scripts/run_rgbd_stream_ttt_fresh_v7.py
#!/usr/bin/env python3
"""Freeze, run and analyze Core B V7 (stream and context-only TTT on the fresh EVAL-V5, GPU 4).
--stage freeze writes the contracts and the preregistration (CPU, CUDA hidden);
--stage run on GPU 4 through cv_gpu_run.sh: the sealed evaluation of every seed, an
independent recomputation of every metric, the normalized readout, the NEAR/FAR
split with model-free masks, the analysis and the README.
Run from the repository root under cv_host_run.py (records store zonghan_o paths).
"""
from __future__ import annotations
import argparse
import datetime
import importlib.util
import json
import os
import subprocess
import sys
from pathlib import Path
import numpy as np
import torch
from PIL import Image
from mcss.dynamic.types import hash_scene_state
from mcss.mechanism_pilot.direct_capacity_statistics import describe, paired
from mcss.mechanism_pilot.geometry_carrier_experiment import read, verify_data
from mcss.mechanism_pilot.rgbd_stream_evaluation_v3 import preseal_depth_audit
from mcss.mechanism_pilot.rgbd_stream_ttt_context_v6 import cohort_records
from mcss.mechanism_pilot.rgbd_stream_ttt_eval_v5 import eval_context
from mcss.mechanism_pilot.rgbd_stream_ttt_v4 import (
RAYS_PER_FRAME,
STREAM_LENGTH,
load_frozen_carrier,
stream_frame_ids,
)
from mcss.mechanism_pilot.rgbd_stream_ttt_v7 import (
CONTROLS,
EXPERIMENT,
LEARNING_RATE,
POLICY_SPECS,
evaluate_v7,
)
from mcss.mechanism_pilot.small_training import sha, write_json
from mcss.mechanism_pilot.statistics import measurement_metrics
COREB_V4 = Path("outputs/EXP-3D-RGBD-DYNAMIC-TTT-V4")
COREB_V6 = Path("outputs/EXP-3D-RGBD-DYNAMIC-TTT-CONTEXT-V6")
V7_SPLIT = Path("outputs/EXP-3D-RGBD-DEPTH-BOUNDS-V7")
V17 = Path("outputs/EXP-3D-RGBD-LONG-TRAIN-V17")
V18 = Path("outputs/EXP-3D-RGBD-LONG-TRAIN-DATA-V18")
EVAL_V5 = Path("outputs/EXP-3D-RGBD-EVAL-V5-DATA")
PLAN = "PLAN_BEFORE_COREB_V7.md"
PLAN_SHA256 = "d48a9b2868a59ec1ae2cc1a1aad3da11ce03c1376dbb007dfdebe36295a848dd"
CARRIER_STEP = 24000
COHORT = "EVAL_V5"
EXPECTED_SPLIT = {COHORT: "EVAL_V5"}
SPECS = dict(POLICY_SPECS)
GPU = "4"
LAUNCHER = Path.home() / "cv_tools/cv_host_run.py"
THREADS = {
"OMP_NUM_THREADS": "1",
"OPENBLAS_NUM_THREADS": "1",
"MKL_NUM_THREADS": "1",
"ATEN_CPU_CAPABILITY": "avx2",
"MKL_CBWR": "AVX2",
"DNNL_MAX_CPU_ISA": "AVX2",
"NVIDIA_TF32_OVERRIDE": "0",
}
CONTRASTS = {
"CONTEXT_TTT_100": ("NO_STREAM", "CTX_TTT_100"),
"STREAM_TTT_100": ("OFF", "TTT_100"),
"STREAM_TTT": ("OFF", "TTT"),
"CONTEXT_TTT": ("NO_STREAM", "CTX_TTT"),
"CONTEXT_DOSE": ("CTX_TTT", "CTX_TTT_100"),
"STREAM_DOSE": ("TTT", "TTT_100"),
"SUPERVISION_SOURCE": ("OFF_CTX_TTT", "TTT"),
"SUPERVISION_SOURCE_100": ("OFF_CTX_TTT_100", "TTT_100"),
"TTT_SPECIFICITY": ("TTT_WRONG_SCENE", "TTT"),
"TTT_100_SPECIFICITY": ("TTT_100_WRONG_SCENE", "TTT_100"),
"CTX_TTT_SPECIFICITY": ("CTX_TTT_WRONG_SCENE", "CTX_TTT"),
"CTX_TTT_100_SPECIFICITY": ("CTX_TTT_100_WRONG_SCENE", "CTX_TTT_100"),
"STREAM_CACHE_EFFECT": ("NO_STREAM", "OFF"),
}
PRIMARY = {
"P1_CONTEXT_TTT_100_STATUS": "CONTEXT_TTT_100",
"P2_STREAM_TTT_100_STATUS": "STREAM_TTT_100",
}
DRAWS, BOOT_SEED = 10000, 20260928
METRICS = ("rgb_mse", "depth_absrel", "depth_rmse", "depth_delta1", "opacity", "coverage")
KEYS = ("seed", "cohort", "scene_id", "role", "query_id", "policy")
SCRIPTS = Path("scripts")
def module(name, filename):
spec = importlib.util.spec_from_file_location(name, SCRIPTS / filename)
loaded = importlib.util.module_from_spec(spec)
spec.loader.exec_module(loaded)
return loaded
def v3_analysis():
"""Core B V3's scene aggregation and frozen V2 gate (as in Core B V4-V6)."""
return module("coreb_v3_analysis_v7", "analyze_rgbd_stream_write_v3.py")
def resolve(path, base):
path = Path(path)
return path if path.is_absolute() else Path(base) / path
def load_cohorts(root):
"""{cohort: (manifest, base)} from the frozen cohort lock, the manifest hash-checked."""
item = read(Path(root) / "cohort_lock.json")[COHORT]
if sha(Path(item["manifest"])) != item["manifest_sha256"]:
raise PermissionError(f"{COHORT} manifest changed")
return {COHORT: (read(Path(item["manifest"])), Path(item["manifest"]).parent)}
def verify_lock_v7(root):
"""The preregistration's input and source hashes (verify_lock without a TRAIN/DEV split)."""
lock = read(Path(root) / "preregistration.json")
if lock["status"] != "FROZEN_BEFORE_FORMAL_TRAINING":
raise PermissionError("Formal preregistration absent")
for section in ("input_sha256", "source_sha256"):
for name, digest in lock[section].items():
if sha(Path(name)) != digest:
raise PermissionError(f"Frozen input/source changed: {name}")
return read(Path(root) / "evaluation_contract.json")
def seal_previous(root):
files = {}
for base in (Path("outputs"), Path("docs/experiments")):
for path in sorted(base.rglob("*")):
if path.is_file() and EXPERIMENT not in path.parts:
files[str(path)] = {"sha256": sha(path), "bytes": path.stat().st_size}
write_json(
root / "audit/previous_experiment_seal.json",
{"created_utc": datetime.datetime.now(datetime.UTC).isoformat(), "files": files},
)
return len(files)
def used_scene_ids():
"""Every scene of TRAIN72, DEV, the holdout, EVAL-V3/V4, FRESH-V2 and TRAIN-EXT."""
split = read(V7_SPLIT / "scene_split.json")
used = {r["scene_id"] for r in split["scenes"]} | set(split.get("protected_scene_ids", []))
cohorts = read(V17 / "training_contract.json")["eval"]["cohorts"]
for name in ("EVAL_V3", "EVAL_V4", "FRESH_V2"):
used |= {r["scene_id"] for r in read(Path(cohorts[name]["manifest"]))["scenes"]}
train_ext = read(V18 / "training_contract.json")["train_ext"]
used |= {r["scene_id"] for r in read(Path(train_ext["manifest"]))["scenes"]}
return used
def freeze(root):
root = Path(root).resolve()
if root.name != EXPERIMENT or (root / "preregistration.json").exists():
raise PermissionError("Exact, unfrozen Core B V7 directory required")
if not (root / "PROTOCOL.md").is_file() or sha(root / PLAN) != PLAN_SHA256:
raise PermissionError("The Core B V7 plan and protocol must be in place")
for done in (COREB_V4, COREB_V6):
if read(done / "integrity.json")["status"] != "PASS":
raise PermissionError(f"{done.name} must be finalized")
selected = read(COREB_V4 / "selection.json")["settings"]
if selected != {"steps": POLICY_SPECS["TTT"][2], "learning_rate": LEARNING_RATE}:
raise PermissionError("TTT must use the setting Core B V4 selected on TRAIN")
if sha(LAUNCHER) != sha(Path("../_archives_cv/cv_host_run.py")):
raise PermissionError("The path-mapping launcher must equal the migrated one")
carriers = read(COREB_V4 / "carrier_lock.json")
for seed, item in carriers["carriers"].items():
if sha(Path(item["path"])) != item["sha256"] or item["step"] != CARRIER_STEP:
raise PermissionError(f"V17 step-{CARRIER_STEP} carrier changed: {seed}")
load_frozen_carrier(item["path"], "cpu")
write_json(root / "carrier_lock.json", carriers)
integrity = read(EVAL_V5 / "preparation_integrity.json")
manifest_path = EVAL_V5 / "manifest_eval_v5.json"
if integrity["status"] != "PASS" or read(manifest_path)["candidate_lock_sha256"] != sha(
EVAL_V5 / "candidate_lock.json"
):
raise PermissionError("EVAL-V5 must be a completed PASS preparation of its lock")
lock = {
COHORT: {"manifest": str(manifest_path.resolve()), "manifest_sha256": sha(manifest_path)}
}
write_json(root / "cohort_lock.json", lock)
entries = cohort_records(load_cohorts(root), EXPECTED_SPLIT)
if used_scene_ids() & set(entries):
raise PermissionError("EVAL-V5 must be disjoint from every earlier cohort")
episodes = {}
for sid, (_, record, _, _) in sorted(entries.items()):
for role in ("A", "B"):
ids = stream_frame_ids(record, role)
if ids is None:
raise PermissionError(f"{sid}/{role} lacks a full stream")
episodes[f"{sid}/{role}"] = list(ids)
write_json(
root / "stream_roles.json",
{
"rule": "Core B V1-V6 frame-id rule",
"stream_length": STREAM_LENGTH,
"episodes": episodes,
},
)
config = {
"experiment": EXPERIMENT,
"seeds": sorted(int(s) for s in carriers["carriers"]),
"policies": {name: list(spec) for name, spec in SPECS.items()},
"controls": CONTROLS,
"learning_rate": LEARNING_RATE,
"settings_source": "10 steps: Core B V4 selection (TRAIN); 100 steps: V23 dose arms on "
"EVAL-V3/V4 (descriptive); no selection on EVAL-V5",
"rays_per_supervised_frame": RAYS_PER_FRAME,
"stream_length": STREAM_LENGTH,
"cohort": {
"name": COHORT,
"scenes": len(entries),
"expected_split": EXPECTED_SPLIT[COHORT],
},
"near_masks": "V11 geometry + INPAINT-V1 near_mask from the role's context depth",
"num_threads": 1,
"thread_environment": THREADS,
"cpu_affinity": list(range(56, 64)),
"device": "cuda",
"gpu": GPU,
"host_launcher": {"path": str(LAUNCHER), "sha256": sha(LAUNCHER)},
}
write_json(root / "evaluation_contract.json", config)
write_json(
root / "decision_rules.json",
{
"aggregation": "per seed: scene mean of role means of query AbsRel; seeds equally "
"weighted",
"bootstrap": {"unit": "scene", "draws": DRAWS, "seed": BOOT_SEED},
"gate": "frozen V2 primary-contrast checks (as Core B V2-V6)",
"P1_CONTEXT_TTT_100_STATUS": "gate on AbsRel(NO_STREAM) - AbsRel(CTX_TTT_100)",
"P2_STREAM_TTT_100_STATUS": "gate on AbsRel(OFF) - AbsRel(TTT_100)",
"interpretation": {
"A": "both SUPPORTED",
"B": "only P2 SUPPORTED",
"C": "only P1 SUPPORTED",
"D": "neither SUPPORTED",
},
"descriptive": list(CONTRASTS),
"final_holdout": "never opened",
},
)
sealed = seal_previous(root)
data_hashes = {}
for _, record, _, base in entries.values():
for frame in record["frames"]:
for key in ("rgb", "depth"):
path = resolve(frame[key], base)
data_hashes[str(path)] = sha(path)
write_json(root / "audit/data_hashes.json", data_hashes)
names = (
"carrier_lock.json",
"cohort_lock.json",
"stream_roles.json",
"evaluation_contract.json",
"decision_rules.json",
"PROTOCOL.md",
PLAN,
"audit/data_hashes.json",
"audit/previous_experiment_seal.json",
)
sources = sorted(Path("src/mcss").rglob("*.py"))
sources += sorted(SCRIPTS.glob("*rgbd_stream_ttt_fresh_v7*.py"))
sources += [
SCRIPTS / "analyze_rgbd_stream_write_v3.py",
SCRIPTS / "run_rgbd_geometric_baselines.py",
SCRIPTS / "run_rgbd_inpaint_baselines.py",
LAUNCHER,
]
write_json(
root / "preregistration.json",
{
"status": "FROZEN_BEFORE_FORMAL_TRAINING",
"experiment": EXPERIMENT,
"input_sha256": {str((root / n).resolve()): sha(root / n) for n in names},
"source_sha256": {str(p.resolve()): sha(p) for p in sources},
"data_sha256": data_hashes,
"formal_training_started": False,
},
)
print(
"FROZEN_CORE_B_V7",
len(entries),
"scenes",
len(episodes),
"episodes",
sealed,
"files sealed",
)
def analyze(rows, seeds):
"""Every contrast with Core B V3's aggregation and the frozen V2 gate; the plan's branch."""
v3 = v3_analysis()
policies = sorted({r["policy"] for r in rows})
absolute = {
p: describe(v3.scene_values(rows, p), draws=DRAWS, seed=BOOT_SEED)["mean"] for p in policies
}
contrasts, statuses = {}, {}
for name, (a, b) in CONTRASTS.items():
stats = paired(
v3.scene_values(rows, a), v3.scene_values(rows, b), draws=DRAWS, seed=BOOT_SEED
)
status, _ = v3.gate(stats, True)
contrasts[name] = {
"mean": stats["mean"],
"ci95": stats["ci95"],
"positive_tied_negative": [stats["positive"], stats["tied"], stats["negative"]],
"gate_descriptive": status,
}
for key, value in PRIMARY.items():
if value == name:
statuses[key] = status
p1 = statuses["P1_CONTEXT_TTT_100_STATUS"] == "SUPPORTED"
p2 = statuses["P2_STREAM_TTT_100_STATUS"] == "SUPPORTED"
branch = {(True, True): "A", (False, True): "B", (True, False): "C", (False, False): "D"}[
p1, p2
]
per_seed = {
str(seed): {
name: float(np.mean(list(v3.scene_values(rows, a, seed=seed).values())))
- float(np.mean(list(v3.scene_values(rows, b, seed=seed).values())))
for name, (a, b) in CONTRASTS.items()
}
for seed in seeds
}
return {
"absolute": absolute,
"contrasts": contrasts,
"statuses": statuses,
"INTERPRETATION_BRANCH": branch,
"per_seed_means": per_seed,
}
def near_masks(cohorts):
"""{(scene, role, query): NEAR mask} from each role's measured context depth (model-free)."""
geo = module("v7_geometry", "run_rgbd_geometric_baselines.py")
inp = module("v7_inpaint", "run_rgbd_inpaint_baselines.py")
masks = {}
for _, record, manifest, base in cohort_records(cohorts, EXPECTED_SPLIT).values():
for role, context in eval_context(record, manifest, base, "cpu", []).items():
points = geo.context_points(context)
for qid in record["roles"]["primary_query"]:
camera = geo.query_camera(record, qid, manifest["image_size"])
_, hit = geo.reproject(points, camera)
masks[record["scene_id"], role, qid] = inp.near_mask(hit)
return masks
def recompute(root, cohorts, rows, masks):
"""Every metric from the saved prediction and GT; NORM and NEAR/FAR rows; the seals."""
entries = cohort_records(cohorts, EXPECTED_SPLIT)
worst = 0.0
derived = {"NORM": [], "FAR": [], "NEAR": []}
for row in rows:
path = Path(root) / "raw/eval_ttt" / f"seed_{row['seed']}" / row["prediction_path"]
if sha(path) != row["prediction_file_sha256"]:
raise AssertionError("Prediction changed")
pred = dict(np.load(path))
_, record, _, base = entries[row["scene_id"]]
frame = next(f for f in record["frames"] if f["frame_id"] == row["query_id"])
with Image.open(resolve(frame["rgb"], base)) as image:
rgb = np.array(image.convert("RGB"), dtype=np.float32) / 255
gt = np.load(resolve(frame["depth"], base)).squeeze()
metrics = measurement_metrics(pred["rgb"], rgb, pred["depth"], gt, pred["opacity"])
for key in METRICS:
worst = max(worst, abs(metrics[key] - row[key]))
common = {k: row[k] for k in KEYS}
depth = pred["depth"] / np.maximum(pred["opacity"], 1e-3)
norm = measurement_metrics(pred["rgb"], rgb, depth, gt, pred["opacity"])
derived["NORM"].append({**common, "depth_absrel": norm["depth_absrel"]})
valid = np.isfinite(gt) & (gt > 0)
near = masks[row["scene_id"], row["role"], row["query_id"]]
for region, mask in (("FAR", valid & ~near), ("NEAR", valid & near)):
if mask.any():
value = np.abs(pred["depth"][mask] - gt[mask]) / gt[mask]
derived[region].append({**common, "depth_absrel": float(value.mean())})
if worst > 1e-6:
raise AssertionError(f"Metric mismatch {worst}")
records = {sid: item[1] for sid, item in entries.items()}
for seed in {r["seed"] for r in rows}:
out = Path(root) / "raw/eval_ttt" / f"seed_{seed}"
audit = read(out / "integrity.json")
if sha(out / "states.pt") != audit["states_file_sha256"]:
raise AssertionError("Sealed states changed")
states = torch.load(out / "states.pt", map_location="cpu", weights_only=False)
index = read(out / "state_hashes.json")
if any(hash_scene_state(s) != index[k]["state_hash"] for k, s in states.items()):
raise AssertionError("State hash mismatch")
access = read(out / "GT_access.json")
marker = next(
i for i, e in enumerate(access) if e.get("event") == "ALL_STREAM_STATES_SEALED"
)
preseal_depth_audit(access[:marker], records)
return worst, derived
def ttt_descriptives(root, seeds):
out = {}
for policy, (_, supervision, _) in SPECS.items():
if supervision is None:
continue
first, last, norms = [], [], []
for seed in seeds:
index = read(Path(root) / "raw/eval_ttt" / f"seed_{seed}" / "state_hashes.json")
for item in index.values():
if item["policy"] == policy:
first.append(item["ttt_loss_first"])
last.append(item["ttt_loss_last"])
norms.append(item["fast_norm"])
drop = np.array(first) - np.array(last)
out[policy] = {
"episodes": len(first),
"loss_first_median": float(np.median(first)),
"loss_last_median": float(np.median(last)),
"share_with_lower_loss": float((drop > 0).mean()),
"fast_norm_median": float(np.median(norms)),
}
return out
def stat(value):
lo, hi = value["ci95"]
pos, tie, neg = value["positive_tied_negative"]
return f"{value['mean']:+.4f} [{lo:+.4f}, {hi:+.4f}]，场景 {pos}/{tie}/{neg}"
def report(root, result):
main = result["primary"]
c, s = main["contrasts"], main["statuses"]
lines = [
f"# {EXPERIMENT}",
"",
"Core B 第七轮：在全新的 EVAL-V5（23 个场景，与 V17 L 的训练集按 volume 不相交）上，"
"检验 100 步的测试时写入（依据 V23 在 EVAL-V3/V4 上的剂量结果选定），"
"并与 Core B V4 选出的 10 步对照。"
"计划见 [PLAN_BEFORE_COREB_V7.md](PLAN_BEFORE_COREB_V7.md)。"
"全部在 GPU（4 号卡，关闭 TF32）上进行，不能逐位复现。",
"",
"- **P1 CONTEXT_TTT_100（NO_STREAM − CTX_TTT_100）：** "
f"{stat(c['CONTEXT_TTT_100'])}，{s['P1_CONTEXT_TTT_100_STATUS']}",
"- **P2 STREAM_TTT_100（OFF − TTT_100）：** "
f"{stat(c['STREAM_TTT_100'])}，{s['P2_STREAM_TTT_100_STATUS']}",
f"- **分支 {main['INTERPRETATION_BRANCH']}**",
f"- **10 步的重现：** OFF − TTT {stat(c['STREAM_TTT'])}；"
f"NO_STREAM − CTX_TTT {stat(c['CONTEXT_TTT'])}",
f"- **剂量（10 步 − 100 步）：** 上下文 {stat(c['CONTEXT_DOSE'])}；"
f"流帧 {stat(c['STREAM_DOSE'])}",
f"- **监督来源（OFF_CTX_TTT − TTT）：** 10 步 {stat(c['SUPERVISION_SOURCE'])}；"
f"100 步 {stat(c['SUPERVISION_SOURCE_100'])}",
f"- **特异性：** TTT {stat(c['TTT_SPECIFICITY'])}；"
f"TTT_100 {stat(c['TTT_100_SPECIFICITY'])}；"
f"CTX_TTT {stat(c['CTX_TTT_SPECIFICITY'])}；"
f"CTX_TTT_100 {stat(c['CTX_TTT_100_SPECIFICITY'])}",
f"- **缓存新帧（NO_STREAM − OFF）：** {stat(c['STREAM_CACHE_EFFECT'])}",
"",
"| 策略 | query AbsRel | 归一化读出 | NEAR | FAR |",
"|---|---|---|---|---|",
]
for policy in (*SPECS, *CONTROLS):
lines.append(
f"| {policy} | {main['absolute'][policy]:.4f} | "
f"{result['normalized_readout']['absolute'][policy]:.4f} | "
f"{result['regions']['NEAR']['absolute'][policy]:.4f} | "
f"{result['regions']['FAR']['absolute'][policy]:.4f} |"
)
lines += [
"",
"| 区域（描述性） | STREAM_TTT_100 | CONTEXT_TTT_100 | SUPERVISION_SOURCE_100 |",
"|---|---|---|---|",
]
for region in ("NEAR", "FAR"):
rc = result["regions"][region]["contrasts"]
lines.append(
f"| {region} | {stat(rc['STREAM_TTT_100'])} | {stat(rc['CONTEXT_TTT_100'])} | "
f"{stat(rc['SUPERVISION_SOURCE_100'])} |"
)
lines += [
"",
"- **每个 seed：** "
+ "；".join(
f"{k}: CONTEXT_TTT_100 {v['CONTEXT_TTT_100']:+.4f}，"
f"STREAM_TTT_100 {v['STREAM_TTT_100']:+.4f}"
for k, v in main["per_seed_means"].items()
),
"- **被监督帧自身的损失（中位数）：** "
+ "；".join(
f"{p} {v['loss_first_median']:.4f} → {v['loss_last_median']:.4f}"
f"（{100 * v['share_with_lower_loss']:.0f}% 下降）"
for p, v in result["ttt"].items()
),
"",
"说明：EVAL-V5 是放宽规则后最后一批从未观察过的 Hypersim 场景，与 DEV 共享 volume，"
"23 个场景来自 7 个 volume；受保护的 final holdout 未打开。",
]
(Path(root) / "README.md").write_text("\n".join(lines) + "\n")
def run(root):
root = Path(root).resolve()
config = verify_lock_v7(root)
verify_data(root)
if read(root / "tests.json")["status"] != "PASS":
raise PermissionError("Full suite must PASS before the run")
if not os.sched_getaffinity(0) <= set(config["cpu_affinity"]):
raise PermissionError("Run must be pinned to the frozen CPU cores (taskset)")
if os.environ.get("CUDA_VISIBLE_DEVICES") != config["gpu"]:
raise PermissionError("Run must hold the frozen GPU through cv_gpu_run.sh")
if sha(LAUNCHER) != config["host_launcher"]["sha256"] or not os.environ.get("CV_ROOT"):
raise PermissionError("The frozen launcher and CV_ROOT are required")
for key, value in config["thread_environment"].items():
if os.environ.get(key) != value:
raise PermissionError(f"Thread environment {key} must be {value}")
torch.set_num_threads(1)
lock = read(root / "carrier_lock.json")["carriers"]
carriers = {int(s): load_frozen_carrier(item["path"], "cuda")[0] for s, item in lock.items()}
cohorts = load_cohorts(root)
rows = []
for seed, carrier in sorted(carriers.items()):
out = root / "raw/eval_ttt" / f"seed_{seed}"
seed_rows = evaluate_v7(carrier, cohorts, EXPECTED_SPLIT, out, seed, "cuda", SPECS)
rows.extend(seed_rows)
print("EVALUATED", seed, len(seed_rows), flush=True)
write_json(root / "raw/eval_ttt_query_results.json", rows)
masks = near_masks(cohorts)
worst, derived = recompute(root, cohorts, rows, masks)
seeds = sorted(carriers)
result = {
"primary": analyze(rows, seeds),
"normalized_readout": analyze(derived["NORM"], seeds),
"regions": {region: analyze(derived[region], seeds) for region in ("NEAR", "FAR")},
"ttt": ttt_descriptives(root, seeds),
"max_rgb_clamp_excess": max(r["rgb_clamp_max_excess"] for r in rows),
}
result["INTERPRETATION_BRANCH"] = result["primary"]["INTERPRETATION_BRANCH"]
write_json(root / "ttt_fresh_results.json", result)
for name, values in derived.items():
write_json(root / f"raw/{name.lower()}_rows.json", values)
old = read(root / "audit/previous_experiment_seal.json")["files"]
changed = [p for p, info in old.items() if sha(Path(p)) != info["sha256"]]
if changed:
raise AssertionError(f"Old experiment files changed: {changed[:3]}")
write_json(
root / "integrity.json",
{
"status": "PASS",
"metrics_recomputed": len(rows),
"max_metric_error": worst,
"old_files_unchanged": len(old),
"settings_reselected": False,
"slow_carrier_updated": False,
"final_holdout_touched": False,
"commit": subprocess.check_output(["git", "rev-parse", "HEAD"], text=True).strip(),
"finished_utc": datetime.datetime.now(datetime.UTC).isoformat(),
},
)
report(root, result)
print(
"CORE_B_V7_COMPLETE",
result["INTERPRETATION_BRANCH"],
json.dumps(result["primary"]["statuses"]),
flush=True,
)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--root", required=True)
parser.add_argument("--stage", choices=("freeze", "run"), required=True)
args = parser.parse_args()
if args.stage == "freeze":
freeze(args.root)
else:
run(args.root)
sys.exit(0)