Publish curated, controls-verified results

README.md

@@ -58,7 +58,7 @@ positive into a correct negative.
 ## Dataset Description
 
 This dataset is the distilled, **verified evidence** from a learnability
-instrument built on top of the [`bfl-asic`](#reproduction) toolkit (a
+instrument built on top of the [`bfl-asic`](https://github.com/bshepp/bfl-asic) toolkit (a
 codebase for a Butterfly Labs BF0005G "Jalapeno" SHA-256 mining ASIC,
 which also contains a numpy-vectorized, `hashlib`-anchored round-reduced
 SHA-256 and a controls-gated train/eval harness).
@@ -266,17 +266,27 @@ are hosted. The results above were produced by the `bfl-asic` toolkit's
 TinyCNN/linear-probe distinguishers, a controls-gated harness), run on
 Hugging Face Jobs (`cpu-xl`, ~16 CPU-hours total).
 
+**Source code:** [github.com/bshepp/bfl-asic](https://github.com/bshepp/bfl-asic) (MIT) — the `bfl_asic/ml/` subsystem and `dataset/build_dataset.py`.
+
 ```bash
-pip install "bfl-asic[ml]"          # PyTorch is isolated behind [ml]
+git clone https://github.com/bshepp/bfl-asic
+cd bfl-asic
+pip install -e ".[ml]"              # PyTorch is isolated behind [ml]
 
 # Regenerate the spine (one seed, scaled down for a laptop):
 bfl-asic ml run sweep --seed 0 --n 20000 --epochs 10
 bfl-asic ml report runs/ml/<timestamp>/sweep_seed0.json
 
-# Rebuild these exact Parquet tables from the synced run JSON:
+# Rebuild these exact Parquet tables from the shipped source JSON
+# (dataset/source/ travels with the repo — no external data needed):
 python dataset/build_dataset.py     # deps: pandas, pyarrow
 ```
 
+This HF dataset repo is itself self-contained: `git clone` it, `pip
+install pandas pyarrow`, run `python build_dataset.py`, and the four
+Parquet rebuild from the bundled `source/` JSON — no external data, no
+GitHub checkout required.
+
 The harness is deterministic: the same seed reproduces the same curve.
 The `dynamics_validated` table is the output of the *fixed* harness
 (real Clopper–Pearson CI + permuted-label control); the earlier
@@ -310,10 +320,11 @@ honest record of the false positive that the control corrected.
   author = {Sheppard, B.},
   year   = {2026},
   publisher = {Hugging Face},
-  url = {https://huggingface.co/datasets/bshepp/round-reduced-sha256-learnability}
+  url = {https://huggingface.co/datasets/bshepp/round-reduced-sha256-learnability},
+  note = {Code: https://github.com/bshepp/bfl-asic}
 }
 ```
 
 ## License
 
-MIT.
+MIT — see the [`bfl-asic` repository](https://github.com/bshepp/bfl-asic/blob/master/LICENSE).

build_dataset.py

@@ -3,8 +3,11 @@
 
 Mirrors the convention of the author's other HF dataset
 (`bshepp/pairwise-poisson-algebras`): a deterministic script that reads
-the raw result JSON and emits one Parquet table per config, sitting next
-to the dataset card (README.md) in this directory.
+the curated result JSON shipped in `source/` and emits one Parquet
+table per config, sitting next to the dataset card (README.md) in this
+directory. The `source/` JSON is the exact, verified output of the HF
+runs that produced the published numbers — it travels with the repo so
+this script runs on a fresh clone with no external data.
 
 Only *verified* results go in (controls passed, or — for the dynamics
 negative — the permuted-label control actively fired). The synthetic
@@ -24,7 +27,7 @@ from pathlib import Path
 import pandas as pd
 
 SCRIPT_DIR = Path(__file__).resolve().parent
-HF = SCRIPT_DIR.parent / "hf_results"
+HF = SCRIPT_DIR / "source"  # curated run JSON shipped with the repo
 _Z = 1.959963984540054  # 97.5th pct of N(0,1); matches the harness
 
 

publish_dataset.py

@@ -5,9 +5,12 @@ Mirrors `bfl_asic/ml/publish.py` (HfApi.create_repo + upload_folder) but
 with `repo_type="dataset"` and public-by-default, matching the author's
 existing HF dataset convention (`bshepp/pairwise-poisson-algebras`).
 
-Uploads only the curated card + Parquet + the two build/publish scripts
-(no synced run JSON, no payloads). Auth comes from the already-configured
-`hf` CLI token (HF_TOKEN env var or `~/.cache/huggingface/token`).
+Uploads a self-contained dataset: the curated card + Parquet + the
+build/publish scripts + the shipped `source/` run JSON, so the HF repo
+rebuilds via `python build_dataset.py` with no external data (no HF
+payloads/buckets). The card's `configs:` pin the four .parquet, so the
+source JSON is inert to the Dataset Viewer / `load_dataset`. Auth comes
+from the configured `hf` CLI token (HF_TOKEN or ~/.cache/huggingface/token).
 
 Usage:
     python dataset/publish_dataset.py                       # default repo, public
@@ -18,23 +21,35 @@ Usage:
 from __future__ import annotations
 
 import argparse
+import fnmatch
 from pathlib import Path
 
 DEFAULT_REPO = "bshepp/round-reduced-sha256-learnability"
-ALLOW = ["README.md", "*.parquet", "build_dataset.py", "publish_dataset.py"]
+ALLOW = [
+    "README.md",
+    "*.parquet",
+    "build_dataset.py",
+    "publish_dataset.py",
+    "source/*",  # fnmatch '*' spans '/', so this is recursive under source/
+]
 
 
 def publish(repo_id: str, *, private: bool, dry_run: bool) -> str:
     folder = Path(__file__).resolve().parent
     url = f"https://huggingface.co/datasets/{repo_id}"
+    # Mirror huggingface_hub's allow_patterns matching exactly (fnmatch
+    # on the repo-relative posix path) so the dry-run is honest.
     files = sorted(
-        p.name
-        for p in folder.iterdir()
-        if p.suffix in (".parquet", ".md", ".py")
+        rel
+        for p in folder.rglob("*")
+        if p.is_file()
+        and any(fnmatch.fnmatch(rel := p.relative_to(folder).as_posix(),
+                                pat) for pat in ALLOW)
     )
     if dry_run:
         print(f"[dry-run] would create dataset repo {repo_id} "
-              f"(private={private}) and upload from {folder}:")
+              f"(private={private}) and upload {len(files)} files "
+              f"from {folder}:")
         for f in files:
             print(f"  + {f}")
         print(f"[dry-run] -> {url}")

source/bfl-ml-tierA/full_structure_seed0.json

@@ -0,0 +1,53 @@
+{
+  "timestamp": "2026-05-16T09:51:12.022788+00:00",
+  "experiment": "full_structure_seed0",
+  "feature": "per-hash",
+  "model": "multi",
+  "points": [
+    {
+      "rounds": 64,
+      "accuracy": 0.5007625,
+      "advantage": 0.0015249999999999986,
+      "auc": null,
+      "accuracy_ci": [
+        0.49830942764154007,
+        0.5032155447256841
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "tiny_cnn"
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.4991125,
+      "advantage": -0.001774999999999971,
+      "auc": null,
+      "accuracy_ci": [
+        0.4966594585498896,
+        0.5015655736128495
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "linear_probe"
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 1.0,
+    "positive_ok": true,
+    "negative_ci": [
+      0.49830942764154007,
+      0.5032155447256841
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {
+    "best_model": "tiny_cnn",
+    "accuracy": 0.5007625,
+    "accuracy_ci": [
+      0.49830942764154007,
+      0.5032155447256841
+    ],
+    "advantage": 0.0015249999999999986,
+    "min_detectable_advantage": 0.004899909961350135,
+    "controls_ok": true,
+    "conclusion": "no structure detected above the detection floor"
+  }
+}

source/bfl-ml-tierA/full_structure_seed1.json

@@ -0,0 +1,53 @@
+{
+  "timestamp": "2026-05-16T10:33:36.555551+00:00",
+  "experiment": "full_structure_seed1",
+  "feature": "per-hash",
+  "model": "multi",
+  "points": [
+    {
+      "rounds": 64,
+      "accuracy": 0.499125,
+      "advantage": -0.0017500000000000293,
+      "auc": null,
+      "accuracy_ci": [
+        0.49667195821544186,
+        0.501578073494301
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "tiny_cnn"
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.500275,
+      "advantage": 0.0005500000000000504,
+      "auc": null,
+      "accuracy_ci": [
+        0.4978219339967231,
+        0.5027280560373577
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "linear_probe"
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 0.99823125,
+    "positive_ok": true,
+    "negative_ci": [
+      0.49667195821544186,
+      0.501578073494301
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {
+    "best_model": "linear_probe",
+    "accuracy": 0.500275,
+    "accuracy_ci": [
+      0.4978219339967231,
+      0.5027280560373577
+    ],
+    "advantage": 0.0005500000000000504,
+    "min_detectable_advantage": 0.004899909961350135,
+    "controls_ok": true,
+    "conclusion": "no structure detected above the detection floor"
+  }
+}

source/bfl-ml-tierA/full_structure_seed2.json

@@ -0,0 +1,53 @@
+{
+  "timestamp": "2026-05-16T11:15:57.238584+00:00",
+  "experiment": "full_structure_seed2",
+  "feature": "per-hash",
+  "model": "multi",
+  "points": [
+    {
+      "rounds": 64,
+      "accuracy": 0.499225,
+      "advantage": -0.0015500000000000513,
+      "auc": null,
+      "accuracy_ci": [
+        0.4967719555949829,
+        0.5016780724907892
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "tiny_cnn"
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.50099375,
+      "advantage": 0.001987500000000031,
+      "auc": null,
+      "accuracy_ci": [
+        0.4985406754413102,
+        0.5034467885454819
+      ],
+      "min_detectable_advantage": 0.004899909961350135,
+      "model": "linear_probe"
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 1.0,
+    "positive_ok": true,
+    "negative_ci": [
+      0.4966282093927081,
+      0.5015343239025218
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {
+    "best_model": "linear_probe",
+    "accuracy": 0.50099375,
+    "accuracy_ci": [
+      0.4985406754413102,
+      0.5034467885454819
+    ],
+    "advantage": 0.001987500000000031,
+    "min_detectable_advantage": 0.004899909961350135,
+    "controls_ok": true,
+    "conclusion": "no structure detected above the detection floor"
+  }
+}

source/bfl-ml-tierA/indistinguishability.json

@@ -0,0 +1,29 @@
+{
+  "timestamp": "2026-05-16T11:54:28.796744+00:00",
+  "experiment": "indistinguishability",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 64,
+      "accuracy": 0.5007625,
+      "advantage": 0.0015249999999999986,
+      "auc": null,
+      "accuracy_ci": [
+        0.49830942764154007,
+        0.5032155447256841
+      ],
+      "min_detectable_advantage": 0.004899909961350135
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 1.0,
+    "positive_ok": true,
+    "negative_ci": [
+      0.49830942764154007,
+      0.5032155447256841
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/bfl-ml-tierA/sweep_seed0.json

@@ -0,0 +1,150 @@
+{
+  "timestamp": "2026-05-16T07:34:49.576047+00:00",
+  "experiment": "sweep_seed0",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 0.99995,
+      "advantage": 0.9999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998193946037771,
+        0.9999939447106779
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 2,
+      "accuracy": 1.0,
+      "advantage": 1.0,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999077822659638,
+        1.0
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 3,
+      "accuracy": 0.999975,
+      "advantage": 0.9999499999999999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998607168748304,
+        0.9999993670550007
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.500475,
+      "advantage": 0.0009500000000000064,
+      "auc": null,
+      "accuracy_ci": [
+        0.4955627069332437,
+        0.5053872240970375
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 6,
+      "accuracy": 0.4994,
+      "advantage": -0.0011999999999999789,
+      "auc": null,
+      "accuracy_ci": [
+        0.4944877862946945,
+        0.5043123008249503
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 8,
+      "accuracy": 0.50295,
+      "advantage": 0.005900000000000016,
+      "auc": null,
+      "accuracy_ci": [
+        0.49803761031634497,
+        0.5078619613453772
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 12,
+      "accuracy": 0.50155,
+      "advantage": 0.0031000000000001027,
+      "auc": null,
+      "accuracy_ci": [
+        0.4966376502202616,
+        0.5064621247206528
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 16,
+      "accuracy": 0.5032,
+      "advantage": 0.006399999999999961,
+      "auc": null,
+      "accuracy_ci": [
+        0.4982876072329792,
+        0.5081119281288841
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 24,
+      "accuracy": 0.50225,
+      "advantage": 0.0044999999999999485,
+      "auc": null,
+      "accuracy_ci": [
+        0.49733762546654603,
+        0.5071620478347755
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 32,
+      "accuracy": 0.50025,
+      "advantage": 0.0004999999999999449,
+      "auc": null,
+      "accuracy_ci": [
+        0.49533772166968126,
+        0.5051622420304669
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 48,
+      "accuracy": 0.50265,
+      "advantage": 0.005300000000000082,
+      "auc": null,
+      "accuracy_ci": [
+        0.49773761563333563,
+        0.5075619995882157
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.502025,
+      "advantage": 0.004050000000000109,
+      "auc": null,
+      "accuracy_ci": [
+        0.49711263237578385,
+        0.5069370735954074
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 1.0,
+    "positive_ok": true,
+    "negative_ci": [
+      0.49643765905663984,
+      0.5062621449241573
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/bfl-ml-tierA/sweep_seed1.json

@@ -0,0 +1,150 @@
+{
+  "timestamp": "2026-05-16T08:16:58.741310+00:00",
+  "experiment": "sweep_seed1",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 0.999925,
+      "advantage": 0.9998499999999999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9997808339737771,
+        0.9999845329298606
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 2,
+      "accuracy": 0.99995,
+      "advantage": 0.9999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998193946037771,
+        0.9999939447106779
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 3,
+      "accuracy": 0.999975,
+      "advantage": 0.9999499999999999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998607168748304,
+        0.9999993670550007
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.498775,
+      "advantage": -0.002449999999999952,
+      "auc": null,
+      "accuracy_ci": [
+        0.49386284284674964,
+        0.5036873350225266
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 6,
+      "accuracy": 0.50565,
+      "advantage": 0.011300000000000088,
+      "auc": null,
+      "accuracy_ci": [
+        0.500737641843862,
+        0.5105615377793082
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 8,
+      "accuracy": 0.49815,
+      "advantage": -0.0037000000000000366,
+      "auc": null,
+      "accuracy_ci": [
+        0.4932379070545073,
+        0.5030623615644031
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 12,
+      "accuracy": 0.495,
+      "advantage": -0.010000000000000009,
+      "auc": null,
+      "accuracy_ci": [
+        0.49008834719199657,
+        0.49991237880516426
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 16,
+      "accuracy": 0.49775,
+      "advantage": -0.0044999999999999485,
+      "auc": null,
+      "accuracy_ci": [
+        0.4928379521652245,
+        0.502662374533454
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 24,
+      "accuracy": 0.5028,
+      "advantage": 0.005600000000000049,
+      "auc": null,
+      "accuracy_ci": [
+        0.4978876127543479,
+        0.5077119806872891
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 32,
+      "accuracy": 0.497825,
+      "advantage": -0.004349999999999965,
+      "auc": null,
+      "accuracy_ci": [
+        0.49291294346810305,
+        0.5027373723406188
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 48,
+      "accuracy": 0.49795,
+      "advantage": -0.0040999999999999925,
+      "auc": null,
+      "accuracy_ci": [
+        0.4930379292178877,
+        0.5028623684409066
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.4992,
+      "advantage": -0.0016000000000000458,
+      "auc": null,
+      "accuracy_ci": [
+        0.4942878035584158,
+        0.504112312601111
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 0.99995,
+    "positive_ok": true,
+    "negative_ci": [
+      0.4927629609725905,
+      0.5025873766160448
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/bfl-ml-tierA/sweep_seed2.json

@@ -0,0 +1,150 @@
+{
+  "timestamp": "2026-05-16T09:08:34.034602+00:00",
+  "experiment": "sweep_seed2",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 0.999975,
+      "advantage": 0.9999499999999999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998607168748304,
+        0.9999993670550007
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 2,
+      "accuracy": 1.0,
+      "advantage": 1.0,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999077822659638,
+        1.0
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 3,
+      "accuracy": 0.9998,
+      "advantage": 0.9996,
+      "auc": null,
+      "accuracy_ci": [
+        0.999605958510213,
+        0.9999136503678103
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.50335,
+      "advantage": 0.006699999999999928,
+      "auc": null,
+      "accuracy_ci": [
+        0.49843760597094716,
+        0.5082619076110002
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 6,
+      "accuracy": 0.501375,
+      "advantage": 0.00275000000000003,
+      "auc": null,
+      "accuracy_ci": [
+        0.49646265790922056,
+        0.5062871424415913
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 8,
+      "accuracy": 0.49895,
+      "advantage": -0.0020999999999999908,
+      "auc": null,
+      "accuracy_ci": [
+        0.49403782624048226,
+        0.5038623262188968
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 12,
+      "accuracy": 0.503325,
+      "advantage": 0.006650000000000045,
+      "auc": null,
+      "accuracy_ci": [
+        0.49841260615066135,
+        0.5082369110612721
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 16,
+      "accuracy": 0.496075,
+      "advantage": -0.007850000000000024,
+      "auc": null,
+      "accuracy_ci": [
+        0.4911631751262332,
+        0.5009873947815012
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 24,
+      "accuracy": 0.500225,
+      "advantage": 0.00045000000000006146,
+      "auc": null,
+      "accuracy_ci": [
+        0.49531272336830834,
+        0.5051372439618249
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 32,
+      "accuracy": 0.49975,
+      "advantage": -0.0004999999999999449,
+      "auc": null,
+      "accuracy_ci": [
+        0.49483775796953317,
+        0.5046622783303187
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 48,
+      "accuracy": 0.502275,
+      "advantage": 0.004550000000000054,
+      "auc": null,
+      "accuracy_ci": [
+        0.4973626247600997,
+        0.5071870449112362
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.500275,
+      "advantage": 0.0005500000000000504,
+      "auc": null,
+      "accuracy_ci": [
+        0.4953627199833032,
+        0.5051872400868596
+      ],
+      "min_detectable_advantage": 0.00979981992270027
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 1.0,
+    "positive_ok": true,
+    "negative_ci": [
+      0.49708763320472354,
+      0.5069120763964532
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/bfl-ml-tierB/sweep_seed0.json

@@ -0,0 +1,205 @@
+{
+  "timestamp": "2026-05-16T15:40:23.083515+00:00",
+  "experiment": "sweep_seed0",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 1.0,
+      "advantage": 1.0,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999631118858421,
+        1.0
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 2,
+      "accuracy": 0.99999,
+      "advantage": 0.9999800000000001,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999442848396523,
+        0.9999997468219523
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 3,
+      "accuracy": 0.99999,
+      "advantage": 0.9999800000000001,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999442848396523,
+        0.9999997468219523
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.49862,
+      "advantage": -0.0027599999999999847,
+      "auc": null,
+      "accuracy_ci": [
+        0.49551611432120896,
+        0.5017239657314021
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 5,
+      "accuracy": 0.49797,
+      "advantage": -0.0040599999999999525,
+      "auc": null,
+      "accuracy_ci": [
+        0.49486614691167463,
+        0.5010739708468777
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 6,
+      "accuracy": 0.4979,
+      "advantage": -0.0041999999999999815,
+      "auc": null,
+      "accuracy_ci": [
+        0.49479615073379235,
+        0.5010039710853998
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 8,
+      "accuracy": 0.49807,
+      "advantage": -0.0038599999999999746,
+      "auc": null,
+      "accuracy_ci": [
+        0.4949661415568716,
+        0.5011739704007666
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 10,
+      "accuracy": 0.49994,
+      "advantage": -0.00012000000000000899,
+      "auc": null,
+      "accuracy_ci": [
+        0.49683606425438664,
+        0.5030439392261616
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 12,
+      "accuracy": 0.49789,
+      "advantage": -0.0042200000000000015,
+      "auc": null,
+      "accuracy_ci": [
+        0.49478615128476744,
+        0.5009939711145162
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 16,
+      "accuracy": 0.4955,
+      "advantage": -0.009000000000000008,
+      "auc": null,
+      "accuracy_ci": [
+        0.49239631852025956,
+        0.4986039425208842
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 20,
+      "accuracy": 0.49895,
+      "advantage": -0.0020999999999999908,
+      "auc": null,
+      "accuracy_ci": [
+        0.4958460997796744,
+        0.5020539611299208
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 24,
+      "accuracy": 0.49973,
+      "advantage": -0.0005399999999999849,
+      "auc": null,
+      "accuracy_ci": [
+        0.49662607077485327,
+        0.502833944887614
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 32,
+      "accuracy": 0.49817,
+      "advantage": -0.0036599999999999966,
+      "auc": null,
+      "accuracy_ci": [
+        0.49506613632602714,
+        0.5012739698306969
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 40,
+      "accuracy": 0.49779,
+      "advantage": -0.0044199999999999795,
+      "auc": null,
+      "accuracy_ci": [
+        0.49468615686269696,
+        0.5008939713375008
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 48,
+      "accuracy": 0.49853,
+      "advantage": -0.0029400000000000537,
+      "auc": null,
+      "accuracy_ci": [
+        0.49542611852136115,
+        0.5016339667520726
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 56,
+      "accuracy": 0.49797,
+      "advantage": -0.0040599999999999525,
+      "auc": null,
+      "accuracy_ci": [
+        0.49486614691167463,
+        0.5010739708468777
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.50157,
+      "advantage": 0.0031399999999999206,
+      "auc": null,
+      "accuracy_ci": [
+        0.4984660322316088,
+        0.5046738766940436
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 0.99999,
+    "positive_ok": true,
+    "negative_ci": [
+      0.4951061342683978,
+      0.5013139695679607
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/bfl-ml-tierB/sweep_seed1.json

@@ -0,0 +1,205 @@
+{
+  "timestamp": "2026-05-16T18:36:44.901723+00:00",
+  "experiment": "sweep_seed1",
+  "feature": "per-hash",
+  "model": "tiny_cnn",
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 0.99995,
+      "advantage": 0.9999,
+      "auc": null,
+      "accuracy_ci": [
+        0.9998833205695797,
+        0.999983764943183
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 2,
+      "accuracy": 0.99999,
+      "advantage": 0.9999800000000001,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999442848396523,
+        0.9999997468219523
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 3,
+      "accuracy": 1.0,
+      "advantage": 1.0,
+      "auc": null,
+      "accuracy_ci": [
+        0.9999631118858421,
+        1.0
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.49701,
+      "advantage": -0.005979999999999985,
+      "auc": null,
+      "accuracy_ci": [
+        0.4939062046248662,
+        0.5001139688224632
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 5,
+      "accuracy": 0.50168,
+      "advantage": 0.0033600000000000296,
+      "auc": null,
+      "accuracy_ci": [
+        0.4985760312568294,
+        0.5047838712878174
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 6,
+      "accuracy": 0.49809,
+      "advantage": -0.0038200000000000456,
+      "auc": null,
+      "accuracy_ci": [
+        0.494986140500786,
+        0.5011939702966695
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 8,
+      "accuracy": 0.49898,
+      "advantage": -0.0020400000000000418,
+      "auc": null,
+      "accuracy_ci": [
+        0.49587609852465336,
+        0.5020839606446676
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 10,
+      "accuracy": 0.49701,
+      "advantage": -0.005979999999999985,
+      "auc": null,
+      "accuracy_ci": [
+        0.4939062046248662,
+        0.5001139688224632
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 12,
+      "accuracy": 0.4997,
+      "advantage": -0.0006000000000000449,
+      "auc": null,
+      "accuracy_ci": [
+        0.49659607175097265,
+        0.5028039456517687
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 16,
+      "accuracy": 0.50064,
+      "advantage": 0.0012799999999999478,
+      "auc": null,
+      "accuracy_ci": [
+        0.497536046467494,
+        0.5037439164066575
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 20,
+      "accuracy": 0.49974,
+      "advantage": -0.0005199999999999649,
+      "auc": null,
+      "accuracy_ci": [
+        0.49663607045195934,
+        0.5028439446304165
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 24,
+      "accuracy": 0.50222,
+      "advantage": 0.0044399999999999995,
+      "auc": null,
+      "accuracy_ci": [
+        0.4991160286470185,
+        0.5053238425726923
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 32,
+      "accuracy": 0.49701,
+      "advantage": -0.005979999999999985,
+      "auc": null,
+      "accuracy_ci": [
+        0.4939062046248662,
+        0.5001139688224632
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 40,
+      "accuracy": 0.49701,
+      "advantage": -0.005979999999999985,
+      "auc": null,
+      "accuracy_ci": [
+        0.4939062046248662,
+        0.5001139688224632
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 48,
+      "accuracy": 0.5002,
+      "advantage": 0.00039999999999995595,
+      "auc": null,
+      "accuracy_ci": [
+        0.4970960569387984,
+        0.5033039314593739
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 56,
+      "accuracy": 0.49856,
+      "advantage": -0.0028799999999999937,
+      "auc": null,
+      "accuracy_ci": [
+        0.4954561171101542,
+        0.5016639664230054
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    },
+    {
+      "rounds": 64,
+      "accuracy": 0.49865,
+      "advantage": -0.0027000000000000357,
+      "auc": null,
+      "accuracy_ci": [
+        0.49554611294347056,
+        0.5017539653688663
+      ],
+      "min_detectable_advantage": 0.006197950323045616
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 0.99999,
+    "positive_ok": true,
+    "negative_ci": [
+      0.4939062046248662,
+      0.5001139688224632
+    ],
+    "negative_ok": true
+  },
+  "bounded_null": {}
+}

source/dynamics_validated_seed0.json

@@ -0,0 +1,80 @@
+{
+  "config": {
+    "seed": 0,
+    "n": 20000,
+    "epochs": 25,
+    "trunc_widths": [
+      1,
+      2,
+      3,
+      4
+    ],
+    "n_bins": 4
+  },
+  "elapsed_s": 2124.1,
+  "points": [
+    {
+      "rounds": 1,
+      "accuracy": 0.3535,
+      "advantage": 0.10349999999999998,
+      "auc": null,
+      "accuracy_ci": [
+        0.33867165024339485,
+        0.3685426437655393
+      ],
+      "min_detectable_advantage": 0.013418956077878678,
+      "chance": 0.25
+    },
+    {
+      "rounds": 2,
+      "accuracy": 0.249,
+      "advantage": -0.0010000000000000009,
+      "auc": null,
+      "accuracy_ci": [
+        0.23566118455531623,
+        0.26270639218531416
+      ],
+      "min_detectable_advantage": 0.013418956077878678,
+      "chance": 0.25
+    },
+    {
+      "rounds": 3,
+      "accuracy": 0.249,
+      "advantage": -0.0010000000000000009,
+      "auc": null,
+      "accuracy_ci": [
+        0.23566118455531623,
+        0.26270639218531416
+      ],
+      "min_detectable_advantage": 0.013418956077878678,
+      "chance": 0.25
+    },
+    {
+      "rounds": 4,
+      "accuracy": 0.24275,
+      "advantage": -0.0072500000000000064,
+      "auc": null,
+      "accuracy_ci": [
+        0.22953001188344008,
+        0.25634675686691016
+      ],
+      "min_detectable_advantage": 0.013418956077878678,
+      "chance": 0.25
+    }
+  ],
+  "controls": {
+    "positive_accuracy": 0.3535,
+    "positive_ci": [
+      0.33867165024339485,
+      0.3685426437655393
+    ],
+    "positive_ok": true,
+    "permuted_label_accuracy": 0.3535,
+    "permuted_label_ci": [
+      0.33867165024339485,
+      0.3685426437655393
+    ],
+    "negative_ok": false,
+    "note": "knob=truncation width (bytes); chance=1/n_bins; advantage=accuracy-chance; positive_ok = CI lower bound > chance; negative_ok = permuted-label model CI lower bound <= chance (no setup/dataset artifact)."
+  }
+}