diff --git a/fastvideo2/evidence/goldens/train-finetune-main/manifest.json b/fastvideo2/evidence/goldens/train-finetune-main/manifest.json
new file mode 100644
index 000000000..f63049f15
--- /dev/null
+++ b/fastvideo2/evidence/goldens/train-finetune-main/manifest.json
@@ -0,0 +1,41 @@
+{
+  "fastvideo_commit": "c459a1897899ffcec3be7765534d81000b9bb9c1",
+  "dataset": "wlsaidhi/crush-smol_processed_t2v",
+  "seed": 42,
+  "steps": 5,
+  "num_latent_t": 8,
+  "config": "1gpu bs1 accum1 lr5e-5 wd1e-4 betas(0.9,0.999) clip1.0 uniform-t cfg_rate0 dit_fp32 mixed_bf16 flow-match target=noise-latents flash_attn",
+  "torch": "2.12.0+cu130",
+  "gpu": "NVIDIA GB200",
+  "losses": [
+    0.1940414160490036,
+    0.94484943151474,
+    0.1001732274889946,
+    0.9069435000419617,
+    0.10525074601173401
+  ],
+  "self_noise_runs": [
+    [
+      0.1940414160490036,
+      0.9445295333862305,
+      0.10048552602529526,
+      0.9095064997673035,
+      0.10758557915687561
+    ],
+    [
+      0.1940414160490036,
+      0.9446913599967957,
+      0.10070198774337769,
+      0.9106038808822632,
+      0.10818696022033691
+    ],
+    [
+      0.1940414160490036,
+      0.94484943151474,
+      0.1001732274889946,
+      0.9069435000419617,
+      0.10525074601173401
+    ]
+  ],
+  "self_noise_max": 0.0036603808403015137
+}
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/params.npz b/fastvideo2/evidence/goldens/train-finetune-main/params.npz
new file mode 100644
index 000000000..f33a78aba
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/params.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/step0.npz b/fastvideo2/evidence/goldens/train-finetune-main/step0.npz
new file mode 100644
index 000000000..4f906037f
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/step0.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/step1.npz b/fastvideo2/evidence/goldens/train-finetune-main/step1.npz
new file mode 100644
index 000000000..e58644e04
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/step1.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/step2.npz b/fastvideo2/evidence/goldens/train-finetune-main/step2.npz
new file mode 100644
index 000000000..c49ab2fa6
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/step2.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/step3.npz b/fastvideo2/evidence/goldens/train-finetune-main/step3.npz
new file mode 100644
index 000000000..397b38fb9
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/step3.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/step4.npz b/fastvideo2/evidence/goldens/train-finetune-main/step4.npz
new file mode 100644
index 000000000..19be91bdd
Binary files /dev/null and b/fastvideo2/evidence/goldens/train-finetune-main/step4.npz differ
diff --git a/fastvideo2/evidence/goldens/train-finetune-main/steps.json b/fastvideo2/evidence/goldens/train-finetune-main/steps.json
new file mode 100644
index 000000000..81e48b26a
--- /dev/null
+++ b/fastvideo2/evidence/goldens/train-finetune-main/steps.json
@@ -0,0 +1,82 @@
+[
+ {
+  "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
+  "latents_hash": "93ae01bc47c0817b",
+  "embeds_hash": "53f384dfed477e28",
+  "noise_hash": "e1f135db4ff97ff8",
+  "noisy_hash": "6a0cabf3c2e98794",
+  "timesteps": [
+   118.0
+  ],
+  "sigmas": [
+   0.1181640625
+  ],
+  "pred_hash": "2903c217cc9279df",
+  "loss": 0.1940414160490036,
+  "grad_norm": 0.23514027893543243
+ },
+ {
+  "caption": "The video shows a colorful sponge being flattened as if it were under a hydraulic press, with the sponge being compressed and eventually flattened into a thin layer.",
+  "latents_hash": "b9d5abc838e815da",
+  "embeds_hash": "82c3f45fec89f678",
+  "noise_hash": "ea2e0427b4584180",
+  "noisy_hash": "5ff0bac18dd8e80e",
+  "timesteps": [
+   85.0
+  ],
+  "sigmas": [
+   0.0849609375
+  ],
+  "pred_hash": "688def082dcbfa60",
+  "loss": 0.94484943151474,
+  "grad_norm": 10.1902437210083
+ },
+ {
+  "caption": "The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is composed of a large, cylindrical metal cylinder with yellow and black stripes, and a metal base. The objects being flattened are two cylindrical blocks of cotton candy, one pink and one blue. The press is positioned on a metal table, and the background features a green wall with a yellow and red sign.",
+  "latents_hash": "840ddc5d78da4b7b",
+  "embeds_hash": "ffce0c1bf664fb3d",
+  "noise_hash": "2f512661cebc8a51",
+  "noisy_hash": "05b1746f31f8d83d",
+  "timesteps": [
+   618.0
+  ],
+  "sigmas": [
+   0.6171875
+  ],
+  "pred_hash": "963b9731aafbe2ea",
+  "loss": 0.1001732274889946,
+  "grad_norm": 0.523212730884552
+ },
+ {
+  "caption": "A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.",
+  "latents_hash": "c5d8005785df08ab",
+  "embeds_hash": "88eea058d58c26f7",
+  "noise_hash": "1022b2e68b615e5f",
+  "noisy_hash": "ee1750fb9fdb7094",
+  "timesteps": [
+   41.0
+  ],
+  "sigmas": [
+   0.041015625
+  ],
+  "pred_hash": "5fd1a77f3716b63b",
+  "loss": 0.9069435000419617,
+  "grad_norm": 3.6921098232269287
+ },
+ {
+  "caption": "The video shows a large, industrial press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing a pile of pink objects into a pile of crumbs. The press is large and metallic, with a yellow and black striped pattern on its side. The background is a green wall with a yellow warning sign.",
+  "latents_hash": "57045eb31ae1e81a",
+  "embeds_hash": "71cee36d5ca3083d",
+  "noise_hash": "cdc1cbc4a1fed9ac",
+  "noisy_hash": "53ada1e537af47ac",
+  "timesteps": [
+   610.0
+  ],
+  "sigmas": [
+   0.609375
+  ],
+  "pred_hash": "315edc5acc47767a",
+  "loss": 0.10525074601173401,
+  "grad_norm": 0.9913093447685242
+ }
+]
diff --git a/fastvideo2/evidence/ledger.jsonl b/fastvideo2/evidence/ledger.jsonl
index 07f08f196..1da926fe8 100644
--- a/fastvideo2/evidence/ledger.jsonl
+++ b/fastvideo2/evidence/ledger.jsonl
@@ -157,3 +157,4 @@
 {"ts": "2026-07-23T03:12:27Z", "gate": "anchor.fastwan-qad-main", "status": "pass", "model_id": "fastwan-qad-fp8-1.3b", "card_digest": "1c8e6f7d1380552d", "metrics": {"dit.bf16.t1000": 0.0, "dit.bf16.t757": 0.0, "dit.bf16.t522": 0.0, "dit.fp8.t1000": 0.0, "dit.fp8.t757": 0.0, "dit.fp8.t522": 0.0, "text.e2e": 0.0, "text.probe": 0.0, "e2e.final": 0.0, "e2e.step1.x": 0.0, "e2e.step2.x": 0.0}, "tolerances": {"all": 0.0}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "0b332ff", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897"}
 {"ts": "2026-07-23T03:16:22Z", "gate": "anchor.fastwan-vsa-main", "status": "pass", "model_id": "fastwan-t2v-1.3b", "card_digest": "0bd5c7771e10ce44", "metrics": {"dit.vsa.t1000": 0.0, "dit.vsa.t757": 0.0, "dit.vsa.t522": 0.0, "text.e2e": 0.0, "text.probe": 0.0, "e2e.final": 0.0, "e2e.step1.x": 0.0, "e2e.step2.x": 0.0}, "tolerances": {"all": 0.0}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "0b332ff", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897"}
 {"ts": "2026-07-23T11:20:04Z", "gate": "anchor.sfwan-main", "status": "pass", "model_id": "sfwan-t2v-1.3b", "card_digest": "9c9b4cc573606028", "metrics": {"text.e2e": 0.0, "e2e.final": 0.0, "fwd.hash_matches": 0.0}, "tolerances": {"all": 0.0}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "8b7285a", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897"}
+{"ts": "2026-07-23T11:47:02Z", "gate": "anchor.train-finetune-main", "status": "pass", "model_id": "wan2.1-t2v-1.3b", "card_digest": "7171d8729ae72ae4", "metrics": {"noisy.step0": 0.0, "pred.step0": 0.0, "loss.step0": 0.0, "gnorm.step0": 7.675588130950928e-05, "loss.step1": 5.221366882324219e-05, "gnorm.step1": 0.05505657196044922, "loss.step2": 0.0007446706295013428, "gnorm.step2": 0.006015121936798096, "loss.step3": 0.0025970935821533203, "gnorm.step3": 0.032111406326293945, "loss.step4": 0.0013115331530570984, "gnorm.step4": 0.0047234296798706055, "params.w0": 0.0, "params.w5": 0.00029868653199637143}, "tolerances": {"step0": 0.0, "later": "main self-noise band (see manifest)"}, "env": {"python": "3.12.13", "machine": "aarch64", "commit": "e14afd7", "torch": "2.12.0+cu130", "cuda": "13.0", "gpu": "NVIDIA GB200", "diffusers": "0.39.0", "transformers": "5.13.0", "attention": {"policy": "AUTO", "available": ["FLASH_ATTN", "SDPA"]}}, "detail": "goldens c459a1897"}
diff --git a/fastvideo2/evidence/train_numerics_report.md b/fastvideo2/evidence/train_numerics_report.md
new file mode 100644
index 000000000..924fde833
--- /dev/null
+++ b/fastvideo2/evidence/train_numerics_report.md
@@ -0,0 +1,37 @@
+# Training parity — wan2.1 finetune vs fastvideo-main (legacy stack)
+
+Gate `anchor.train-finetune-main`: PASS (2026-07-23, main @ c459a189).
+
+Config (goldens manifest): crush-smol processed parquet, 1 GPU, batch 1,
+grad_accum 1, num_latent_t 8, lr 5e-5, wd 1e-4, betas (0.9,0.999), clip 1.0,
+uniform timestep sampling, target = noise - latents, dit fp32 masters + bf16
+compute (FSDP2 world-1 mixed precision on main; explicit fp32-master/bf16-
+compute chain in `fastvideo2/train/finetune.py`), cfg_rate 0, seed 42.
+
+| row | result |
+|---|---|
+| noisy/pred/loss @ step0 | **bitwise 0.0** |
+| loss steps 1–4 | 5.2e-5 … 2.6e-3, inside the measured band |
+| params.w0 | bitwise |
+| gnorm / params.w5 | informational (grad-noise carriers) |
+
+Tolerance is MEASURED, not chosen: main-vs-main across three capture runs
+drifts up to 3.66e-3 in per-step loss (flash-attention backward atomics);
+the gate band is 1.5x that. Step-0 loss is exactly repeatable on main
+(0.0 across runs) and we match it bitwise — i.e., the forward, input
+construction, loss, and optimizer-chain math are exact; only the
+nondeterministic backward separates runs, ours indistinguishable from
+main's own reruns.
+
+Root cause found by the gate: main's `train_one_step` runs
+`_normalize_dit_input` (VAE scaling) BETWEEN batch fetch and noise prep —
+capturing latents at fetch time silently records the wrong tensors (3.03
+loss delta). Recorded post-normalization.
+
+Reuse invariant held: the trained module is the SAME `WanModelFV` the
+serving path loads (training-mode forward proved bitwise against main's
+FSDP2-wrapped port via pred.step0).
+
+Next gates (see fastvideo2/train/PORT_NOTES.md): dataloader order parity,
+VSA training, DMD2 (legacy CFG convention), self-forcing (rollout via
+WanCausalDMDLoop), attn-QAT/QAD, DiffusionNFT+VideoAlign.