Files
SkillFactory/scripts/audit_contact_cycle.py
chenlin 9fe9d27865 feat: checkpoint contact-cycle evaluation and RootXY PPO progress
Consolidate next-stage docs, archive superseded narratives, and preserve frozen document inputs. Retain invalid P2 pairs, partial RootXY training, failed attempts and nonselectable checkpoints as explicit limitations.

Include supervised evaluation/training, desktop GPU admission, bounded evidence and atomic-publication monitor repairs with CPU regressions. Ignore local dotenv configuration.

Validation: 1403 passed, 6 skipped for contact-cycle and related contracts/CLI tests. Full-suite legacy ABI and registration tests require Python 3.12/Torch 2.10/pxr; no runtime checks relaxed. No new physical run or accepted skill.
2026-10-15 14:18:17 +08:00

80 lines
3.9 KiB
Python

#!/usr/bin/env python3
"""Offline CPU trace readback. Audit consistency is NOT skill acceptance."""
import argparse
import hashlib
import json
from pathlib import Path
import torch
from runtime_common import write_json
from SkillFactory.legs.contact_cycle.metrics import summarize
def audit(run):
run = Path(run)
report = json.loads((run/"result.json").read_text())
cfg = json.loads((run/"config.json").read_text())
locked = json.loads((run/"invocation.json").read_text())["thresholds"]["gates"]
trace = torch.load(run/"physics_trace.pt", map_location="cpu", weights_only=True)
rows = summarize(trace, report["physics_dt"], locked, cfg["scenarios"])
samples = round(sum(cfg["phase_durations_s"])/report["physics_dt"])
gates = {
"not_skill_acceptance": report["accepted_skill"] is False and report["trained_skill"] is False,
"same_rows_from_raw_trace": rows == report["rows"],
"summary_status_consistent": (
report["passed"] == all(report["gates"].values())
and report["gates"]["all_scenario_metrics"] == all(r["success"] for r in rows)
and report["status"] == ("completed" if report["passed"] else "failed")
),
"complete_duration": len(trace["time"]) == samples == report["physics_steps"],
"monotonic_physics_clock": bool(torch.allclose(trace["time"][1:]-trace["time"][:-1],
torch.full_like(trace["time"][1:], report["physics_dt"]),
atol=1e-6, rtol=0)),
}
identity = json.loads((run/"identity.json").read_text())
snapshot_files = list((run/"source_snapshot").rglob("*"))
hashes = {hashlib.sha256(p.read_bytes()).hexdigest() for p in snapshot_files if p.is_file()}
gates["bound_source_snapshots"] = all(digest in hashes for digest in identity.values())
# Independent scalar loop for the slip integral (does not call summarize's implementation).
slip = []
for e in range(trace["phase"].shape[1]):
total = 0.0
for k in range(len(trace["time"])):
phase, progress = int(trace["phase"][k, e]), float(trace["progress"][k, e])
stance = (phase == 0 and float(trace["time"][k, e]) >= 2) or (phase == 5 and progress >= 0.8)
if stance and float(trace["normal"][k, e, 2]) > 2:
total += float(trace["slip_speed"][k, e]) * report["physics_dt"]
slip.append(total)
gates["independent_slip_integral"] = all(
not r["complete"] or not r["finite"] or abs(slip[i]-r["stance_slip_m"]) <= 1e-5
for i, r in enumerate(rows))
failures = []
for row in rows:
if not row["success"]:
e = row["env"]
indices = torch.nonzero(trace["would_fail"][:, e]).flatten()
failures.append({"env": e, "scenario": row["scenario"], "reasons": row["failure_reasons"],
"first_task_failure_s": float(trace["time"][indices[0], e]) if len(indices) else None,
"peak_height_error_s": float(trace["time"][trace["height_error"][:, e].argmax(), e])})
return {"passed": all(gates.values()), "accepted_skill": False, "audit_only": True,
"original_physical_passed": report["passed"], "gates": gates, "failure_cases": failures,
"trace_sha256": hashlib.sha256((run/"physics_trace.pt").read_bytes()).hexdigest(),
"report_sha256": hashlib.sha256((run/"result.json").read_bytes()).hexdigest()}
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--run", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
if args.output.exists():
parser.error("Refusing to overwrite audit evidence")
result = audit(args.run)
write_json(args.output, result)
raise SystemExit(0 if result["passed"] else 2)
if __name__ == "__main__":
main()