#!/usr/bin/env python3
"""三引擎裸描述盲测报告页生成器。
- 逐图随机匿名(甲/乙/丙), 映射写 results/blind_map.json
- 页面元素: 大图 + 程序化事实锚点(details折叠) + 三张匿名描述卡 + 单选判定 + 揭晓按钮
"""
import json, base64, random, hashlib
from pathlib import Path
NL = Path('/home/kai/shared/qwen-eye-exp/nltest')
RES = NL / 'results'
OUT_DIR = Path('/home/kai/shared/2026-08-27')
OUT = OUT_DIR / 'nl-blind-test.html'
IMAGES = ['L1_ocr', 'L2_scene', 'L3_night', 'L4_grid']
IMG_TITLE = {
'L1_ocr': ('L1 · 特写+文字', '难度① 主体特写带手写字条 —— 考 OCR 与微观细节'),
'L2_scene': ('L2 · 中景场景', '难度② 单主体环境互动 —— 考方位与主体状态'),
'L3_night': ('L3 · 夜景远景', '难度③ 低光远景多光源 —— 考光照与色温感知'),
'L4_grid': ('L4 · 四格拼图', '难度④ 合成 2×2 多格 —— 考网格分解与逐格分析'),
}
ARMS = ['local', 'hosted', '46v']
def b64(p):
return 'data:image/jpeg;base64,' + base64.b64encode(Path(p).read_bytes()).decode()
# ── 载入结果 ──
data = {a: json.loads((RES / f'nl_{a}.json').read_text()) for a in ARMS}
facts = json.loads((NL / 'facts.json').read_text())
# ── 盲测映射 (稳定随机: seed=内容哈希, 复现一致) ──
random.seed(hashlib.sha256(json.dumps({k: sorted(v) for k, v in facts.items()}, sort_keys=True).encode()).hexdigest())
blind_map = {}
for img in IMAGES:
arms_shuffled = ARMS[:]
random.shuffle(arms_shuffled)
blind_map[img] = dict(zip(['甲', '乙', '丙'], arms_shuffled))
(RES / 'blind_map.json').write_text(json.dumps(blind_map, ensure_ascii=False, indent=1))
DIM_LABELS = [
('overall', '整体概览'), ('subject', '主体'), ('layout', '方位布局'),
('colors', '颜色'), ('lighting', '光照'), ('composition', '构图'),
('text', '文字转录'), ('defects', '缺陷检查'),
]
def fact_html(name):
f = facts[name]
doms = ' '.join(f"{c} {round(p*100)}%"
for c, p in [(d[0], d[1]/100) for d in f['dominant_colors']])
extra = ''
if name == 'L4_grid':
extra = f"
真实拼图顺序: {facts['_grid_truth']}
"
return f"""
平均亮度: {f['avg_brightness']}/255
尺寸: {f['size'][0]}×{f['size'][1]}
程序化主色 (PIL 自适应调色板): {doms}
{extra}
"""
cards_all = ''
for idx, img in enumerate(IMAGES):
title, sub = IMG_TITLE[img]
label_of = blind_map[img] # {'甲': arm, ...}
descs = []
for i, lbl in enumerate(['甲', '乙', '丙']):
arm = label_of[lbl]
rec = data[arm][img]
dims_html = ''
if rec.get('dims'):
rows = ''
for key, zh in DIM_LABELS:
val = str(rec['dims'].get(key) or '').strip()
if not val:
val = '(未作答)'
rows += f"
{zh}
{val}
"
dims_html = f"
{rows}
"
else:
dims_html = f"
解析失败原始输出:\n{rec.get('raw','')[:800]}
"
descs.append(f"""
{lbl} 号引擎
{dims_html}
""")
cards_all += f"""
{title}
{sub}
{ ''.join(descs) }
🧑⚖️ 你的裁决
对照原图读完三份描述后，选出你心中最强的引擎：
{''.join(f"" for l in ['甲','乙','丙'])}
📐 程序化事实锚点（判完再看）
{fact_html(img)}
"""
ARM_NAME = {'local': 'qwen-eye :8125', 'hosted': 'glm-5.3-flash', '46v': 'glm-4.6v'}
map_rows = ''
for k in IMAGES:
tds = ''.join(f"
{lbl} = {ARM_NAME[blind_map[k][lbl]]}
" for lbl in ['甲', '乙', '丙'])
map_rows += f"
{IMG_TITLE[k][0]}
{tds}
"
SCRIPT_BLOCK = ''''''
html = f"""
三引擎视觉描述盲测 · Kai 裁判局
⚖️ 已裁定: 0/4 张 · 你选中的赢家: 尚无
👁🗨 三引擎视觉描述盲测 · 裁判局
2026-08-27 · qwen-eye(:8125 本地) vs glm-5.3-flash vs glm-4.6v · 同图同 prompt(temp 0.1) · 八维强制框架
})