Rebuild rl/: restore the proven shaped reward, lean campaign machinery

data/README.md

@@ -29,6 +29,12 @@ The benchmark input fixtures (silhouette + footprint per device) are
 `trajectories/` (cohort-pinned to the audited model set; `--all` widens to
 later runs the trajectories also carry).
 
+`training/qwen-rl-v2/` is the validated evidence for the released model
+([`qpaig-mit/pixcell`](https://huggingface.co/qpaig-mit/pixcell), a GRPO-trained
+LoRA for Qwen3.6-35B-A3B): the F1–F8 benchmark records and the
+attempt–measure–revise deployment-loop record with every champion program;
+the render gallery regenerates from it (`rl/track_a/make_gallery.py`). Its
+README maps each number to its file.
 `training/representation-curriculum-v1/` preserves superseded Qwen and Inkling
 diagnostic runs. Those runs used the retired v1 direct prompt contract and are
 not canonical baselines or paper results. Their protocol and artifacts remain
@@ -135,10 +141,12 @@ Every number in the paper's Results, mapped to its backing file.
 
 ### Sec V — verifier-derived training
 
-The repository contains the training recipes under `rl/` and the derived
-paper table and gallery under `arxiv/`. The full benchmark-attempt and
-training-run ledgers are reserved for the companion training-artifact
-release, as stated in the paper's data-availability paragraph.
+The repository contains the training recipes under `rl/`, the derived
+paper table and gallery under `arxiv/`, and the validated evidence for the
+released model under `training/qwen-rl-v2/` (benchmark JSONs, deployment-loop
+record, champion programs). Raw step-level training ledgers and
+checkpoints stay outside the repository; the released adapter itself is
+[`qpaig-mit/pixcell`](https://huggingface.co/qpaig-mit/pixcell).
 
 ## Curation notes (read before reusing)
 

data/training/qwen-rl-v2/README.md

@@ -0,0 +1,38 @@
+# Qwen RL v2 — evidence for the released model
+
+The validated record behind the released PixCell model
+([`qpaig-mit/pixcell`](https://huggingface.co/qpaig-mit/pixcell)): a LoRA
+adapter for `Qwen/Qwen3.6-35B-A3B` trained with GRPO alone (no SFT) on the
+released curriculum L0→L4 at the non-thinking 4096-token operating point —
+Run B of the campaign recipe in [`rl/`](../../../rl/README.md). Checkpoint of
+record: `tinker://d8269c70-0dd9-5776-9cae-53e8b24e647a:train:0/sampler_weights/final`.
+
+Per the repository rule, every number here was recomputed from artifacts by
+the deterministic evaluator; nothing model-reported enters the record.
+
+## Files
+
+| Path | Protocol | Headline |
+|---|---|---|
+| `bench/base-nothink.json` | F1–F8, 8 attempts @ T=1.0, best-of-8 raw IoU | base model: 0/64 attempts executable |
+| `bench/b-final.json` | same | released model: 39/64 executable, mean IoU 0.228, best-of-8 0.467 |
+| `bench/a-post-l1.json`, `bench/d-post-l0.json` | same | mid-campaign controls (L0-SFT arm after L1; thinking arm after L0) |
+| `agentloop/b-final.json` | attempt–measure–revise loop: 8 attempts, then 3 rounds × 4 revisions of the champion under number-free feedback | **mean champion IoU 0.491**; includes every champion program and per-round trajectory |
+
+The agent loop is the deployment protocol — the released model is the
+searcher inside that loop, and 0.491 is its number. The single-pass benchmark
+is the ablation that shows what a bare forward pass gives up.
+
+## Recompute
+
+```bash
+uv run python -m rl.track_a.benchmark_eval --model-path <checkpoint> --out bench.json
+uv run python -m rl.track_a.agent_loop     --model-path <checkpoint> --out agentloop.json
+uv run python -m rl.track_a.make_gallery   --champions agentloop.json --out-dir gallery/
+```
+
+The render gallery (per device: target | champion render | difference map)
+is not stored here: `make_gallery` re-executes each stored champion program
+from source and re-renders it against the frozen `data/benchmark` fixtures,
+so it regenerates from this record alone, no checkpoint required. The
+composite ships on the [model card](https://huggingface.co/qpaig-mit/pixcell).

data/training/qwen-rl-v2/bench/a-post-l1.json

@@ -0,0 +1,62 @@
+{
+  "model_name": "Qwen/Qwen3.6-35B-A3B",
+  "model_path": "tinker://da5c0ccb-55cb-547c-aa0a-b5169801186a:train:0/sampler_weights/final",
+  "attempts": 8,
+  "temperature": 1.0,
+  "per_target": {
+    "final_1": {
+      "executable": 2,
+      "attempts": 8,
+      "mean_iou": 0.04667048709503287,
+      "best_iou": 0.30297820032053774
+    },
+    "final_2": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_3": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_4": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_5": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_6": {
+      "executable": 7,
+      "attempts": 8,
+      "mean_iou": 0.39151451399245063,
+      "best_iou": 0.6434092192544019
+    },
+    "final_7": {
+      "executable": 6,
+      "attempts": 8,
+      "mean_iou": 0.16893688302640955,
+      "best_iou": 0.45974476777216944
+    },
+    "final_8": {
+      "executable": 4,
+      "attempts": 8,
+      "mean_iou": 0.18510123992188554,
+      "best_iou": 0.46525473940028256
+    }
+  },
+  "overall": {
+    "executable": 19,
+    "total": 64,
+    "mean_iou": 0.09902789050447232,
+    "mean_best_of_k": 0.23392336584342394
+  }
+}

data/training/qwen-rl-v2/bench/b-final.json

@@ -0,0 +1,62 @@
+{
+  "model_name": "Qwen/Qwen3.6-35B-A3B",
+  "model_path": "tinker://d8269c70-0dd9-5776-9cae-53e8b24e647a:train:0/sampler_weights/final",
+  "attempts": 8,
+  "temperature": 1.0,
+  "per_target": {
+    "final_1": {
+      "executable": 7,
+      "attempts": 8,
+      "mean_iou": 0.24696953452861198,
+      "best_iou": 0.43563673082449117
+    },
+    "final_2": {
+      "executable": 8,
+      "attempts": 8,
+      "mean_iou": 0.42891641561783983,
+      "best_iou": 0.5225050774837938
+    },
+    "final_3": {
+      "executable": 4,
+      "attempts": 8,
+      "mean_iou": 0.0815497234827478,
+      "best_iou": 0.21433973354480168
+    },
+    "final_4": {
+      "executable": 5,
+      "attempts": 8,
+      "mean_iou": 0.31383181845503727,
+      "best_iou": 0.6632708820396994
+    },
+    "final_5": {
+      "executable": 3,
+      "attempts": 8,
+      "mean_iou": 0.06311243742627864,
+      "best_iou": 0.24391132501379273
+    },
+    "final_6": {
+      "executable": 3,
+      "attempts": 8,
+      "mean_iou": 0.20205123415863877,
+      "best_iou": 0.6433213909378293
+    },
+    "final_7": {
+      "executable": 2,
+      "attempts": 8,
+      "mean_iou": 0.10999817415811014,
+      "best_iou": 0.48292264524221046
+    },
+    "final_8": {
+      "executable": 7,
+      "attempts": 8,
+      "mean_iou": 0.3748545833063247,
+      "best_iou": 0.5264526994671637
+    }
+  },
+  "overall": {
+    "executable": 39,
+    "total": 64,
+    "mean_iou": 0.22766049014169865,
+    "mean_best_of_k": 0.4665450605692228
+  }
+}

data/training/qwen-rl-v2/bench/base-nothink.json

@@ -0,0 +1,62 @@
+{
+  "model_name": "Qwen/Qwen3.6-35B-A3B",
+  "model_path": null,
+  "attempts": 8,
+  "temperature": 1.0,
+  "per_target": {
+    "final_1": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_2": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_3": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_4": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_5": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_6": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_7": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_8": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    }
+  },
+  "overall": {
+    "executable": 0,
+    "total": 64,
+    "mean_iou": 0.0,
+    "mean_best_of_k": 0.0
+  }
+}

data/training/qwen-rl-v2/bench/d-post-l0.json

@@ -0,0 +1,62 @@
+{
+  "model_name": "Qwen/Qwen3.6-35B-A3B",
+  "model_path": "tinker://b1eeca1b-b1c1-55b8-87c7-fd19a38a3088:train:0/sampler_weights/final",
+  "attempts": 8,
+  "temperature": 1.0,
+  "per_target": {
+    "final_1": {
+      "executable": 1,
+      "attempts": 8,
+      "mean_iou": 0.01881755829330524,
+      "best_iou": 0.15054046634644191
+    },
+    "final_2": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_3": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_4": {
+      "executable": 6,
+      "attempts": 8,
+      "mean_iou": 0.37353030776040386,
+      "best_iou": 0.6464024496792284
+    },
+    "final_5": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_6": {
+      "executable": 5,
+      "attempts": 8,
+      "mean_iou": 0.4082847723288638,
+      "best_iou": 0.764634685414781
+    },
+    "final_7": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    },
+    "final_8": {
+      "executable": 0,
+      "attempts": 8,
+      "mean_iou": 0.0,
+      "best_iou": 0.0
+    }
+  },
+  "overall": {
+    "executable": 12,
+    "total": 64,
+    "mean_iou": 0.10007907979782162,
+    "mean_best_of_k": 0.19519720018005643
+  }
+}