83 lines
2.9 KiB
Python
83 lines
2.9 KiB
Python
"""Stage 1 of PREREG_flat_regime_2026-09-21.md: is the 'unkept promise' flag selective?
|
|
|
|
Flag after a step when: the step was the model's own (the evaluation that followed a traced
|
|
state is exactly x + raw, so no gate or safeguard replaced it), f did not change sign, and
|
|
r = f(x_new)/f(x) >= 1/4 - the break-even of a tangent step against a halving at m = 2.
|
|
|
|
python -u diag_flat_flag.py
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from scipy.optimize._tstutils import _APS_TESTS_DICTS, _CHANDRUPATLA_TESTS_DICTS
|
|
|
|
from simroot.pure import pure_root
|
|
|
|
XTOL, MAXFEV, ACC = 1e-12, 200, 1e-6
|
|
R_STAR = 0.25
|
|
CFG = dict(similarity="ratio", scale="derived3", guard="project")
|
|
FLAT = ("fun3", "fun4", "fun5", "fun6")
|
|
|
|
|
|
def flags(f, a, b, r_true):
|
|
"""(model steps judged, flags raised) for one run of the proposed rule."""
|
|
evals = []
|
|
|
|
def g(t):
|
|
v = f(t)
|
|
evals.append((t, v))
|
|
return v
|
|
st = {"trace": []}
|
|
x, _n, conv = pure_root(g, a, b, XTOL, MAXFEV, stats=st, **CFG)
|
|
if not (conv and abs(float(x) - r_true) <= ACC):
|
|
return 0, 0
|
|
order = [e[0] for e in evals]
|
|
judged = raised = 0
|
|
for s in st["trace"]:
|
|
try:
|
|
i = order.index(s["x"])
|
|
except ValueError:
|
|
continue
|
|
if i + 1 >= len(evals):
|
|
continue
|
|
x_next, f_next = evals[i + 1]
|
|
if x_next != s["x"] + s["raw"] or s["fx"] == 0.0:
|
|
continue # not the model's own step
|
|
judged += 1
|
|
r = f_next / s["fx"]
|
|
if r > 0.0 and r >= R_STAR:
|
|
raised += 1
|
|
return judged, raised
|
|
|
|
|
|
def main():
|
|
groups = {"APS": [0, 0], "CHA fun1": [0, 0], "CHA flat fun3-6": [0, 0],
|
|
"CHA fun7": [0, 0], "CHA other": [0, 0]}
|
|
for d in _APS_TESTS_DICTS:
|
|
a, b = (float(v) for v in d["bracket"])
|
|
f = (lambda ff, aa: lambda t: ff(t, *aa))(d["f"], d.get("args", ()))
|
|
j, r = flags(f, a, b, float(d["root"]))
|
|
groups["APS"][0] += j
|
|
groups["APS"][1] += r
|
|
for d in _CHANDRUPATLA_TESTS_DICTS:
|
|
a, b = (float(v) for v in d["bracket"])
|
|
fam = str(d["ID"]).split(".")[0]
|
|
g = ("CHA flat fun3-6" if fam in FLAT else "CHA fun1" if fam == "fun1"
|
|
else "CHA fun7" if fam == "fun7" else "CHA other")
|
|
j, r = flags(d["f"], a, b, float(d["root"]))
|
|
groups[g][0] += j
|
|
groups[g][1] += r
|
|
print(f"{'group':<18} {'model steps':>12} {'flagged':>8} {'share':>7}")
|
|
out = {}
|
|
for g, (j, r) in groups.items():
|
|
out[g] = (j, r, 100.0 * r / max(j, 1))
|
|
print(f"{g:<18} {j:12d} {r:8d} {out[g][2]:6.1f}%")
|
|
flat, aps, f1 = (out["CHA flat fun3-6"][2], out["APS"][2], out["CHA fun1"][2])
|
|
print(f"\nF1 flat > APS and flat > fun1: {flat:.1f}% vs {aps:.1f}% / {f1:.1f}% -> "
|
|
f"{'PASS' if flat > aps and flat > f1 else 'FAIL'}")
|
|
print("bar: m>=2 fires 0.2% canonical / 5.8% flat")
|
|
return out
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|