#!/usr/bin/env python3
"""从 round{N}_verdicts.json 汇总 A 轨 8 维 / B 轨 11 场景分数（分数可复算的唯一入口）。
用法: aggregate.py scores_v1/round1_verdicts.json
"""
import json, sys

SCN = json.load(open('/app/system/agents/deep-research/train/digest/test_scenarios_v2.json'))
WEIGHT = {'D1': 15, 'D2': 15, 'D3': 15, 'D4': 15, 'D5': 10, 'D6': 10, 'D7': 10, 'D8': 10}
EW = {'E1': 20, 'E2': 20, 'E3': 30, 'E4': 20, 'E5': 10}

def main(path):
    V = json.load(open(path))
    meta = SCN['meta']; crit_meta = {}
    for s in SCN['scenarios']:
        for c in s['criteria']:
            crit_meta[c['id']] = c
    dims = {d: [0, 0] for d in WEIGHT}
    rows = []
    a_total = 0.0
    for sid, sv in sorted(V['scenarios'].items()):
        a_hit = a_all = 0
        for cid, cv in sv['criteria'].items():
            cm = crit_meta[cid]
            pts = cm['points']
            if cm['track'] != 'A':
                continue
            a_all += pts
            if cv['hit']:
                a_hit += pts
                d = dims[cm['dim']]; d[0] += pts
            dims[cm['dim']][1] += pts
        scn_a = round(a_hit / a_all * 100, 2) if a_all else 0.0
        b = sum(sv['B'][e]['score'] * w for e, w in EW.items()) / 100.0
        veto = not sv['B']['E4']['score'] == 100
        if veto:
            b = 0.0  # E4 一票否决：该场景 B 轨直接 0 分（benchmark §三）
        rows.append((sid, scn_a, round(b, 2), veto))
        a_total += 0  # placeholder
    dimscore = {d: round(h / t * 100, 2) if t else 0.0 for d, (h, t) in dims.items()}
    a_track = round(sum(dimscore[d] * WEIGHT[d] / 100 for d in WEIGHT), 2)
    scn_a_mean = round(sum(r[1] for r in rows) / len(rows), 2)
    print(f'== {path}  skill {V["skill_version"]} ==')
    for d in WEIGHT:
        h, t = dims[d]
        print(f'  {d} 权重{WEIGHT[d]:>3}: {h}/{t} = {dimscore[d]:>6}  ×{WEIGHT[d]/100:.2f} = {round(dimscore[d]*WEIGHT[d]/100,2):>5}')
    print(f'  A 轨总分 = {a_track}   （11 场景 A 分简单均值 = {scn_a_mean}）')
    print(f'  弱维度（低于 60）: {[d for d in WEIGHT if dimscore[d] < 60]}')
    print('  场景明细: 场景A分 / B分 / E4否决')
    for sid, a, b, veto in rows:
        flag = '✓' if (a >= 90 and b >= 85 and not veto) else '✗'
        print(f'   {sid}: A={a:>6}  B={b:>6}  veto={veto}  {flag}')
    ok = a_track >= 90 and all(b >= 85 and not v for _, _, b, v in rows)
    print(f'  双门槛: A轨 {a_track} ≥90 ? {a_track >= 90}；B轨逐场景 ≥85 ? {all(b >= 85 and not v for _, _, b, v in rows)} → {"放行" if ok else "不放行"}')
    out = path.replace('_verdicts.json', '_summary.json')
    json.dump({'skill_version': V['skill_version'], 'dimensions': {d: {'hit': dims[d][0], 'total': dims[d][1], 'score': dimscore[d], 'weight': WEIGHT[d]} for d in WEIGHT},
               'a_track_total': a_track, 'a_scenario_mean': scn_a_mean,
               'scenarios': {sid: {'A': a, 'B': b, 'veto': v} for sid, a, b, v in rows},
               'gate_pass': ok}, open(out, 'w'), ensure_ascii=False, indent=1)
    print(f'  -> {out}')

if __name__ == '__main__':
    main(sys.argv[1])