot/legacy/training/lora",
+    "dcp_object": "checkpoints/native_lora_v1_step338_stopped/objects/step-000000000338-1f2c404a246ebabc",
+    "object_manifest_sha256": "1f2c404a246ebabc77215999c898a7a848cf3f0656834e4962adcae08864c5dd"
+  },
+  "implementation": {
+    "source": "/home/ubuntu/lwb/Projects/RLinf_prompt_following/toolkits/standalone_eval_scripts/openpi/behavior_prompt_following_checkpoint_export.py",
+    "source_sha256": "2672a734516a8c9a57b310e6c4b1b629dd6203d172409ee8ee87beb80f36802c",
+    "operations_reused": [
+      "verify DCP manifest payload size and SHA-256",
+      "CPU no_dist DCP model-state load",
+      "deterministic ordered BF16 materialization and mmap readback"
+    ],
+    "python": "3.10.12",
+    "torch": "2.5.1+cu124"
+  },
+  "model_state": {
+    "global_step": 338,
+    "run_id": "native-lora-v1",
+    "tensor_count": 1137,
+    "floating_tensor_count": 1137,
+    "all_floating_tensors_bfloat16": true,
+    "native_lora_enabled": true,
+    "native_lora_tensor_count": 468,
+    "vision_lora_tensor_count": 324,
+    "language_lora_tensor_count": 144,
+    "completion_head_is_default": true,
+    "materialized_size_bytes": 6719691106,
+    "materialized_sha256": "d9c4f27f12a7ed0948e96ac4e9a969e6af02f236e72da96366fe4a3d1e68bfbc",
+    "historical_evaluated_bf16_sha256": "d9c4f27f12a7ed0948e96ac4e9a969e6af02f236e72da96366fe4a3d1e68bfbc",
+    "matches_historical_evaluated_bf16": true
+  },
+  "training_state_structure": {
+    "dcp_metadata_key_count": 3181,
+    "contains_model": true,
+    "contains_optimizer": true,
+    "contains_scheduler": true,
+    "rank_local_data_state_count": 8,
+    "rank_local_rng_state_count": 8,
+    "runner_state_global_step": 338,
+    "runner_state_checkpoint_interval_seconds": 7200,
+    "runner_state_next_checkpoint_active_seconds": 14400
+  },
+  "associated_cpu_validation": {
+    "status": "passed_historical_export",
+    "forward_count": 32,
+    "anchor_count": 32,
+    "masked_val_flow_loss": 0.007970079779624939,
+    "receipt_sha256": "dabb937aced2bcad08c97bd0e209c84990f504a12d8563b32aeae515a74fa7fd",
+    "scope_note": "The prior validator evaluated the byte-identical BF16 model export. It did not restore optimizer, scheduler, data, or RNG state from DCP."
+  },
+  "production_equivalent_resume": {
+    "status": "not_run",
+    "reason": "No isolated matching eight-rank GPU environment was available without interrupting the active production training job.",
+    "not_proven": [
+      "optimizer application restore",
+      "scheduler application restore",
+      "per-rank data and RNG application restore",
+      "eight-rank resume receipts",
+      "controlled step 338 to 339 progression"
+    ]
+  }
+}

manifests/verification/native-lora-step338-isolated-redownload.json

@@ -0,0 +1,28 @@
+{
+  "schema_version": "scilwb.behavior_explore.isolated_redownload.v1",
+  "generated_at": "2026-08-13T12:51:27+08:00",
+  "status": "passed",
+  "repository": {
+    "repo_id": "scilwb/behavior-explore",
+    "repo_type": "dataset",
+    "revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+    "hub_root": "legacy/training/lora"
+  },
+  "download": {
+    "isolated_root": "/dev/shm/rpent-lora-step338-redownload.BJFXKk/snapshot",
+    "transport": "fixed-revision HTTP/LFS download with Xet disabled after the first isolated Xet attempt stalled",
+    "file_count": 21,
+    "total_size_bytes": 16978218437
+  },
+  "comparison": {
+    "source_file_count": 21,
+    "download_file_count": 21,
+    "matched_path_count": 21,
+    "matched_size_count": 21,
+    "matched_sha256_count": 21,
+    "missing_paths": [],
+    "unexpected_paths": []
+  },
+  "file_inventory": "manifests/archive-manifest-v2.json#files",
+  "note": "The isolated temporary download may be removed after this receipt and the manifest are published; the fixed revision and per-file digests are the durable reproduction contract."
+}

manifests/verification/native-lora-step338-remote-verification.json

@@ -0,0 +1,31 @@
+{
+  "schema_version": "scilwb.behavior_explore.remote_verification.v1",
+  "generated_at": "2026-08-13T12:51:27+08:00",
+  "status": "passed",
+  "repository": {
+    "repo_id": "scilwb/behavior-explore",
+    "repo_type": "dataset",
+    "revision": "d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141",
+    "hub_root": "legacy/training/lora"
+  },
+  "source": {
+    "root": "/home/ubuntu/lwb/Artifacts/RLinf_prompt_following/lora",
+    "file_count": 21,
+    "total_size_bytes": 16978218437
+  },
+  "remote": {
+    "file_count": 21,
+    "total_size_bytes": 16978218437,
+    "matched_path_count": 21,
+    "matched_size_count": 21,
+    "matched_sha256_count": 21,
+    "missing_paths": [],
+    "unexpected_paths": []
+  },
+  "method": {
+    "path_set": "fixed-revision repository tree",
+    "size": "fixed-revision file metadata",
+    "sha256": "LFS object sha256 for LFS files; fixed-revision content download and sha256 for non-LFS files"
+  },
+  "file_inventory": "manifests/archive-manifest-v2.json#files"
+}