| """Score any S-parameter physical-validity checker against the corpus. |
| |
| The metric is deliberately NOT a single accuracy number. A checker has two |
| independent ways to be wrong and they carry very different costs: |
| |
| FALSE PASS the checker admits a non-physical network. |
| This is the one that ships a broken model downstream. |
| |
| FALSE FAIL the checker rejects a realizable network. |
| Annoying, erodes trust, but nothing bad reaches a simulator. |
| |
| So both are reported, per law, and the headline verdict requires ZERO false |
| passes. A checker that flags everything has no false passes and is useless -- |
| which is why the false-fail count sits beside it, exactly as in the |
| abstain-bench design. |
| |
| Adapter contract: point --checker at `module:function` where the function takes |
| a Touchstone path and returns {law_name: bool_passed}. |
| |
| Usage: |
| python score.py --checker sparam_lint_adapter:check |
| python score.py --checker mypkg.mychecker:run --json |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import importlib |
| import json |
| import sys |
| from pathlib import Path |
|
|
| HERE = Path(__file__).resolve().parent |
| DATA = HERE / "data" |
|
|
|
|
| def load_manifest(data_dir: Path = DATA) -> dict: |
| mf = data_dir / "manifest.json" |
| if not mf.exists(): |
| raise SystemExit(f"no manifest at {mf} -- run generate.py first") |
| return json.loads(mf.read_text(encoding="utf-8")) |
|
|
|
|
| def _load_checker(spec: str): |
| if ":" not in spec: |
| raise ValueError(f"expected 'module:function', got {spec!r}") |
| mod, fn = spec.split(":", 1) |
| f = getattr(importlib.import_module(mod), fn, None) |
| if f is None or not callable(f): |
| raise ValueError(f"{mod} has no callable {fn!r}") |
| return f |
|
|
|
|
| def score(checker, manifest: dict, data_dir: Path = DATA) -> dict: |
| laws = manifest["laws"] |
| per_case = [] |
| false_pass = 0 |
| false_fail = 0 |
| errors = 0 |
| per_law = {law: {"false_pass": 0, "false_fail": 0, "correct": 0} for law in laws} |
|
|
| for c in manifest["cases"]: |
| path = data_dir / c["file"] |
| try: |
| got = checker(str(path)) |
| except Exception as exc: |
| errors += 1 |
| per_case.append({"name": c["name"], "error": f"{type(exc).__name__}: {exc}"}) |
| continue |
|
|
| if not isinstance(got, dict): |
| errors += 1 |
| per_case.append({"name": c["name"], |
| "error": f"checker returned {type(got).__name__}, expected dict"}) |
| continue |
|
|
| case_fp, case_ff, missing = [], [], [] |
| for law in laws: |
| if law not in got: |
| missing.append(law) |
| continue |
| expected = c["expect"][law] |
| actual = bool(got[law]) |
| if actual == expected: |
| per_law[law]["correct"] += 1 |
| elif actual and not expected: |
| per_law[law]["false_pass"] += 1 |
| case_fp.append(law) |
| else: |
| per_law[law]["false_fail"] += 1 |
| case_ff.append(law) |
|
|
| false_pass += len(case_fp) |
| false_fail += len(case_ff) |
| per_case.append({ |
| "name": c["name"], |
| "physical": c["physical"], |
| "false_pass_laws": case_fp, |
| "false_fail_laws": case_ff, |
| "missing_laws": missing, |
| "correct": not case_fp and not case_ff and not missing, |
| }) |
|
|
| n_law_checks = len(manifest["cases"]) * len(laws) |
| n_missing = sum(len(c.get("missing_laws", [])) for c in per_case) |
| return { |
| "corpus_version": manifest["version"], |
| "n_cases": len(manifest["cases"]), |
| "n_law_checks": n_law_checks, |
| "false_pass": false_pass, |
| "false_fail": false_fail, |
| "checker_errors": errors, |
| "n_missing": n_missing, |
| "per_law": per_law, |
| "per_case": per_case, |
| |
| |
| |
| |
| "verdict": ("NOT CONFORMING" if (false_pass or errors) |
| else "INCOMPLETE" if n_missing |
| else "CONFORMING"), |
| "passed": false_pass == 0 and errors == 0 and n_missing == 0, |
| "verdict_rule": ( |
| "A conforming checker has ZERO false passes, ZERO errors, and " |
| "reports every law on every case. False fails are reported but do " |
| "not fail the verdict, because a conservative checker is safe and a " |
| "permissive one is not. A checker that omits a law is INCOMPLETE " |
| "rather than conforming: an unreported law was not checked, and " |
| "this corpus cannot certify what it never saw." |
| ), |
| } |
|
|
|
|
| def main(argv: list[str] | None = None) -> int: |
| ap = argparse.ArgumentParser(description="Score a checker against sparam-conformance.") |
| ap.add_argument("--checker", required=True, help="module:function") |
| ap.add_argument("--data", default=str(DATA)) |
| ap.add_argument("--json", action="store_true") |
| args = ap.parse_args(argv) |
|
|
| data_dir = Path(args.data) |
| manifest = load_manifest(data_dir) |
| try: |
| checker = _load_checker(args.checker) |
| except Exception as exc: |
| print(f"error: {exc}", file=sys.stderr) |
| return 2 |
|
|
| result = score(checker, manifest, data_dir) |
| result["checker"] = args.checker |
|
|
| if args.json: |
| print(json.dumps(result, indent=2)) |
| else: |
| print(f"sparam-conformance v{result['corpus_version']} " |
| f"{result['n_cases']} cases x {len(manifest['laws'])} laws") |
| print() |
| for c in result["per_case"]: |
| if "error" in c: |
| print(f" [ERR ] {c['name']:<22} {c['error']}") |
| elif c["correct"]: |
| print(f" [ OK ] {c['name']}") |
| else: |
| bits = [] |
| if c["false_pass_laws"]: |
| bits.append("FALSE PASS: " + ", ".join(c["false_pass_laws"])) |
| if c["false_fail_laws"]: |
| bits.append("false fail: " + ", ".join(c["false_fail_laws"])) |
| if c["missing_laws"]: |
| bits.append("not reported: " + ", ".join(c["missing_laws"])) |
| print(f" [FAIL] {c['name']:<22} {'; '.join(bits)}") |
| print() |
| print(f" false passes : {result['false_pass']} <- must be 0") |
| print(f" false fails : {result['false_fail']}") |
| print(f" errors : {result['checker_errors']}") |
| if result["n_missing"]: |
| print(f" laws not reported : {result['n_missing']} <- must be 0") |
| print(f" verdict : {result['verdict']}") |
|
|
| return 0 if result["passed"] else 1 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|