JG1310's picture
Upload folder using huggingface_hub
e85b663 verified
Raw
History Blame Contribute Delete
7.05 kB
#!/usr/bin/env python3
"""Procedure-fidelity gates for kcnuX4xEpL reproduction.
Usage:
python3 gates.py exp01 --toy # structural checks, relaxed params
python3 gates.py exp01 --full # structural checks + exact full-scale params
python3 gates.py exp01 --full --report
Gates verify STRUCTURE / SHAPES / SCHEMA / RANGES / (synthetic) PROVENANCE only.
They NEVER encode expected paper outcomes (no beta==1/(d+2) assertions).
"""
import json, os, sys, math
RESULTS = os.path.join(os.path.dirname(os.path.abspath(__file__)), "results", "exp01.json")
FULL = {
"N": 2000, "M": 2000, "R": 10, "d_grid": [100, 200, 500, 1000],
"eps_multipliers": [1e-8, 5e-8, 1e-7, 5e-7, 1e-6, 5e-6, 1e-5, 5e-5, 1e-4, 5e-4],
"initTol": 0.01, "tau": 1e-12, "base": 1.00005, "a": 0,
"solvers": ["nonlinear_gauss_seidel", "semismooth_newton"],
}
K = 10
def _num(x):
return isinstance(x, (int, float)) and not isinstance(x, bool)
def check(results, toy, report):
C = [] # (name, ok, detail)
def add(name, ok, detail=""):
C.append((name, bool(ok), detail))
m = results.get("meta", {})
add("meta present", isinstance(m, dict) and len(m) > 0)
# ---- synthetic provenance (this paper uses a SYNTHETIC family, no real dataset) ----
add("provenance: base==1.00005", m.get("base") == 1.00005, str(m.get("base")))
add("provenance: a==0", m.get("a") == 0)
add("provenance: tau==1e-12", m.get("tau") == 1e-12)
add("provenance: initTol==0.01", m.get("initTol") == 0.01)
add("provenance: solvers set",
m.get("solvers") == FULL["solvers"], str(m.get("solvers")))
em = m.get("eps_multipliers", [])
add("eps_multipliers length 10", isinstance(em, list) and len(em) == K)
if isinstance(em, list) and len(em) == K:
okm = all(abs(a - b) <= 1e-9 * max(1, abs(b)) for a, b in zip(em, FULL["eps_multipliers"]))
add("eps_multipliers match spec grid", okm, str(em))
if not toy:
add("full: N==2000", m.get("N") == 2000)
add("full: M==2000", m.get("M") == 2000)
add("full: R==10", m.get("R") == 10)
add("full: d_grid exact", m.get("d_grid") == FULL["d_grid"], str(m.get("d_grid")))
else:
add("toy: N present int", _num(m.get("N")))
add("toy: d_grid nonempty", isinstance(m.get("d_grid"), list) and len(m.get("d_grid")) >= 1)
d_grid = m.get("d_grid", []) if isinstance(m.get("d_grid"), list) else []
solvers = m.get("solvers", []) if isinstance(m.get("solvers"), list) else []
R = m.get("R", 0)
# ---- records ----
recs = results.get("records", [])
add("records is list nonempty", isinstance(recs, list) and len(recs) > 0)
if not isinstance(recs, list):
recs = []
if not toy and d_grid and solvers and _num(R):
add("full: record count == len(d_grid)*R*len(solvers)",
len(recs) == len(d_grid) * R * len(solvers),
f"{len(recs)} vs {len(d_grid) * R * len(solvers)}")
fields = ["d", "seed", "solver", "c_med", "eps_actual", "dbias", "converged",
"n_active", "beta_hat", "alpha_hat", "rel_err"]
all_shape_ok = True
all_range_ok = True
seen = set()
for r in recs:
if not isinstance(r, dict):
all_shape_ok = False
continue
if not all(f in r for f in fields):
all_shape_ok = False
continue
seen.add((r.get("d"), r.get("seed"), r.get("solver")))
for arr in ("eps_actual", "dbias", "converged", "n_active"):
v = r.get(arr)
if not (isinstance(v, list) and len(v) == K):
all_shape_ok = False
# ranges
if not (_num(r.get("c_med")) and r["c_med"] > 0):
all_range_ok = False
ea = r.get("eps_actual", [])
if isinstance(ea, list) and all(_num(x) and x > 0 for x in ea):
pass
else:
all_range_ok = False
db = r.get("dbias", [])
if isinstance(db, list):
for x in db:
if x is None:
continue
if not (_num(x) and x >= 0 and math.isfinite(x)):
all_range_ok = False
if not (_num(r.get("beta_hat")) and math.isfinite(r["beta_hat"])):
all_range_ok = False
if r.get("solver") not in solvers:
all_range_ok = False
if r.get("d") not in d_grid:
all_range_ok = False
add("all records have required fields", all_shape_ok)
add("all record arrays length 10", all_shape_ok)
add("record ranges sane (c_med>0, eps>0, dbias>=0/null, beta finite)", all_range_ok)
if not toy and d_grid and solvers and _num(R):
expected = {(d, s, sv) for d in d_grid for s in range(R) for sv in solvers}
add("full: (d,seed,solver) grid complete", seen == expected,
f"missing {len(expected - seen)}")
# ---- summary ----
summ = results.get("summary", [])
add("summary is list nonempty", isinstance(summ, list) and len(summ) > 0)
if isinstance(summ, list):
sok = True
for s in summ:
if not isinstance(s, dict):
sok = False
continue
for f in ("d", "solver", "theory", "beta_mean", "beta_std",
"rel_err_mean", "rel_err_std"):
if f not in s:
sok = False
if _num(s.get("d")) and _num(s.get("theory")):
if abs(s["theory"] - 1.0 / (s["d"] + 2)) > 1e-9:
sok = False # theory column must equal 1/(d+2) (definition, not outcome)
add("summary rows well-formed; theory==1/(d+2)", sok)
if not toy and d_grid and solvers:
add("full: summary has len(d_grid)*len(solvers) rows",
len(summ) == len(d_grid) * len(solvers),
f"{len(summ)} vs {len(d_grid) * len(solvers)}")
ok = all(c[1] for c in C)
if report:
print(f"=== gates report exp01 ({'toy' if toy else 'full'}) ===")
for name, cok, detail in C:
print(f" [{'PASS' if cok else 'FAIL'}] {name}" + (f" ({detail})" if detail and not cok else ""))
print(f"=== {'ALL PASS' if ok else 'FAILURES PRESENT'} ({sum(c[1] for c in C)}/{len(C)}) ===")
return ok
def main():
args = sys.argv[1:]
if not args or args[0] != "exp01":
print("usage: python3 gates.py exp01 --toy|--full [--report]")
sys.exit(2)
toy = "--toy" in args
full = "--full" in args
if toy == full:
print("specify exactly one of --toy / --full")
sys.exit(2)
report = "--report" in args
if not os.path.exists(RESULTS):
print(f"[FAIL] results file missing: {RESULTS}")
sys.exit(1)
try:
with open(RESULTS) as f:
results = json.load(f)
except Exception as e:
print(f"[FAIL] cannot parse {RESULTS}: {e}")
sys.exit(1)
ok = check(results, toy=toy, report=report or True)
sys.exit(0 if ok else 1)
if __name__ == "__main__":
main()