Datasets:
DOI:
License:
| # PROVENANCE — Lazy_private | |
| *Generated by `experiments/build_provenance.py` on 2026-10-21. Every figure is read from an artefact on disk; none is retyped. Tags: `battery_v2.0` (the test battery as used), `validation_v2.0` (the eight references, licence tables and verdict). The registration cites `validation_v2.0` by hash; the phase battery will be a new tag.* | |
| ## Freeze scope | |
| The eight references below. **Nothing further is recorded before posting** — the Qwen arm, the read-site fix and the phase battery all follow it. | |
| ## Inventory | |
| - CPython `Parser/Python.asdl` **3.12.13**, sha256 `ecc26bc2818ebdde` | |
| - predicate R: **63** AST constructors kept, **48** excluded; **143** builtin names; total **206** | |
| - tier: deep **76**, shallow **130** — union of four declared sets, criterion 40, outcome 76 (`calibration/tier.json`) | |
| - verification: all **35** Python keywords map to ≥1 construct | |
| ## Calibration | |
| Corpus: codeparrot-clean, 253,398 parsed files of 300,000. Table hashes (sha256, first 16): | |
| | table | sha256 | | |
| |---|---| | |
| | `companions.json` | `e32513a78fe323d9` | | |
| | `construct_rates.json` | `9faac903de18c578` | | |
| | `core_axis_rates.json` | `8c9363adaac023ae` | | |
| | `corpus.json` | `9cadc21633600b56` | | |
| | `depth.json` | `c1f9d05973e29b01` | | |
| | `exposure.json` | `57e658ac11c2fda4` | | |
| | `identifiers.json` | `4d5ef93560fc9267` | | |
| | `identifiers_raw_top.json` | `1d0a9ddf64fc64ac` | | |
| | `length_bands.json` | `a6e1c9fe4adb63c1` | | |
| | `literals.json` | `9cc2dbffc86f0519` | | |
| | `padding_pool.json` | `ea507b30bc46043a` | | |
| | `scope.json` | `62c257957b77aa98` | | |
| ## Generation | |
| - prompts: **{'twin1/AnnAssign/A': {'path': '$ATLAS_DATA/battery/twin1/AnnAssign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/AnnAssign/B': {'path': '$ATLAS_DATA/battery/twin1/AnnAssign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 106.0, 'surplus': 1250, 'paired': 143}, 'twin1/Assert/A': {'path': '$ATLAS_DATA/battery/twin1/Assert_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Assert/B': {'path': '$ATLAS_DATA/battery/twin1/Assert_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 45.65, 'surplus': 1800, 'paired': 0}, 'twin1/Assign/A': {'path': '$ATLAS_DATA/battery/twin1/Assign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Assign/B': {'path': '$ATLAS_DATA/battery/twin1/Assign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 33.0, 'surplus': 847, 'paired': 0}, 'twin1/Attribute/A': {'path': '$ATLAS_DATA/battery/twin1/Attribute_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Attribute/B': {'path': '$ATLAS_DATA/battery/twin1/Attribute_B.jsonl.gz', 'n': 632, 'valid': 632, 'offset_residual': 112.65, 'surplus': 632, 'paired': 0}, 'twin1/AugAssign/A': {'path': '$ATLAS_DATA/battery/twin1/AugAssign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/AugAssign/B': {'path': '$ATLAS_DATA/battery/twin1/AugAssign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 43.1, 'surplus': 866, 'paired': 0}, 'twin1/BinOp/A': {'path': '$ATLAS_DATA/battery/twin1/BinOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/BinOp/B': {'path': '$ATLAS_DATA/battery/twin1/BinOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 64.9, 'surplus': 968, 'paired': 0}, 'twin1/BoolOp/A': {'path': '$ATLAS_DATA/battery/twin1/BoolOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/BoolOp/B': {'path': '$ATLAS_DATA/battery/twin1/BoolOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 43.3, 'surplus': 1800, 'paired': 0}, 'twin1/Call/A': {'path': '$ATLAS_DATA/battery/twin1/Call_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Call/B': {'path': '$ATLAS_DATA/battery/twin1/Call_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 127.0, 'surplus': 1306, 'paired': 0}, 'twin1/Compare/A': {'path': '$ATLAS_DATA/battery/twin1/Compare_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Compare/B': {'path': '$ATLAS_DATA/battery/twin1/Compare_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 31.75, 'surplus': 1350, 'paired': 0}, 'twin1/Delete/A': {'path': '$ATLAS_DATA/battery/twin1/Delete_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Delete/B': {'path': '$ATLAS_DATA/battery/twin1/Delete_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 43.3, 'surplus': 1800, 'paired': 0}, 'twin1/Dict/A': {'path': '$ATLAS_DATA/battery/twin1/Dict_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Dict/B': {'path': '$ATLAS_DATA/battery/twin1/Dict_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 21.0, 'surplus': 1084, 'paired': 63}, 'twin1/For/A': {'path': '$ATLAS_DATA/battery/twin1/For_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/For/B': {'path': '$ATLAS_DATA/battery/twin1/For_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 66.3, 'surplus': 2250, 'paired': 179}, 'twin1/FunctionDef/A': {'path': '$ATLAS_DATA/battery/twin1/FunctionDef_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/FunctionDef/B': {'path': '$ATLAS_DATA/battery/twin1/FunctionDef_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 72.15, 'surplus': 2160, 'paired': 113}, 'twin1/GeneratorExp/A': {'path': '$ATLAS_DATA/battery/twin1/GeneratorExp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/GeneratorExp/B': {'path': '$ATLAS_DATA/battery/twin1/GeneratorExp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 119.9, 'surplus': 2250, 'paired': 143}, 'twin1/If/A': {'path': '$ATLAS_DATA/battery/twin1/If_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/If/B': {'path': '$ATLAS_DATA/battery/twin1/If_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 70.8, 'surplus': 2070, 'paired': 108}, 'twin1/IfExp/A': {'path': '$ATLAS_DATA/battery/twin1/IfExp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/IfExp/B': {'path': '$ATLAS_DATA/battery/twin1/IfExp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 35.3, 'surplus': 2160, 'paired': 116}, 'twin1/ImportFrom/A': {'path': '$ATLAS_DATA/battery/twin1/ImportFrom_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/ImportFrom/B': {'path': '$ATLAS_DATA/battery/twin1/ImportFrom_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 29.5, 'surplus': 1800, 'paired': 0}, 'twin1/List/A': {'path': '$ATLAS_DATA/battery/twin1/List_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/List/B': {'path': '$ATLAS_DATA/battery/twin1/List_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 34.65, 'surplus': 901, 'paired': 14}, 'twin1/ListComp/A': {'path': '$ATLAS_DATA/battery/twin1/ListComp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/ListComp/B': {'path': '$ATLAS_DATA/battery/twin1/ListComp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 106.0, 'surplus': 2070, 'paired': 109}, 'twin1/NamedExpr/A': {'path': '$ATLAS_DATA/battery/twin1/NamedExpr_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/NamedExpr/B': {'path': '$ATLAS_DATA/battery/twin1/NamedExpr_B.jsonl.gz', 'n': 630, 'valid': 630, 'offset_residual': 34.75, 'surplus': 630, 'paired': 0}, 'twin1/Pass/A': {'path': '$ATLAS_DATA/battery/twin1/Pass_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Pass/B': {'path': '$ATLAS_DATA/battery/twin1/Pass_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 38.4, 'surplus': 1800, 'paired': 0}, 'twin1/Return/A': {'path': '$ATLAS_DATA/battery/twin1/Return_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Return/B': {'path': '$ATLAS_DATA/battery/twin1/Return_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 68.25, 'surplus': 1800, 'paired': 0}, 'twin1/Set/A': {'path': '$ATLAS_DATA/battery/twin1/Set_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Set/B': {'path': '$ATLAS_DATA/battery/twin1/Set_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 37.5, 'surplus': 1081, 'paired': 62}, 'twin1/Slice/A': {'path': '$ATLAS_DATA/battery/twin1/Slice_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Slice/B': {'path': '$ATLAS_DATA/battery/twin1/Slice_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 146.15, 'surplus': 1495, 'paired': 174}, 'twin1/Starred/A': {'path': '$ATLAS_DATA/battery/twin1/Starred_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Starred/B': {'path': '$ATLAS_DATA/battery/twin1/Starred_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 40.1, 'surplus': 1189, 'paired': 0}, 'twin1/Subscript/A': {'path': '$ATLAS_DATA/battery/twin1/Subscript_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Subscript/B': {'path': '$ATLAS_DATA/battery/twin1/Subscript_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 39.9, 'surplus': 1137, 'paired': 55}, 'twin1/Tuple/A': {'path': '$ATLAS_DATA/battery/twin1/Tuple_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/Tuple/B': {'path': '$ATLAS_DATA/battery/twin1/Tuple_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 46.0, 'surplus': 1086, 'paired': 0}, 'twin1/UnaryOp/A': {'path': '$ATLAS_DATA/battery/twin1/UnaryOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/UnaryOp/B': {'path': '$ATLAS_DATA/battery/twin1/UnaryOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 42.0, 'surplus': 1800, 'paired': 0}, 'twin1/While/A': {'path': '$ATLAS_DATA/battery/twin1/While_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/While/B': {'path': '$ATLAS_DATA/battery/twin1/While_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 87.35, 'surplus': 1800, 'paired': 0}, 'twin1/With/A': {'path': '$ATLAS_DATA/battery/twin1/With_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin1/With/B': {'path': '$ATLAS_DATA/battery/twin1/With_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 59.65, 'surplus': 2160, 'paired': 114}, 'twin1/C': {'path': '$ATLAS_DATA/battery/twin1/C.jsonl.gz', 'n': 1000, 'tokens': 95421}, 'twin2/AnnAssign/A': {'path': '$ATLAS_DATA/battery/twin2/AnnAssign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/AnnAssign/B': {'path': '$ATLAS_DATA/battery/twin2/AnnAssign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 97.25, 'surplus': 1455, 'paired': 122}, 'twin2/Assert/A': {'path': '$ATLAS_DATA/battery/twin2/Assert_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Assert/B': {'path': '$ATLAS_DATA/battery/twin2/Assert_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 29.4, 'surplus': 1800, 'paired': 0}, 'twin2/Assign/A': {'path': '$ATLAS_DATA/battery/twin2/Assign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Assign/B': {'path': '$ATLAS_DATA/battery/twin2/Assign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 32.25, 'surplus': 936, 'paired': 0}, 'twin2/Attribute/A': {'path': '$ATLAS_DATA/battery/twin2/Attribute_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Attribute/B': {'path': '$ATLAS_DATA/battery/twin2/Attribute_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 55.65, 'surplus': 907, 'paired': 0}, 'twin2/AugAssign/A': {'path': '$ATLAS_DATA/battery/twin2/AugAssign_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/AugAssign/B': {'path': '$ATLAS_DATA/battery/twin2/AugAssign_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 46.25, 'surplus': 840, 'paired': 0}, 'twin2/BinOp/A': {'path': '$ATLAS_DATA/battery/twin2/BinOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/BinOp/B': {'path': '$ATLAS_DATA/battery/twin2/BinOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 61.1, 'surplus': 1120, 'paired': 0}, 'twin2/BoolOp/A': {'path': '$ATLAS_DATA/battery/twin2/BoolOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/BoolOp/B': {'path': '$ATLAS_DATA/battery/twin2/BoolOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 40.4, 'surplus': 1800, 'paired': 0}, 'twin2/Call/A': {'path': '$ATLAS_DATA/battery/twin2/Call_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Call/B': {'path': '$ATLAS_DATA/battery/twin2/Call_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 132.0, 'surplus': 1294, 'paired': 0}, 'twin2/Compare/A': {'path': '$ATLAS_DATA/battery/twin2/Compare_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Compare/B': {'path': '$ATLAS_DATA/battery/twin2/Compare_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 36.6, 'surplus': 1261, 'paired': 0}, 'twin2/Delete/A': {'path': '$ATLAS_DATA/battery/twin2/Delete_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Delete/B': {'path': '$ATLAS_DATA/battery/twin2/Delete_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 41.15, 'surplus': 1800, 'paired': 0}, 'twin2/Dict/A': {'path': '$ATLAS_DATA/battery/twin2/Dict_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Dict/B': {'path': '$ATLAS_DATA/battery/twin2/Dict_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 24.9, 'surplus': 1261, 'paired': 43}, 'twin2/For/A': {'path': '$ATLAS_DATA/battery/twin2/For_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/For/B': {'path': '$ATLAS_DATA/battery/twin2/For_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 62.9, 'surplus': 2250, 'paired': 175}, 'twin2/FunctionDef/A': {'path': '$ATLAS_DATA/battery/twin2/FunctionDef_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/FunctionDef/B': {'path': '$ATLAS_DATA/battery/twin2/FunctionDef_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 68.25, 'surplus': 2160, 'paired': 118}, 'twin2/GeneratorExp/A': {'path': '$ATLAS_DATA/battery/twin2/GeneratorExp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/GeneratorExp/B': {'path': '$ATLAS_DATA/battery/twin2/GeneratorExp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 111.8, 'surplus': 2250, 'paired': 163}, 'twin2/If/A': {'path': '$ATLAS_DATA/battery/twin2/If_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/If/B': {'path': '$ATLAS_DATA/battery/twin2/If_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 75.3, 'surplus': 2070, 'paired': 108}, 'twin2/IfExp/A': {'path': '$ATLAS_DATA/battery/twin2/IfExp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/IfExp/B': {'path': '$ATLAS_DATA/battery/twin2/IfExp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 37.3, 'surplus': 2160, 'paired': 112}, 'twin2/ImportFrom/A': {'path': '$ATLAS_DATA/battery/twin2/ImportFrom_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/ImportFrom/B': {'path': '$ATLAS_DATA/battery/twin2/ImportFrom_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 28.05, 'surplus': 1800, 'paired': 0}, 'twin2/List/A': {'path': '$ATLAS_DATA/battery/twin2/List_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/List/B': {'path': '$ATLAS_DATA/battery/twin2/List_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 32.9, 'surplus': 1147, 'paired': 110}, 'twin2/ListComp/A': {'path': '$ATLAS_DATA/battery/twin2/ListComp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/ListComp/B': {'path': '$ATLAS_DATA/battery/twin2/ListComp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 97.3, 'surplus': 2250, 'paired': 143}, 'twin2/NamedExpr/A': {'path': '$ATLAS_DATA/battery/twin2/NamedExpr_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/NamedExpr/B': {'path': '$ATLAS_DATA/battery/twin2/NamedExpr_B.jsonl.gz', 'n': 630, 'valid': 630, 'offset_residual': 51.5, 'surplus': 630, 'paired': 0}, 'twin2/Pass/A': {'path': '$ATLAS_DATA/battery/twin2/Pass_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Pass/B': {'path': '$ATLAS_DATA/battery/twin2/Pass_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 37.9, 'surplus': 1800, 'paired': 0}, 'twin2/Return/A': {'path': '$ATLAS_DATA/battery/twin2/Return_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Return/B': {'path': '$ATLAS_DATA/battery/twin2/Return_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 56.95, 'surplus': 1800, 'paired': 0}, 'twin2/Set/A': {'path': '$ATLAS_DATA/battery/twin2/Set_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Set/B': {'path': '$ATLAS_DATA/battery/twin2/Set_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 33.5, 'surplus': 1088, 'paired': 2}, 'twin2/Slice/A': {'path': '$ATLAS_DATA/battery/twin2/Slice_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Slice/B': {'path': '$ATLAS_DATA/battery/twin2/Slice_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 142.5, 'surplus': 1662, 'paired': 160}, 'twin2/Starred/A': {'path': '$ATLAS_DATA/battery/twin2/Starred_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Starred/B': {'path': '$ATLAS_DATA/battery/twin2/Starred_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 30.65, 'surplus': 1276, 'paired': 0}, 'twin2/Subscript/A': {'path': '$ATLAS_DATA/battery/twin2/Subscript_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Subscript/B': {'path': '$ATLAS_DATA/battery/twin2/Subscript_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 38.0, 'surplus': 1287, 'paired': 54}, 'twin2/Tuple/A': {'path': '$ATLAS_DATA/battery/twin2/Tuple_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/Tuple/B': {'path': '$ATLAS_DATA/battery/twin2/Tuple_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 40.9, 'surplus': 1175, 'paired': 0}, 'twin2/UnaryOp/A': {'path': '$ATLAS_DATA/battery/twin2/UnaryOp_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/UnaryOp/B': {'path': '$ATLAS_DATA/battery/twin2/UnaryOp_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 40.0, 'surplus': 1800, 'paired': 0}, 'twin2/While/A': {'path': '$ATLAS_DATA/battery/twin2/While_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/While/B': {'path': '$ATLAS_DATA/battery/twin2/While_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 77.0, 'surplus': 1800, 'paired': 0}, 'twin2/With/A': {'path': '$ATLAS_DATA/battery/twin2/With_A.jsonl.gz', 'n': 720, 'valid': 720}, 'twin2/With/B': {'path': '$ATLAS_DATA/battery/twin2/With_B.jsonl.gz', 'n': 720, 'valid': 720, 'offset_residual': 68.5, 'surplus': 2160, 'paired': 118}, 'twin2/C': {'path': '$ATLAS_DATA/battery/twin2/C.jsonl.gz', 'n': 1000, 'tokens': 96864}, 'alt/AnnAssign/A': {'path': '$ATLAS_DATA/battery/alt/AnnAssign_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Assert/A': {'path': '$ATLAS_DATA/battery/alt/Assert_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Assign/A': {'path': '$ATLAS_DATA/battery/alt/Assign_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Attribute/A': {'path': '$ATLAS_DATA/battery/alt/Attribute_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/AugAssign/A': {'path': '$ATLAS_DATA/battery/alt/AugAssign_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/BinOp/A': {'path': '$ATLAS_DATA/battery/alt/BinOp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/BoolOp/A': {'path': '$ATLAS_DATA/battery/alt/BoolOp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Call/A': {'path': '$ATLAS_DATA/battery/alt/Call_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Compare/A': {'path': '$ATLAS_DATA/battery/alt/Compare_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Delete/A': {'path': '$ATLAS_DATA/battery/alt/Delete_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Dict/A': {'path': '$ATLAS_DATA/battery/alt/Dict_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/For/A': {'path': '$ATLAS_DATA/battery/alt/For_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/FunctionDef/A': {'path': '$ATLAS_DATA/battery/alt/FunctionDef_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/GeneratorExp/A': {'path': '$ATLAS_DATA/battery/alt/GeneratorExp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/If/A': {'path': '$ATLAS_DATA/battery/alt/If_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/IfExp/A': {'path': '$ATLAS_DATA/battery/alt/IfExp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/ImportFrom/A': {'path': '$ATLAS_DATA/battery/alt/ImportFrom_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/List/A': {'path': '$ATLAS_DATA/battery/alt/List_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/ListComp/A': {'path': '$ATLAS_DATA/battery/alt/ListComp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/NamedExpr/A': {'path': '$ATLAS_DATA/battery/alt/NamedExpr_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Pass/A': {'path': '$ATLAS_DATA/battery/alt/Pass_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Return/A': {'path': '$ATLAS_DATA/battery/alt/Return_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Set/A': {'path': '$ATLAS_DATA/battery/alt/Set_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Slice/A': {'path': '$ATLAS_DATA/battery/alt/Slice_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Starred/A': {'path': '$ATLAS_DATA/battery/alt/Starred_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Subscript/A': {'path': '$ATLAS_DATA/battery/alt/Subscript_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/Tuple/A': {'path': '$ATLAS_DATA/battery/alt/Tuple_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/UnaryOp/A': {'path': '$ATLAS_DATA/battery/alt/UnaryOp_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/While/A': {'path': '$ATLAS_DATA/battery/alt/While_A.jsonl.gz', 'n': 60, 'valid': 60}, 'alt/With/A': {'path': '$ATLAS_DATA/battery/alt/With_A.jsonl.gz', 'n': 60, 'valid': 60}}** sets, 720 per set in the validation subset, twins [1, 2], wall 232 s | |
| - cells **90**, cores per construct **8** | |
| - **Seeding defect, recorded as caught.** Seeds were derived with `abs(hash((twin, core_id, cell_key))) % 2**31`. Python salts `hash()` for strings per process, so every build produced a different battery and the seeds recorded here could never regenerate it — verified by staging a rebuild: **152 of 152 sets differed, none identical**. `profilebattery/generate/seeds.py` replaces it with a keyed blake2b digest for future builds, with a cross-process regression test. **This battery cannot be regenerated; its provenance is the per-set content hash below.** The phase battery is generated under the fixed seeding, with each twin drawing its own cores. | |
| - **Twins share all eight cores**, contrary to the spec's "each twin draws its own eight from a space of ≥24". Twin disagreement therefore carries no core-sampling variance: the twin unit is a draw-and-yardstick unit, and every figure quoted in twin units inherits that. | |
| - access-form labels record the **realised** level, not the nominal one (`self_attribute` degrades outside a method); relabelled in place, each set verified text-identical. | |
| ### Per-set content hashes (212 sets) | |
| `battery/CONTENT_HASHES.json` holds a sha256 over each set's prompt texts in order. A set is re-recorded only if its content hash changed. `run_sweep.py` stamps `prompt_sha256` into every record MANIFEST and refuses to run without the hash file. Manifest of hashes: `9bb00f2b78bece7c` | |
| ## Record | |
| - primary hook: `model.layers.{i}.mlp.down_proj` **forward pre-hook, `args[0]`** — the gated product act(W_gate·x) ⊙ (W_up·x), 6,144 wide. Verified against recomputation at max abs difference **0.0**. | |
| - secondary: `.mlp` output (m_k, 2,240 — the v1 basis); residual stream at the same positions. | |
| - sites: construct token (primary), span end, last token. C read at every position, labelled by (relative-position band × token class), 20 cells. | |
| - dtype: per-prompt fp16, accumulation fp32, model bfloat16 on MPS. | |
| | reference | tokens (B) | commit | sets | wall (h) | fp16 overflow | per-prompt | hashes | | |
| |---|---|---|---|---|---|---|---| | |
| **fp16 overflow is zero at every reference, layer and site.** No layer needed the fp32 fallback. | |
| **The untrained reference's weights do not regenerate.** It is not a checkpoint but a null -- the architecture instantiated from the `pretrain_iter_0008000` config under `torch.manual_seed(20260920)` (`MANIFEST.json: untrained_seed`), recorded on **cuda** in **bfloat16**. Re-instantiating it under the same seed on torch 2.14.0 / transformers 5.17.0 gives different weights: the re-run reproduces a *trained* checkpoint's recorded accumulators to cosine 0.999997 and the untrained one to **cosine 0.004** -- orthogonal. Module initialisation draws a different number of values across library versions, so the seed alone does not pin the weights. **The torch, transformers and Python versions of the recording run were not stored** -- no manifest, lockfile or environment record in the repository carries them -- so the untrained reference cannot be re-recorded, and any pass needing new statistics from it is limited to what is already in the record. Future runs should store the interpreter and library versions beside `dtype` and `device`. | |
| **The local `record/` tree at the eight S1 references is a stale S1-era recording and is kept, untouched, behind a guard** (A50; the decision of 19 Oct). It carries three read sites, 25 B cores, 51,840 summed positions and no `hs|` scope keys, where the dataset's files at the same references carry one site, 21 B cores, 17,280 positions and 4,800 `hs|` keys; `RecordStore.path()` prefers the local copy, so any pass that reads accumulators would silently get the old recording at those eight references and the new one everywhere else — which is how A48 came to record a C0.5 departure that half exists. Every accumulator and yardstick read in the Paper-1 code goes through `experiments/paper1.py:dataset_record()`, which reads the dataset and ignores the local copy. | |
| Storage rule (DECIDED 6): per-prompt arrays at `pretrain_iter_0003000` and the untrained reference; per-core, per-cell accumulators everywhere. The accumulator route reproduces the per-prompt profile to cosine **1.000000**, max|diff| **0.0** — asserted by `test_accumulator_route_equals_the_per_prompt_route`. | |
| ## Profiles | |
| - yardstick: **24** layers × **20** cells, rebuilt at every checkpoint | |
| - ε (smooth floor) = 0.01, grid {0.003, 0.01, 0.03}. **ε does no work**: worst displacement across the grid is **0.0008 twin units**, three orders below twin disagreement. | |
| - reported SE is the **between-core** spread of core means / √cores (`Profile.se`); the per-prompt SE is kept alongside as `se_prompt`. | |
| ## Validation | |
| | reference | checks run | median twin cosine | | |
| |---|---|---| | |
| | `untrained` | 7/10 | 0.680 | | |
| | `pretrain_iter_0001000` | 5/10 | 0.981 | | |
| | `pretrain_iter_0002000` | 5/10 | 0.984 | | |
| | `pretrain_iter_0003000` | 9/10 | 0.986 | | |
| | `pretrain_iter_0005000` | 6/10 | 0.984 | | |
| | `pretrain_iter_0010000` | 6/10 | 0.983 | | |
| | `pretrain_iter_0015000` | 6/10 | 0.981 | | |
| | `pretrain_iter_0020000` | 6/10 | 0.980 | | |
| Checks 2, 3, 7a and 7b require per-prompt arrays and are named NOT RUN with that reason at the accumulator-only references. Check 5 (Qwen2.5-Coder) is NOT RUN and deferred past the freeze by decision, not by failure. | |
| ### Licensed n — a decision, with the rule that produced it | |
| **120 per set for the deep tier, 60 for shallow.** | |
| Rule: *the smallest grid n at which the **raw** cosine of the subsample profile to the full-set profile reaches 0.99 for every construct in the tier, read at `pretrain_iter_0003000`*. Measured on the grid {8, 24, 56, 120, 176}: 0/30 constructs at n=8, 13/30 at 24, 25/30 at 56, **30/30 at 120** with the minimum 0.9904 (`Assign`). | |
| Two things this rule is **not**, both recorded so the choice is reviewable: | |
| 1. It is not stated on the *disattenuated* cosine. That statistic is saturated here — 133 of 150 grid cells sit at or above 0.9999, and 22 of 30 constructs are already there at n=8 — so a 0.999 criterion on it would license **n=8** for most of the tier and yields 120 only because `FunctionDef` and `If` are the last to reach the ceiling. The number would be set by where the cap bites rather than by convergence. | |
| 2. It is not report rule 9.2 as originally written, *"within twin noise of the full set"*. Twin cosine varies **0.811–0.998** by construct for reasons unrelated to sample size, so 9.2 licenses **n=8 for `Attribute`** (twin cosine 0.832) and **n=176 for `For`** (0.998). Rule 9.2 is superseded; the report and method handout carry the replacement. | |
| **The shallow tier's 60 is a declared relaxation, not this rule applied to a different tier.** It is not a grid point, and at the nearest one (n=56) five of thirty constructs fall below 0.99, `Assign` lowest at 0.9791. It is a budget decision and is recorded as one. | |
| **The untrained reference cannot license n.** Its raw cosine at n=120 is **0.618** median (range 0.575–0.806): a random-feature regime with no structure to converge to. | |
| ### Interval calibration | |
| Coverage is **bimodal** and stably so, against a nominal 0.95. At the three trained checkpoints the nine verdict constructs run 0.848–0.872 and the other twenty-one 0.990–0.993; `FunctionDef` needs ×2.03. **A single factor is wrong in both directions.** Per-construct factors are published in `validation/calibration/COVERAGE_FACTORS.json` and never applied silently. | |
| The cause is the **yardstick's own estimation error**, not core clustering: holding the prompt data fixed and swapping only the twin's yardstick moves the profile 0.474 against a full twin gap of 0.940 — a quarter of the twin variance, invisible to any within-twin SE. | |
| ## Go/no-go verdict | |
| **PASS — 9 of 9**, read at `pretrain_iter_0003000`, two-thirds bar. Verdict constructs are the keyword-token constructs that also pass the read-site acceptance rule (overlap ≥0.80, minimum over twins): `Assert`, `Delete`, `For`, `FunctionDef`, `If`, `ImportFrom`, `Pass`, `While`, `With`. | |
| Excluded by the read-site rule: `BoolOp` 0.515, `GeneratorExp` 0.597, `IfExp` 0.613, `ListComp` 0.569, `Return` 0.514, `UnaryOp` 0.25. | |
| Turnover, for G4: top-5% unit-set Jaccard rises 0.176 → 0.396 across six steps against a **twin ceiling of 0.595** (pilot-era: the median over the 30 pilot constructs at layer 22; the pilot recording, not the phase record — A52), and the last three steps span 21B tokens each against the first two's 4.19B while turning over *less*. A date criterion of the form "the set stops turning over" must be stated as a fraction of that ceiling; against an absolute threshold near 1 it cannot fire. | |
| ## Deliverable hashes | |
| | artefact | sha256 | | |
| |---|---| | |
| | `inventory/inventory_v2.json` | `944521b421a6f11b` | | |
| | `calibration/MANIFEST.json` | `1fe92304c4151f6a` | | |
| | `calibration/tier.json` | `4b2ed797d581b722` | | |
| | `battery/CONTENT_HASHES.json` | `9bb00f2b78bece7c` | | |
| | `battery/MANIFEST.json` | `5d49cfc5dc363518` | | |
| | `validation/go-no-go/VERDICT.json` | `47e4c6c47a5d3df2` | | |
| | `validation/ceiling/JACCARD_CEILING.json` | `da867963a0f1ca72` — pilot-derived; valid for the pilot-era battery reports, not phase data (A52) | | |
| | `validation/calibration/COVERAGE_FACTORS.json` | `7271a4d16848e770` | | |
| | `validation/length/INVARIANCE.json` | `fb988b6cb5d6ba8b` | | |
| | `validation/read-site/OVERLAP.json` | `6c81ade619e1fe9a` | | |
| Code: GitHub `<the working repository>` (renamed from `ProfileBattery_private`, 3 Oct 2026) at `7ade464ea11a`. Record: HuggingFace dataset `<the record repository>` -- the same name on a different platform, so the platform is named here and not left to the reader. PROVENANCE hashes file contents, never commit SHAs. | |
| ## Known limitations carried into the freeze | |
| 1. The v1 bridge is **not derived**. v1's mask rule is a per-prompt consistency criterion and the record keeps per-prompt vectors for one site and one basis only, so v1's "72% of circuits wired by 8.4B" has not been re-derived here. | |
| 2. The nine verdict constructs are 55–73% length-flagged against 0.2% elsewhere. Only **2.3%** of the concept profile lies along the length axis at the median (max 19.7%) and the top-5% set is not enriched for length-sensitive units (0.93×), so the concept direction is nearly orthogonal to the length direction — but a claim about *which units* carry these constructs is not licensed. | |
| 3. The ν-sensitivity alternative arm has n=60, reliability 0.187, so its disattenuated cosine goes over unity and is reported as raw distance. | |
| 4. Checks 5 and, at four references, 2/3/7a/7b are NOT RUN, each with its reason in the licence table. | |
| ## The pilot recording at the phase recording's path (A52, A53) | |
| *Reserved for the Settle chat. The register entries are A52 and A53 in `8-Settle/2609-30Sep_DEVIATIONS-CC.md`: the 2x2 that localises the damage, the thirteen affected constructs, the seven references, the figures that moved (92 -> 90 dated, 145 -> 148 falling, P7 91/92 -> 90/90), and the verdict table over every reader that preferred a local tree. The exclusion list the repo build enforces is `experiments/repo_manifest.py`; it fails the build rather than warning.* | |