Download code/severity_scaled_case.py from ai4he/headroom-decision-procedure: direct link, hf CLI and curl.
- Browser
- Download file 3.38 kB
-
https://huggingface.co/datasets/ai4he/headroom-decision-procedure/resolve/main/code/severity_scaled_case.py
- Command line
-
hf download hf://datasets/ai4he/headroom-decision-procedure/code/severity_scaled_case.py
-
curl -L -o severity_scaled_case.py https://huggingface.co/datasets/ai4he/headroom-decision-procedure/resolve/main/code/severity_scaled_case.py
3.38 kB
| #!/usr/bin/env python3 | |
| """A declared synthetic case where the accuracy gate and the loss gate disagree. | |
| Addresses a reviewer point: the opportunity gate is stated in accuracy units, | |
| but accuracy is often not the decision-relevant outcome. Here accuracy headroom | |
| is negligible while severity-weighted loss headroom is material, so the two | |
| scales return opposite verdicts on identical data. | |
| Construction, not an observation. | |
| """ | |
| import math, random, statistics | |
| from pathlib import Path | |
| import csv | |
| SEED=20260829; B=50_000; ALPHA=0.05 | |
| MIN_BENEFIT=0.02; HARM_TOL=0.02 | |
| # 100 items: 10 high-severity (weight 1.0), 90 low-severity (weight 0.1). | |
| # Baseline errs on one HIGH-severity item; protocol fixes it but errs on one | |
| # LOW-severity item. Accuracy is identical; severity-weighted loss is not. | |
| items=[] | |
| for i in range(10): # high severity | |
| base_ok = (i!=0) # baseline wrong on item 0 | |
| prot_ok = True # protocol right on all high-severity items | |
| items.append((1.0, base_ok, prot_ok)) | |
| for i in range(90): # low severity | |
| base_ok = True | |
| prot_ok = (i!=0) # protocol wrong on one low-severity item | |
| items.append((0.1, base_ok, prot_ok)) | |
| def acc(sample): | |
| n=len(sample) | |
| return (sum(1 for _,b,_ in sample if b)/n, sum(1 for _,_,p in sample if p)/n) | |
| def loss(sample): | |
| tw=sum(w for w,_,_ in sample) | |
| l0=sum(w for w,b,_ in sample if not b)/tw | |
| l1=sum(w for w,_,p in sample if not p)/tw | |
| return l0,l1 | |
| def pct(v,q): | |
| o=sorted(v); pos=q*(len(o)-1); lo=math.floor(pos); hi=math.ceil(pos) | |
| return o[lo] if lo==hi else o[lo]*(1-(pos-lo))+o[hi]*(pos-lo) | |
| a0,a1=acc(items); l0,l1=loss(items) | |
| H_acc=1-a0; D_acc=a1-a0 | |
| L_STAR=0.0 # declared attainable loss floor | |
| H_loss=l0-L_STAR; D_loss=l0-l1 # loss reduction is a benefit | |
| rng=random.Random(f"{SEED}:severity") | |
| da=[]; dl=[] | |
| for _ in range(B): | |
| s=[items[rng.randrange(len(items))] for _ in range(len(items))] | |
| x0,x1=acc(s); m0,m1=loss(s) | |
| da.append(x1-x0); dl.append(m0-m1) | |
| da_lo,da_hi=pct(da,ALPHA/2),pct(da,1-ALPHA/2) | |
| dl_lo,dl_hi=pct(dl,ALPHA/2),pct(dl,1-ALPHA/2) | |
| def verdict(d_lo,d_hi,H): | |
| if H==0: return "no-opportunity" | |
| if MIN_BENEFIT>H: return "design-infeasible" | |
| if d_lo>=MIN_BENEFIT: return "benefit-supported" | |
| if d_hi<=-HARM_TOL: return "harm-supported" | |
| if d_lo>=-HARM_TOL and d_hi<=MIN_BENEFIT: return "practically-negligible" | |
| return "unresolved" | |
| print("=== identical data, two outcome scales ===") | |
| print(f" ACCURACY : a0={a0:.3f} a1={a1:.3f} H={H_acc:.3f} D={D_acc:+.3f} " | |
| f"CI=[{da_lo:+.3f},{da_hi:+.3f}] -> {verdict(da_lo,da_hi,H_acc)}") | |
| print(f" LOSS : l0={l0:.4f} l1={l1:.4f} H={H_loss:.4f} (floor {L_STAR}) " | |
| f"D={D_loss:+.4f} CI=[{dl_lo:+.4f},{dl_hi:+.4f}] -> {verdict(dl_lo,dl_hi,H_loss)}") | |
| print(f" R_loss = D/H = {D_loss/H_loss:.3f}") | |
| print(f"\n weights: 10 items at 1.0, 90 items at 0.1 (total {sum(w for w,_,_ in items):.1f})") | |
| with Path("severity_scaled_results.csv").open("w",newline="",encoding="utf-8") as fh: | |
| wr=csv.writer(fh) | |
| wr.writerow(["scale","baseline","protocol","headroom","delta","ci_low","ci_high","verdict"]) | |
| wr.writerow(["accuracy",a0,a1,H_acc,D_acc,da_lo,da_hi,verdict(da_lo,da_hi,H_acc)]) | |
| wr.writerow(["severity_weighted_loss",l0,l1,H_loss,D_loss,dl_lo,dl_hi,verdict(dl_lo,dl_hi,H_loss)]) | |
| print("wrote severity_scaled_results.csv") | |