dollar_table.py

"""Audit fix #9: the fused operating-point dollar table (the money number) + the
honest triage value, and a Malinois-baseline availability check."""

import os, glob
from pathlib import Path
import numpy as np, pandas as pd

HERE = Path(__file__).resolve().parent

df = pd.read_csv(HERE / "data" / "gosai_designed" / "designed_scored.csv")
score = -df["pred_gap"].values
fail = df["measured_fail"].values.astype(int)
n = len(df)
base = fail.mean()
LO, HI = (
    250,
    1000,
)  # $ per design, DIRECT synthesis + individual validation only. This is a NARROWER basis than the
# full-committed-spend model in docs/cisfalcon_triage_economics.csv ($950-$3,500 per pursued design,
# which adds downstream follow-up); both are reported with their basis stated. This script uses the
# direct-cost basis; the shipped economics figure uses the full-committed-spend basis.

print(
    f"n={n}  base fail={base:.4f}  (~{base * 100:.1f} wasted syntheses per 100 designs if uncaught)"
)
print(
    "\nTRIAGE dollar table, POOLED basis (scrutinize/redesign the flagged riskiest BEFORE synthesis):"
)
print(
    "  Every figure below is POOLED across a mixed batch. The project disowns the pooled basis\n"
    "  as evidence of per-design signal: a sequence-free stratum-prior rule beats it. The\n"
    "  conditioned money numbers (2.59x macro enrichment, $458-$1,796 net) are in\n"
    "  economics_table.py, and that conditioned column is the one to quote."
)
order = np.argsort(-score)
ys = fail[order]
tp = np.cumsum(ys)
fp = np.cumsum(1 - ys)
recall = tp / fail.sum()
ppv = tp / (tp + fp)
for topk in [0.02, 0.05, 0.10, 0.20]:
    i = int(topk * n) - 1
    fails_caught = recall[i] * base * 100
    print(
        f"  flag riskiest {topk * 100:4.0f}% -> PPV {ppv[i]:.2f}, catches {recall[i] * 100:4.1f}% of failures "
        f"({fails_caught:.1f}/100); redesign-first avoids ${recall[i] * base * 100 * LO:.0f}-{recall[i] * base * 100 * HI:.0f} per 100 designs"
    )

print(
    "\nSAFEST-FIRST framing, POOLED basis (limited synthesis budget -> make lowest-risk first):"
)
safe = np.argsort(score)
for frac in [0.5, 0.7, 0.9]:
    m = int(frac * n)
    fr = fail[safe[:m]].mean()
    print(
        f"  synthesize safest {frac * 100:.0f}%: failure rate {fr * 100:.2f}% (vs {base * 100:.2f}% random) "
        f"= {(1 - fr / base) * 100:.0f}% fewer failures in what you make"
    )
print(
    "  Conditioned within (cell x generator), the safest-half reduction is 41% macro, not 70%\n"
    "  (triage_conditioning_check.py). A sequence-free stratum-prior rule reaches about 90% pooled."
)

print("\n=== Malinois self-prediction availability (killer comparator) ===")
files = [
    os.path.basename(c)
    for c in glob.glob(
        str(HERE / "data" / "gosai_designed" / "**" / "*"), recursive=True
    )
    if os.path.isfile(c)
]
print("gosai_designed files:", files)
b = pd.read_csv(HERE / "data" / "gosai_designed" / "designed_benchmark.csv", nrows=2)
print("benchmark cols (any Malinois pred?):", list(b.columns))