radixark · Rockdu · Aug 17, 2026 · Aug 17, 2026 · Aug 17, 2026
diff --git a/docs/index.md b/docs/index.md
@@ -37,7 +37,8 @@ reproducible.
 
 ## Supported models
 
-Each model name links to its recipe page.
+Each model name links to its recipe page. Every documented recipe is labeled with a
+[recipe verification level](/user-guide/recipe-verification).
 
 
 | Model                                                   | Task      | Canonical recipes                         |
@@ -54,8 +55,8 @@ Each model name links to its recipe page.
 
 ## Feature support matrix
 
-- ✅ **Verified** — exercised by a canonical recipe in `scripts/` or a CI test.
-- 🟡 **Supported** — the code path exists, but no shipped recipe or test covers this combination yet.
+- ✅ **Recipe-backed** — exercised by a canonical recipe in `scripts/` or a CI test.
+- 🟡 **Implemented** — the code path exists, but no shipped recipe or test covers this combination yet.
 - ❌ **Not supported** — no working code path for this combination today.
 
 
@@ -67,7 +68,7 @@ Each model name links to its recipe page.
 | LoRA + IPC weight sync                   | ✅     | ✅          | ✅      | 🟡    | ✅       |
 | Single-prompt multi-gen (microgroup > 1) | ✅     | ✅          | ✅      | ❌       | ❌      |
 | USP sequence parallelism                 | ❌     | ❌          | ✅    | ❌       | ❌       |
-| Deterministic mode                       | ✅     | ✅          | ❌      | ✅       | ❌       |
+| Deterministic mode                       | ✅     | ✅          | ✅      | ✅       | ❌       |
 
 
 

diff --git a/docs/models/cosmos/cosmos3.md b/docs/models/cosmos/cosmos3.md
@@ -55,28 +55,33 @@ From `miles/backends/fsdp_utils/configs/cosmos3.py`:
 
 ## 4. Launch
 
-Canonical recipe: `scripts/run_diffusion_grpo_cosmos3_pickscore_t2i_5gpu.py` — 4 colocate + 1 reward GPU, T2I (832×480,
-1 frame), PickScore reward.
+Canonical recipe: `scripts/run_diffusion_grpo_cosmos3_pickscore_t2i_4gpu.py` — train, rollout,
+and PickScore colocated on 4 GPUs; T2I (832×480, 1 frame).
+
+**Status:** [📈 V — Verified](/user-guide/recipe-verification#v)
 
 ```bash
 export SGLANG_DISABLE_COSMOS3_GUARDRAILS=1   # RL scores raw samples; skip serving-side guardrail models
-python3 scripts/run_diffusion_grpo_cosmos3_pickscore_t2i_5gpu.py
+python3 scripts/run_diffusion_grpo_cosmos3_pickscore_t2i_4gpu.py
 ```
 
 
 
-## 5. Recipe Note
-`epoch_global_window` draws a 2-step window per rollout from `--diffusion-sde-candidate-steps 4-15`.
+## 5. Recipe notes
+
+`epoch_global_random_choice` draws two steps per epoch from
+`--diffusion-sde-candidate-steps 8,9,10,11`.
 
-The Cosmos3 checkpoint's Karras flow-sigma grid puts head steps 1–3 at `sigma > 0.96` with `|dt| < 0.02` — they
-basically train nothing. Step numbers are **not transferable across sigma-grid families**: re-derive candidates from
-`|dt|` when changing model or grid.
+The Cosmos3 checkpoint's Karras flow-sigma grid puts head steps 1–7 at
+`sigma > 0.96` with `|dt| < 0.02`; steps 8–11 are the useful high-noise segment.
+Step numbers are **not transferable across sigma-grid families**: re-derive candidates
+from `|dt|` when changing model or grid.
 
 
 ## 6. Reference results
 
-Flow-GRPO + PickScore on Cosmos3-Nano, run with the native Cosmos3 recipe scripts in this repo. PickScore
-(`rollout/reward/raw_mean`) climbs from ~0.77 to ~0.85 over 250 rollout steps:
+The 4-GPU colocated Cosmos3-Nano recipe raises PickScore
+(`rollout/reward/raw_mean`) from ~0.77 to ~0.85 over 250 rollouts:
 
 ![Cosmos3 PickScore reward mean](../../assets/images/cosmos3/reward_mean.png)
 

diff --git a/docs/models/h3/h3.md b/docs/models/h3/h3.md
@@ -60,6 +60,8 @@ The 2-GPU recipe lives on [PR #154](https://github.com/radixark/miles_diffusion/
 (sglang pins `short_edge=768`, so the canvas is 1344×768 / 107 frames). Requires `--use-lora --lora-ipc-weight-sync`
 (the recipe already sets both).
 
+**Status:** [📈 V — Verified](/user-guide/recipe-verification#v)
+
 ```bash
 # alignment check (optimizer frozen)
 python3 scripts/run_diffusion_grpo_h3_t2va_2gpu.py \

diff --git a/docs/models/ltx/ltx2.md b/docs/models/ltx/ltx2.md
@@ -48,6 +48,8 @@ Registered in `miles/backends/fsdp_utils/configs/ltx.py`:
 Canonical recipe: `scripts/run_diffusion_grpo_ltx23_sglang.py` — 4 colocate GPUs + 1 PickScore
 GPU, 57 frames @ 24 fps, 512×768, PickScore reward.
 
+**Status:** [🛡️ FG — Fully gated](/user-guide/recipe-verification#fg)
+
 ```bash
 python3 scripts/run_diffusion_grpo_ltx23_sglang.py
 ```

diff --git a/docs/models/qwen-image/qwen-image.md b/docs/models/qwen-image/qwen-image.md
@@ -48,6 +48,8 @@ Registered in `miles/backends/fsdp_utils/configs/qwen_image.py`:
 Canonical recipe: `scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py` — 4 colocate
 GPUs + 1 PickScore GPU, 512×512, PickScore reward.
 
+**Status:** [📈 V — Verified](/user-guide/recipe-verification#v)
+
 ```bash
 python3 scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py
 ```

diff --git a/docs/models/sd3/sd3.md b/docs/models/sd3/sd3.md
@@ -108,6 +108,8 @@ All recipes are Python modules under `scripts/`. Each exposes a Typer CLI
 
 Canonical script: `scripts/run_diffusion_grpo_sd3_ocr_sglang.py`
 
+**Status:** [🛡️ FG — Fully gated](/user-guide/recipe-verification#fg)
+
 ```bash
 export HF_TOKEN=...
 python3 scripts/run_diffusion_grpo_sd3_ocr_sglang.py \
@@ -125,6 +127,8 @@ E2E test: `tests/e2e/short/test_sd3_ocr_grpo_2xGPU.py`.
 
 Script: `scripts/run_diffusion_nft_sd3_pickscore.py`
 
+**Status:** [📈 V — Verified](/user-guide/recipe-verification#v)
+
 ```bash
 export HF_TOKEN=...
 python3 scripts/run_diffusion_nft_sd3_pickscore.py \

diff --git a/docs/models/wan/wan2-2.md b/docs/models/wan/wan2-2.md
@@ -46,6 +46,8 @@ Registered in `miles/backends/fsdp_utils/configs/wan2_2.py`:
 
 Canonical recipe: `scripts/run_diffusion_grpo_wan22_pickscore_5gpu.py`
 
+**Status:** [○ NV — Not verified](/user-guide/recipe-verification#nv)
+
 ```bash
 python3 scripts/run_diffusion_grpo_wan22_pickscore_5gpu.py
 ```
@@ -57,6 +59,8 @@ Recipe: `scripts/run_diffusion_grpo_wan22_pickscore_17gpu_multinode.py`
 (full finetune, no LoRA, true on-policy). Start the
 [multi-node Ray cluster](/user-guide/launch-script#multi-node-training), then run on the head node:
 
+**Status:** [🧩 PG — Proxy gated](/user-guide/recipe-verification#pg)
+
 ```bash
 MILES_SCRIPT_EXTERNAL_RAY=1 python3 scripts/run_diffusion_grpo_wan22_pickscore_17gpu_multinode.py
 ```
@@ -67,6 +71,8 @@ MILES_SCRIPT_EXTERNAL_RAY=1 python3 scripts/run_diffusion_grpo_wan22_pickscore_1
 
 Recipe: `scripts/run_diffusion_sft_wan22.py`
 
+**Status:** [○ NV — Not verified](/user-guide/recipe-verification#nv)
+
 ```bash
 MILES_SCRIPT_DATA_JSONL=/abs/data.jsonl python3 scripts/run_diffusion_sft_wan22.py
 ```

diff --git a/docs/user-guide/recipe-verification.md b/docs/user-guide/recipe-verification.md
@@ -0,0 +1,54 @@
+---
+title: Recipe Verification Levels
+description: Evidence levels for documented training recipes.
+---
+# Recipe verification levels
+
+Every recipe is assigned one evidence level. The level applies to the exact script and
+topology named in the model guide, not to the model family as a whole.
+
+<a id="fg"></a>
+## 🛡️ FG — Fully gated
+
+- A complete training curve has been run.
+- Deterministic E2E CI runs the canonical recipe itself for two rollouts.
+- Every registered metric matches the committed standard exactly.
+
+<a id="pg"></a>
+## 🧩 PG — Proxy gated
+
+- A complete training curve has been run.
+- Deterministic E2E CI runs a documented, scaled single-node proxy for two rollouts.
+- Every registered metric matches the committed standard exactly.
+
+<a id="v"></a>
+## 📈 V — Verified
+
+A complete training curve has been run, but the recipe has no E2E CI gate.
+
+<a id="nv"></a>
+## ○ NV — Not verified
+
+No complete training curve has been run. Smoke tests and short debugging runs do not
+count as verification.
+
+## Current levels
+
+- **🛡️ FG**
+  - `run_diffusion_grpo_sd3_ocr_sglang.py` — SD3.5 Flow-GRPO + OCR.
+  - `run_diffusion_grpo_ltx23_sglang.py` — LTX-2.3 Flow-GRPO + PickScore.
+- **🧩 PG**
+  - `run_diffusion_grpo_wan22_pickscore_17gpu_multinode.py` — Wan2.2 17-GPU
+    full-finetune Flow-GRPO + PickScore.
+- **📈 V**
+  - `run_diffusion_nft_sd3_pickscore.py` — SD3.5 DiffusionNFT + PickScore.
+  - `run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py` — Qwen-Image
+    Flow-GRPO + PickScore.
+  - `run_diffusion_grpo_cosmos3_pickscore_t2i_4gpu.py` — Cosmos3-Nano
+    Flow-GRPO + PickScore.
+  - `run_diffusion_grpo_h3_t2va_2gpu.py` — MiniMax H3 Flow-GRPO + PickScore
+    ([implementation PR #154](https://github.com/radixark/miles_diffusion/pull/154)).
+- **○ NV**
+  - `run_diffusion_grpo_wan22_pickscore_5gpu.py` — Wan2.2 5-GPU LoRA
+    Flow-GRPO + PickScore.
+  - `run_diffusion_sft_wan22.py` — Wan2.2 4-GPU LoRA SFT.