Remove auxiliary records

1.3B/docs/validation.md

@@ -22,7 +22,6 @@ After removing the launcher preflight gates, core reference checks, and selected
 - Generator weights were finite, and the model, optimizer, and four per-worker random states were saved successfully. All 825 saved Adam parameter states recorded one update, and their gradients were not all zero.
 - The new checkpoint was loaded by the shared inference implementation and generated a decodable **81-frame, 832 × 480, 16 FPS** video.
 
-Machine-readable results are saved in [three-encoder-smoke.json](../results/validation/three-encoder-smoke.json).
 
 The run used `bash scripts/train.sh --smoke`. This overrides only the training length to one update; the distribution batch, differentiation batch, encoders, and reference assets remain those of the released recipe.
 

1.3B/results/validation/three-encoder-smoke.json

@@ -1,36 +0,0 @@
-{
-  "training_updates": 1,
-  "distribution_batch": 256,
-  "differentiation_batch": 128,
-  "reference_batch": 128,
-  "encoders": [
-    "DINOv3",
-    "V-JEPA2",
-    "VideoMAEv2"
-  ],
-  "checkpoint_finite": true,
-  "optimizer_saved": true,
-  "rng_states": 4,
-  "video_frames": 81,
-  "video_shape": [
-    480,
-    832,
-    3
-  ],
-  "fps": 16.0,
-  "cpu_tests_passed": 31,
-  "optimizer_parameter_states": 825,
-  "adam_steps_all_one": true,
-  "first_frame_pixel_std": 58.5123941148281,
-  "hardware": "4 x NVIDIA H200",
-  "torch": "2.5.1+cu121",
-  "flash_attention": "2.7.4.post1",
-  "logged_generator_loss": 3.995,
-  "logged_generator_grad_norm_before_clipping": 44.62,
-  "feature_replay_max_abs": {
-    "dinov3": 0.0,
-    "vjepa2": 0.0,
-    "videomae_v2": 0.0
-  },
-  "scope": "One full-batch training update and checkpoint inference; not a full 150-update or VBench rerun"
-}

14B/docs/validation.md

@@ -4,7 +4,7 @@
 
 The restored recipe is the **three-encoder, 8,000-reference-video run**. The original checkpoint header records step **80**, **1,095 generator parameter tensors**, and all three branches: DINOv3, V-JEPA 2, and VideoMAEv2. Its bank identity matches the released training config. The historical human-evaluation generation records and the VLM source-video run point to this same checkpoint.
 
-The original step-80 log reports loss **3.854**, generator gradient norm **10.98**, three encoder VJPs, batch geometry **256 / 128 / 128**, zero reported feature replay error for all three encoders, and a completed optimizer update. These are historical measurements, not a new run of this package. See [`original-step80.json`](../results/original-step80.json).
+The original step-80 log reports loss **3.854**, generator gradient norm **10.98**, three encoder VJPs, batch geometry **256 / 128 / 128**, zero reported feature replay error for all three encoders, and a completed optimizer update. These are historical measurements, not a new run of this package.
 
 ## Checks of the corrected release
 
@@ -16,7 +16,6 @@ The original step-80 log reports loss **3.854**, generator gradient norm **10.98
 - The single-encoder MMD kernel, training rollout, and encoder VJP implementation match the original three-encoder snapshot's executable AST after removing documentation and restoring readability aliases.
 - Python parsing, shell syntax, and diff whitespace checks passed.
 
-Path-free evidence is in [`release-validation.json`](../results/release-validation.json). Original source-fingerprint manifests are omitted from this anonymous release.
 
 **Full 14B GPU training, GPU video generation, and VBench were not rerun for this correction.** CPU sampler comparisons use a small test generator, not the 14B network, and do not establish identical generated videos or a new benchmark score.
 

14B/results/original-step80.json

@@ -1,24 +0,0 @@
-{
-  "scope": "Historical three-encoder training log; not a rerun of this package. Values retain the precision printed by the original logger.",
-  "checkpoint_step": 80,
-  "active_encoders": [
-    "dinov3",
-    "vjepa2",
-    "videomae_v2"
-  ],
-  "reference_count": 8000,
-  "use_ema": false,
-  "metrics": {
-    "generator_loss": 3.854,
-    "generator_grad_norm": 10.98,
-    "learning_rate": 1.818e-06,
-    "fresh_video_optimizer_step_skipped": 0.0,
-    "triple_dense_global_video_samples": 256.0,
-    "triple_dense_gradient_global_video_samples": 128.0,
-    "triple_dense_current_teacher_global_samples": 128.0,
-    "triple_dense_encoder_vjp_count": 3.0,
-    "dinov3_dense_replay_max_abs": 0.0,
-    "vjepa2_dense_replay_max_abs": 0.0,
-    "videomae_v2_dense_replay_max_abs": 0.0
-  }
-}

14B/results/release-validation.json

@@ -1,24 +0,0 @@
-{
-  "checkpoint_step": 80,
-  "generator_parameter_tensors": 1095,
-  "active_encoders": [
-    "dinov3",
-    "vjepa2",
-    "videomae_v2"
-  ],
-  "reference_count": 8000,
-  "teacher_bank_content_hash": "6b8f249733c39d865a8a9d4b1963779d5578c6a6d68aa5d96002a4b706c766e9",
-  "human_motion_and_camera_protocols_same_checkpoint": true,
-  "vlm_source_vbench_same_checkpoint": true,
-  "cpu_sampler_matches_evaluated_source": {
-    "continuous_two_prompts": true,
-    "paired_two_prompts": true,
-    "scope": "Small recording generator and real scheduler; no 14B model weights were executed."
-  },
-  "actual_dinov3_metadata_relocation_exact": true,
-  "cpu_tests": {
-    "passed": 51,
-    "skipped_cuda": 1
-  },
-  "five_sample_noise_order_cpu_test": true
-}

ANONYMIZATION.txt

@@ -1,35 +0,0 @@
-Anonymous source release
-
-Removed author names, affiliations, contact details, personal repository and
-project links, tracking badges, the identifying paper PDF, shared-chat links,
-cluster-specific paths, private collaboration acknowledgments, original source
-fingerprint manifests, and Git metadata/history. Project-specific module names,
-class names and environment variables now use neutral names.
-
-Hosted model repository URLs and hard-coded repository IDs were removed.
-Model and tokenizer loaders require existing local directories. The legacy-named
-VideoMAEv2 preparation utility stages an existing local snapshot with --source-dir;
-it does not download models. Existing valid local manifests can be preserved for
-frozen-bank compatibility. New manifests require matching reference assets and
-configuration identities. Local preparation reads bank identities into private
-generated configuration files; do not include those files or model assets when
-sharing this source package.
-
-Third-party source attribution and license notices are retained. Scientific
-training settings and reported results are retained. This package contains source
-code, configurations, example prompts and documentation, without model weights
-or training data.
-
-Verification passed:
-- Python syntax: 180 files; shell syntax: 6 scripts.
-- Both training recipes passed the existing source/configuration checks.
-- Both local snapshot tools passed staging, occupied-output rejection, missing
-  source rejection, nested-directory rejection and manifest-preservation checks.
-- Model adapters passed local-only loading and missing-path checks with mocks.
-- Both preparation scripts passed path relocation and bank-identity checks.
-- Final content and archive checks found no targeted author/project identifiers,
-  personal links, hosted model repository addresses, Git metadata or paper PDF.
-
-No GPU training or model inference was run for this source anonymization task.
-Previously public source code and retained scientific results can still be matched
-to public material; removing identifying metadata cannot guarantee anonymity.