docs/experiments/EXP-3D-RGBD-DATA-SCALE-V14/audit/claude_v14_derive.py
"""Derive the V14 pipeline from the frozen V12 files (sources never modified).
Usage (repo root of a dry copy): python3 v14_derive.py. The EVAL stage script, the report and
the tests are written separately; the prepare, finalize and runner are derived here.
"""
from pathlib import Path
def read(path):
with open(path, encoding="utf-8", newline="") as handle:
return handle.read()
def write(target, text):
path = Path(target)
if path.exists():
raise SystemExit(f"DERIVE FAILED {target}: target exists")
with open(path, "w", encoding="utf-8", newline="") as handle:
handle.write(text)
print("derived", target)
def apply(text, target, replacements):
for old, new, count in replacements:
found = text.count(old)
if found != count:
raise SystemExit(f"DERIVE FAILED {target}: expected {count}, found {found}: {old[:80]!r}")
text = text.replace(old, new)
return text
def derive(source, target, replacements, leftovers=()):
text = apply(read(source), target, replacements)
for word in leftovers:
if word in text:
raise SystemExit(f"DERIVE FAILED {target}: leftover {word!r}")
write(target, text)
LEFT = ("rgbd_width", "WIDTH-V12", '"""V12', "width64", "WIDTH64", "width 64", "V12 ")
CARRIER = (
"from mcss.mechanism_pilot.rgbd_width_carrier import ",
"from mcss.mechanism_pilot.rgbd_data_scale_carrier import ",
)
EVALUATION = (
"from mcss.mechanism_pilot.rgbd_width_evaluation import evaluate_model\n",
"from mcss.mechanism_pilot.rgbd_data_scale_evaluation import evaluate_model\n",
)
STREAM_COMMENT = (
" # V12 widths differ by design, so only the data/ray stream must match per seed.\n",
" # V14 TRAIN sets differ by design; variants sharing a TRAIN set share the stream.\n",
)
TRAIN_SETS = (
' for train_set in ("TRAIN24", "TRAIN72")\n',
' for train_set in ("TRAIN24", "TRAIN72", "TRAIN_EXT")\n',
)
# ---------------------------------------------------------------- carrier module
derive(
"src/mcss/mechanism_pilot/rgbd_width_carrier.py",
"src/mcss/mechanism_pilot/rgbd_data_scale_carrier.py",
[
(
'"""V12: carrier width 64 for completion, on the V11 C1 recipe fixed before any V11 result.\n'
"\n"
"PLAN_BEFORE_V11_RESULTS.md runs V12 only if the V11 EVAL-V3 COMPLETION_GAIN point estimate is\n"
"> 0. C0 retrains the V11 C1 recipe at width 32 (a machinery control that must equal V11 C1\n"
"bit-for-bit); C1 is the same recipe (TRAIN72, 32^3, CONTEXT_DEPTH bounds, the frozen V2 C1\n"
"loss, 6000 steps, the V11 extra-view rule) with hidden_dim 64 and expansion_dim 128. The primary\n"
"EVAL-V4 control is V11 C1's own selected checkpoints, loaded with the V11 loader. Every module\n"
"width follows CarrierConfig, so the V8 computation, state construction and loss are reused;\n"
"only the frozen width guard is generalized to the two V12 widths.\n"
'"""\n',
'"""V14: more training scenes for completion, on the unchanged V11 C1 recipe (data scale).\n'
"\n"
"PLAN_BEFORE_TRAIN_EXT_DATA.md fixed V14 before any new training scene was read. C0 retrains\n"
"the V11 C1 recipe on TRAIN72 (a machinery control that must equal V11 C1 bit-for-bit); C1 is\n"
"the same recipe (width 32, 32^3, CONTEXT_DEPTH bounds, the frozen V2 C1 loss, 6000 steps, the\n"
"V11 extra-view rule) trained on TRAIN72 plus the TRAIN-EXT scenes. Both variants share the\n"
"V11 C1 parameter draw; only the variant's frozen TRAIN set differs. V8's train_records looks\n"
"the TRAIN set up in the V8 variant table (TRAIN72 for every V8 variant), so the variant's own\n"
"TRAIN set is resolved here instead.\n"
'"""\n',
1,
),
(
'EXPERIMENT = "EXP-3D-RGBD-WIDTH-V12"\nSCHEMA = "mcss.rgbd_width_carrier.v12"\n',
'EXPERIMENT = "EXP-3D-RGBD-DATA-SCALE-V14"\nSCHEMA = "mcss.rgbd_data_scale_carrier.v14"\n',
1,
),
(
' "C0": {"label": "WIDTH32", "hidden": 32, "expansion": 64, "model": "C1", "train": "TRAIN72"},\n'
' "C1": {"label": "WIDTH64", "hidden": 64, "expansion": 128, "model": "C1", "train": "TRAIN72"},\n',
' "C0": {"label": "TRAIN72", "hidden": 32, "expansion": 64, "model": "C1", "train": "TRAIN72"},\n'
' "C1": {\n'
' "label": "TRAIN_EXT",\n'
' "hidden": 32,\n'
' "expansion": 64,\n'
' "model": "C1",\n'
' "train": "TRAIN_EXT",\n'
" },\n",
1,
),
("train_records = v8.train_records\n", "", 1),
(
"def carrier_config(variant):\n",
"def train_records(manifest, config, variant):\n"
' """Exactly the variant\'s own frozen TRAIN set (V14 table), sorted by scene id."""\n'
' names = set(config["train_sets"][VARIANT_SPECS[variant]["train"]])\n'
" records = sorted(\n"
' (r for r in manifest["scenes"] if r["split"] == "TRAIN" and r["scene_id"] in names),\n'
' key=lambda r: r["scene_id"],\n'
" )\n"
" if len(records) != len(names) or len(records) < 24:\n"
" raise PermissionError(\"The variant's frozen TRAIN set is incomplete\")\n"
" return records\n"
"\n"
"\n"
"def carrier_config(variant):\n",
1,
),
(
' raise ValueError("A registered V12 width at 32^3 (one candidate per voxel) is required")\n',
' raise ValueError("The registered V14 width at 32^3 (one candidate per voxel) is required")\n',
1,
),
(
" \"\"\"The V8 resolution carrier at the variant's width; at width 32 this is the V11 C1 draw.\"\"\"\n",
' """The V8 resolution carrier at width 32; for both V14 variants this is the V11 C1 draw."""\n',
1,
),
(
' """The V8 build_state with the width guard generalized to the registered V12 widths."""\n',
' """The V8 build_state under the width guard of the registered V14 width."""\n',
1,
),
(
'raise TypeError("V12 states are built from an RGBDContext")',
'raise TypeError("V14 states are built from an RGBDContext")',
1,
),
(
'raise PermissionError("V12 static RGB-D width carrier schema required")',
'raise PermissionError("V14 static RGB-D data-scale carrier schema required")',
1,
),
(
'raise PermissionError("V12 checkpoints always carry the measured-depth bypass")',
'raise PermissionError("V14 checkpoints always carry the measured-depth bypass")',
1,
),
],
(*LEFT, "V12", "v8.train_records"),
)
# ---------------------------------------------------------------- DEV evaluation module
derive(
"src/mcss/mechanism_pilot/rgbd_width_evaluation.py",
"src/mcss/mechanism_pilot/rgbd_data_scale_evaluation.py",
[
(
'"""V12 DEV evaluation: the V11 Amendment 1 evaluator (float-rounding RGB clip), V12 widths.\n',
'"""V14 DEV evaluation: the V11 Amendment 1 evaluator; only diagnostics runs save states.\n',
1,
),
(CARRIER[0] + "build_state, scene_data\n", CARRIER[1] + "build_state, scene_data\n", 1),
(
" out.mkdir(parents=True)\n"
' torch.save({key: barrier.state(key) for key in keys}, out / "states.pt")\n',
" out.mkdir(parents=True)\n"
" if diagnostics: # V14: only the selected-checkpoint evaluation keeps states (disk)\n"
' torch.save({key: barrier.state(key) for key in keys}, out / "states.pt")\n',
1,
),
(
' "states_file_sha256": sha(out / "states.pt"),\n',
' "states_file_sha256": sha(out / "states.pt") if diagnostics else None,\n',
1,
),
],
(*LEFT, "V12"),
)
# ---------------------------------------------------------------- training
derive(
"scripts/train_rgbd_width_carrier.py",
"scripts/train_rgbd_data_scale_carrier.py",
[
(
'"""V12 carrier-width runs of the V11 C1 recipe; no writer or test cohort access."""',
'"""V14 data-scale runs of the V11 C1 recipe; no writer or test cohort access."""',
1,
),
(CARRIER[0] + "(\n", CARRIER[1] + "(\n", 1),
(" " + EVALUATION[0], " " + EVALUATION[1], 1),
],
(*LEFT, "V12"),
)
# ---------------------------------------------------------------- selected DEV evaluation
derive(
"scripts/evaluate_rgbd_width_carrier.py",
"scripts/evaluate_rgbd_data_scale_carrier.py",
[
(
'"""V12: evaluate frozen per-seed DEV selections and derive post-seal diagnostics."""',
'"""V14: evaluate frozen per-seed DEV selections and derive post-seal diagnostics."""',
1,
),
(CARRIER[0] + "VARIANT_SPECS, load_checkpoint\n", CARRIER[1] + "VARIANT_SPECS, load_checkpoint\n", 1),
(EVALUATION[0], EVALUATION[1], 1),
(*STREAM_COMMENT, 1),
(*TRAIN_SETS, 1),
(
' "C0": "width32_training",\n "C1": "width64_training",\n',
' "C0": "train72_training",\n "C1": "train_ext_training",\n',
1,
),
],
(*LEFT, "V12"),
)
derive(
"scripts/diagnose_rgbd_width_states.py",
"scripts/diagnose_rgbd_data_scale_states.py",
[
(
'"""Preregistered descriptive state-correlation diagnostic on sealed V12 DEV states.',
'"""Preregistered descriptive state-correlation diagnostic on sealed V14 DEV states.',
1,
)
],
(*LEFT, "V12"),
)
derive(
"scripts/analyze_rgbd_width_carrier.py",
"scripts/analyze_rgbd_data_scale_carrier.py",
[
(
'"""V12 raw-only DEV statistics (secondary) with the frozen V2 estimator and figures."""',
'"""V14 raw-only DEV statistics (secondary) with the frozen V2 estimator and figures."""',
1,
),
(
' "C0": "C0 width 32",\n "C1": "C1 width 64 (primary)",\n',
' "C0": "C0 TRAIN72",\n "C1": "C1 TRAIN72+EXT (primary)",\n',
1,
),
],
(*LEFT, "V12"),
)
derive(
"scripts/audit_rgbd_width_statistics.py",
"scripts/audit_rgbd_data_scale_statistics.py",
[
('"analyze_rgbd_width_carrier.py"', '"analyze_rgbd_data_scale_carrier.py"', 1),
('"rgbd_width_reproduction_analyzer"', '"rgbd_data_scale_reproduction_analyzer"', 1),
('prefix="rgbd-width-statistics-audit-"', 'prefix="rgbd-data-scale-statistics-audit-"', 1),
],
(*LEFT, "V12", "rgbd-width"),
)
derive(
"scripts/reference_rgbd_width_carrier.py",
"scripts/reference_rgbd_data_scale_carrier.py",
[
(
'"""V12 preregistered geometry-free reference for the sealed selected DEV predictions."""',
'"""V14 preregistered geometry-free reference for the sealed selected DEV predictions."""',
1,
),
(
' "V12": {\n'
' "labels": {"C0": "WIDTH32", "C1": "WIDTH64"},\n'
' "primary": "WIDTH_GAIN_DEV",\n',
' "V14": {\n'
' "labels": {"C0": "TRAIN72", "C1": "TRAIN_EXT"},\n'
' "primary": "DATA_GAIN_DEV",\n',
1,
),
(' "source": "V12",\n', ' "source": "V14",\n', 1),
],
(*LEFT, "V12", "WIDTH"),
)
exec(Path(__file__).with_name("v14_derive_part2.py").read_text())