Record clean frm download tests and exact selected-seed configs

README.md

@@ -38,6 +38,9 @@ All model files, benchmark arrays and calibration are downloaded automatically a
 verified by SHA-256. The default is calibrated sequential restarts. Add
 `--max-samples 8` for a functional smoke test, not a reported benchmark score.
 Use `--revision <Hub commit>` to pin the bundle. Outputs resume automatically.
+The evaluator computes a fixed-depth comparison too: evaluation wall-clock work
+includes that extra baseline; reported NFE/FLOPs describe the selected restart
+policy, not the total cost of generating all three comparison rows.
 
 ## Selected models and measured results
 
@@ -72,8 +75,7 @@ that recorded SDE evaluation, not a comparison of restart policies across seeds.
 Each folder contains `model.pt`, `training.yaml`, `threshold.json`, `test.npz`,
 `calibration.npz`, `manifest.json`, `provenance.json`, and `expected_results.json`.
 The manifest records the original checkpoint hash and the compact inference-checkpoint
-hash. Optimizer state is omitted; raw weights and EMA are retained. These files are
-our own PyTorch checkpoints and should be loaded only from this trusted release.
+hash. Optimizer state is omitted; raw weights and EMA are retained. These are compact PyTorch checkpoints from the recorded training runs.
 For training from scratch, download `training.yaml` and run
 `frm-train --config training.yaml --seed <recorded seed>` on the matching data.
 

maze_unique/manifest.json

@@ -19,5 +19,5 @@
     "calibration.npz": "3f0d8261ef2977c8a8d219dcb3f6cd146d3e7195438cfbe8939e774d1d8ff0dc",
     "source_evaluation.json": "cc68b4b64c9aa49b7a829ba36206dc05fd928668a349955b9263c3a1a4b4bcac"
   },
-  "code_revision": "0e4c69c84416c708241c58b211ab225fa346c6d9"
+  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
 }

release.json

@@ -1,7 +1,7 @@
 {
   "code_repo": "helblazer811/FlowMapReasoning",
   "code_ref": "public-simplified",
-  "code_revision": "0e4c69c84416c708241c58b211ab225fa346c6d9",
+  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3",
   "package": "frm",
   "pypi_published": false,
   "scope": "private draft"

sudoku/manifest.json

@@ -19,5 +19,5 @@
     "calibration.npz": "e1f3b99cd5ce91dfe1e7de44995e1e38bbf1372766457025eca589625e89281e",
     "source_evaluation.json": "6085f2d71fe4aa39d253820326922de1236a142527fabf5746c8202269e0251c"
   },
-  "code_revision": "0e4c69c84416c708241c58b211ab225fa346c6d9"
+  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
 }

verification.json

@@ -0,0 +1,24 @@
+{
+  "mode": "functional smoke, not full benchmark re-evaluation",
+  "device": "cpu",
+  "num_puzzles_per_task": 1,
+  "installed_from": "GitHub/public-simplified",
+  "package": "frm",
+  "package_version": "0.2.0",
+  "tested_code_revision": "5c40fd89631c5f7384589b408f6591dd98057621",
+  "tested_bundle_revision": "3e8c45fd28b56f65c072334fc25c51328aa57850",
+  "tasks": {
+    "maze_unique": {
+      "passed": true,
+      "model_sha256": "3f0023f50352d395ca97a1d4a5c08a0f4bae4140fb53411a0a9c3840f7a7a4e3"
+    },
+    "sudoku": {
+      "passed": true,
+      "model_sha256": "5952ab8fb4184cffa7580b2035bc269531350d27e95cbdb4d26c9d5b241066c2"
+    },
+    "zebra": {
+      "passed": true,
+      "model_sha256": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4"
+    }
+  }
+}

zebra/manifest.json

@@ -9,7 +9,7 @@
   "checkpoint_format": "raw model plus EMA/config/provenance; optimizer omitted",
   "scope": "5x5 Zebra, repacked answers (not paper mixed-size pinned)",
   "files": {
-    "training.yaml": "f84ddff47ae2208ea1795e21c0b14790aa9f29c8ca1750555ccb674d8810def0",
+    "training.yaml": "29d7da5265a805190959d234de1f3702591c15632b80ac713f3744706147d3c2",
     "model.pt": "2a94118070b94af73c9233ecf81400638ea5d3b4bc99b5b57169b796fcc847c4",
     "provenance.json": "476fdce0c4bdd67e05c3e7554c9a6f99a9b54406eac859cc59ca48dc293e8974",
     "expected_results.json": "46a370d09021d2f900475851296b17fcd7fbe332dd5b4ac48fe7878dcf051cfd",
@@ -21,5 +21,5 @@
     "source_evaluation.json": "8466835dc4e90919c24cfefcc5de98814544e14c42b9f076631fda0d97397375"
   },
   "calibration_note": "256 unique 5x5 training rows, seed20260910, explicitly excludes test puzzles; not the mixed-size paper policy.",
-  "code_revision": "0e4c69c84416c708241c58b211ab225fa346c6d9"
+  "code_revision": "e18449e2c089e83f2069a750fe78aad091f2a3e3"
 }

zebra/training.yaml

@@ -1,7 +1,7 @@
-# Source: public / zebra_train / zebra_25m_stage{A,B}.
 data:
   dataset: zebra
-  sizes: [5x5]
+  sizes:
+  - 5x5
   repack: true
   pin_answer: false
 model:
@@ -12,7 +12,7 @@ model:
   dropout: 0.1
   self_cond: true
 train:
-  lr: 3.0e-4
+  lr: 0.0003
   weight_decay: 0.3
   ema_decay: 0.9999
   local_batch_size: 128
@@ -22,8 +22,8 @@ train:
   val_every: 5000
   save_best_metric: val/solved
 pipeline:
-  seed: 0
-  num_workers: 4
+  seed: 1
+  num_workers: 2
   stage_a_steps: 200000
   stage_b_steps: 200000
   stage_b_weight_decay: 0.0