verify.py

#!/usr/bin/env python3
"""Supervisor recomputation for one model dir: line counts, per-tier accuracy (failures count wrong),
schema validity, ECE (top-label, 10 bins), Brier, latency p50/p95, status codes. Reads only."""
import json, sys, os, math, statistics, collections
root = sys.argv[1]
tiers = ['easy', 'original', 'hard']
expect = {'easy': 48, 'original': 72, 'hard': 111}
def pct(xs, p):
    if not xs: return None
    xs = sorted(xs); k = (len(xs) - 1) * p; f = math.floor(k); c = math.ceil(k)
    return xs[f] if f == c else xs[f] + (xs[c] - xs[f]) * (k - f)
def ece(pairs, bins=10):
    if not pairs: return None
    tot = 0.0
    for b in range(bins):
        lo, hi = b / bins, (b + 1) / bins
        sel = [(c, o) for c, o in pairs if (lo < c <= hi) or (b == 0 and c == 0)]
        if sel:
            conf = sum(c for c, _ in sel) / len(sel); acc = sum(o for _, o in sel) / len(sel)
            tot += len(sel) / len(pairs) * abs(conf - acc)
    return tot
out = {}
all_lat = []
for t in tiers:
    p = os.path.join(root, f'results-{t}.jsonl')
    if not os.path.exists(p): out[t] = 'MISSING'; continue
    rows = [json.loads(l) for l in open(p) if l.strip()]
    n = len(rows); ok = sum(1 for r in rows if r['ok']); valid = sum(1 for r in rows if r['valid'])
    strict = sum(1 for r in rows if r.get('strict_valid')); correct = sum(1 for r in rows if r['correct'])
    codes = collections.Counter(str(r.get('status_code')) for r in rows)
    lat = [r['latency_s'] for r in rows if r['ok']]; all_lat += lat
    pairs = []; brier = []
    for r in rows:
        if r['valid'] and r.get('probs'):
            probs = r['probs']; top = max(probs, key=lambda k: probs[k])
            pairs.append((probs[top], 1.0 if r['correct'] else 0.0))
            # Brier over exact labels: expected one-hot vs probs (expected label = predicted iff correct)
    out[t] = {'rows': n, 'expected_rows': expect[t], 'ok': ok, 'valid': valid, 'strict_valid': strict,
              'correct': correct, 'accuracy_over_all': round(correct / n, 4) if n else None,
              'status_codes': dict(codes), 'p50_s': pct(lat, .5), 'p95_s': pct(lat, .95),
              'ece_top_label_10bin': round(ece(pairs), 4) if pairs else None, 'n_calib': len(pairs),
              'ids_unique': len({r['task_id'] for r in rows}) == n}
out['all_tiers'] = {'n': len(all_lat), 'p50_s': pct(all_lat, .5), 'p95_s': pct(all_lat, .95)}
# cross-check against jevbench summarize output when present
for t in tiers:
    sp = os.path.join(root, f'summary-{t}.json')
    if os.path.exists(sp):
        s = json.load(open(sp)); out[t + '_jevbench_summary'] = {k: s.get(k) for k in ('accuracy', 'n_scorable', 'n_correct', 'schema_validity', 'schema_validity_strict', 'ece', 'brier_mean', 'coverage', 'latency')}
print(json.dumps(out, indent=1))