Download code/guard_sensitivity.py from ai4he/headroom-decision-procedure: direct link, hf CLI and curl.
- Browser
- Download file 3.21 kB
-
https://huggingface.co/datasets/ai4he/headroom-decision-procedure/resolve/main/code/guard_sensitivity.py
- Command line
-
hf download hf://datasets/ai4he/headroom-decision-procedure/code/guard_sensitivity.py
-
curl -L -o guard_sensitivity.py https://huggingface.co/datasets/ai4he/headroom-decision-procedure/resolve/main/code/guard_sensitivity.py
3.21 kB
| #!/usr/bin/env python3 | |
| """Sensitivity of the procedure's verdicts to its own prespecified thresholds. | |
| Answers a reviewer question: how do the four synthetic verdicts move as the | |
| H >= 0.05 ratio-stability guard and the 5 percent undefined-draw rule vary? | |
| Also varies the minimum benefit and harm tolerance, which are the thresholds | |
| that actually enter the verdict function. | |
| All records are declared synthetic constructions, not observations. | |
| """ | |
| import csv, math, random, statistics | |
| from pathlib import Path | |
| import worked_headroom as W | |
| CASES = W.CASES | |
| def boot(name, counts, alpha=0.05, B=50_000): | |
| both, p_only, b_only, neither = counts | |
| pairs=([(1,1)]*both+[(0,1)]*p_only+[(1,0)]*b_only+[(0,0)]*neither) | |
| n=len(pairs); rng=random.Random(f"{W.SEED}:{name}") | |
| d=[]; r=[]; undef=0 | |
| for _ in range(B): | |
| dr=[pairs[rng.randrange(n)] for _ in range(n)] | |
| a0=sum(x for x,_ in dr)/n; a1=sum(y for _,y in dr)/n | |
| d.append(a1-a0) | |
| if a0==1: undef+=1 | |
| else: r.append((a1-a0)/(1-a0)) | |
| return (W.percentile(d,alpha/2), W.percentile(d,1-alpha/2), | |
| undef/B, (both+b_only)/n, (both+p_only)/n) | |
| def verdict(dl,dh,H,min_benefit,harm_tol): | |
| if H==0: return "no-opportunity" | |
| if min_benefit>H: return "design-infeasible" | |
| if dl>=min_benefit: return "benefit-supported" | |
| if dh<=-harm_tol: return "harm-supported" | |
| if dl>=-harm_tol and dh<=min_benefit: return "practically-negligible" | |
| return "unresolved" | |
| print("=== cached bootstraps ===") | |
| B={} | |
| for nm,c in CASES.items(): | |
| B[nm]=boot(nm,c); print(f" {nm:18} dCI=[{B[nm][0]:+.3f},{B[nm][1]:+.3f}] undef={B[nm][2]:.3%} a0={B[nm][3]:.2f}") | |
| print("\n=== (a) vary the RATIO-STABILITY GUARDS: H floor and max undefined ===") | |
| rows=[] | |
| for hf in (0.02,0.05,0.10): | |
| for mu in (0.01,0.05,0.10,0.25,0.40,1.00): | |
| line=[] | |
| for nm,c in CASES.items(): | |
| dl,dh,undef,a0,a1=B[nm]; H=1-a0 | |
| r_rep = (H>=hf and undef<=mu and H>0) | |
| v=verdict(dl,dh,H,W.MIN_BENEFIT,W.HARM_TOLERANCE) | |
| line.append(f"{nm.split('_')[0]}:{'R' if r_rep else '-'}/{v[:6]}") | |
| rows.append({"H_floor":hf,"max_undefined":mu,"case":nm, | |
| "R_reported":r_rep,"verdict":v}) | |
| print(f" Hfloor={hf:.2f} maxundef={mu:.2f} "+" ".join(line)) | |
| vs={(r['case'],) for r in rows} | |
| import collections | |
| print("\n verdicts per case across ALL guard settings:") | |
| for nm in CASES: | |
| s={r['verdict'] for r in rows if r['case']==nm} | |
| rr={r['R_reported'] for r in rows if r['case']==nm} | |
| print(f" {nm:18} verdict(s)={s} R_reported varies={len(rr)>1}") | |
| print("\n=== (b) vary MIN_BENEFIT (delta_b) and HARM_TOLERANCE (delta_h) ===") | |
| for mb in (0.01,0.02,0.05,0.10): | |
| for ht in (0.01,0.02,0.05): | |
| line=[] | |
| for nm,c in CASES.items(): | |
| dl,dh,undef,a0,a1=B[nm]; H=1-a0 | |
| line.append(f"{nm.split('_')[0]}:{verdict(dl,dh,H,mb,ht)[:9]}") | |
| print(f" d_b={mb:.2f} d_h={ht:.2f} "+" ".join(line)) | |
| rows2=[] | |
| out=Path("guard_sensitivity_results.csv") | |
| with out.open("w",newline="",encoding="utf-8") as fh: | |
| wr=csv.DictWriter(fh,fieldnames=list(rows[0])); wr.writeheader(); wr.writerows(rows) | |
| print(f"\nwrote {out}") | |