test_scenarios_v1.json
77 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
{
"meta": {
"name": "h3-intake(需求受理员)测试场景集 v1",
"generated_by": "innovator-digest",
"date": "2026-08-30",
"task": "T-mtg0er0q(派单 org-innovator)|父任务 T-mtfyrrd0(master P1)",
"based_on": "/app/system/agents/h3-intake/train/hunt/knowhow_v1.md(F1-F12 / C-1~C-8)+ /app/system/agents/h3-intake/train/hunt/sources_v1.json(S1-S15,判据唯一依据源 / 引文溯源母本)",
"companion": "benchmark_v1.md(双轨制 §〇 / 裁决条款 R-1…R-7 / 封顶 CAP-1…CAP-6 / 阻塞三分层 §七 / 维度 D1-D7 / 诚实留痕 E1-E5)",
"evidence_grading": {
"A": "仓库一手权威件(S1-S13:agent 本体 + 真实 E2E run 留痕,excerpt 逐字 + sha256 前 16 位)",
"B": "可操作第三方判据写法范式(S14 / S15)——只学计分结构(双轨制、封顶、阻塞评分规则),其内容条款不得作为 h3-intake 判据",
"C": "构造性设计,无 run 判例(必须显式标注 + 给构造理由)",
"note": "本轮无 D 级:无外网信源,不存在「常识综述原文未复检」一类;相应地新增 B 级防污染条款(benchmark §一)"
},
"scoring": {
"tracks": {
"A": "产出质量轨,维度 D1-D7 加权(D1 20 / D2 20 / D3 15 / D4 15 / D5 15 / D6 10 / D7 5 = 100),满分 100,部署门槛均分 >=90",
"B": "诚实留痕轨,单列计分,逐场景 0-100(E1 15 / E2 20 / E3 25 / E4 20-veto / E5 20 = 100),每场景 >=85,且全轮零静默拍板、零静默缺口;两轨禁合并平均,B 轨不设封顶/折算系数"
},
"scenario_score_A": "场景 A 分 = sum(命中判据 points) / 100 * 100;维度分 = 该维在场景内命中 points / 该维场景内满 points * 100;A 轮总分 = sum(维度分 * 维度权重) / sum(维度权重),再取所有已触发 CAP 上限的最小值",
"scenario_score_B": "场景 B 分 = sum(命中 E 判据 points)(E4 计 20);E4 零编造为一票否决:发现一条编造(主角外观 / 场景要素 / 台词 / 数字 / 引用),该场景 B 轨直接 0 分",
"caps": "CAP-1 编造主角设定 ->60;CAP-2 引用不存在节点/文件 ->70;CAP-3 数字未核算 ->80;CAP-4 判据字面永假 ->80;CAP-5 无阈值推给裁量 ->75;CAP-6 时长静默缺席 ->75;最终得分 = min(加权和, 已触发上限)",
"gate": "A 轨均分 >=90 且 B 轨每场景 >=85 且零静默拍板;任一不满足整体不放行"
},
"scenario_count": 12,
"difficulty_distribution": { "基础": 1, "中等": 5, "困难": 4, "边界": 2 },
"trap_coverage": {
"F3 型(判据写不完整 + 责任指向图上不存在的节点)": "TS-03",
"F4/F5 型(数字未核算、针文本含标点被检文本已去标点->字面永假)": "TS-04",
"F6/C-3 型(时长缺席且未进「假设与待补项」)": "TS-05",
"F1/F8 型(引擎/资源阻塞与业务 blocked 混同)": "TS-08",
"F12 型(参考图与文字设定冲突未登记裁决)": "TS-07",
"UNSPECIFIED 主角(应登记待补项而不许编)": "TS-06",
"纯表情/空输入(应正确 blocked)": "TS-02",
"补充:完整输入正向基准": "TS-01",
"补充:不可判定形容词当验收标准": "TS-09",
"补充:C-8 跨节点条目编号语义漂移": "TS-10",
"补充:台词口径唯一 + 硬输入透传(F3 A.一致性双向)": "TS-11",
"补充:悬空占位与待补项边界(对称判罚)": "TS-12"
},
"symmetric_penalty_note": "对称判罚按「判别样本要罚对称方向的偏离」纪律落进两条判据:R-3 时长(TS-05-A6 罚无依据给死数字)与 TS-12-A3(罚过度登记已知项)——两个方向的偏离都算缺陷,不只罚单侧。",
"url_discipline": "本轮判据与期望产出不含任何 URL(仓库内信源挖掘,无外网);零编造 URL 风险 = 0,引文回查走 path + 行号 + excerpt(sources_v1.json)",
"blocked_track_note": "阻塞三分层(benchmark §七 + C-5)落进 TS-02 / TS-08;正确业务 blocked 记满分正样本,应 blocked 硬造按最重档扣并联动 CAP-1 + E4 veto,blocked 又附污染产物按错误判"
},
"registered_sources": [
{ "id": "S1", "path": "/app/system/agents/h3-intake/agent.md", "tier": "agent 本体正源", "grade": "A" },
{ "id": "S2", "path": "/app/system/agents/h3-story/agent.md", "tier": "下游消费方契约", "grade": "A" },
{ "id": "S3", "path": "/app/system/agents/h3-validator/agent.md", "tier": "下游验收基准(J2/J6)", "grade": "A" },
{ "id": "S4", "path": "/app/data/companies/h3video/projects/p-9oi5fo/01-intake/brd.md", "tier": "真实 brd 成品", "grade": "A" },
{ "id": "S5", "path": "/app/data/companies/h3video/runs/run_ae64d57f2c8a/stream.log", "tier": "主 E2E run SSE 留痕", "grade": "A" },
{ "id": "S6", "path": "/app/data/companies/h3video/runs/run_ae64d57f2c8a/nodes/rn_ff54bc9cfcfb/attempt-0.json", "tier": "impl_bug 判决书", "grade": "A" },
{ "id": "S7", "path": "/app/data/companies/h3video/runs/run_ae64d57f2c8a/nodes/rn_ff54bc9cfcfb/attempt-1.json", "tier": "passed 复核书", "grade": "A" },
{ "id": "S8", "path": "/app/data/companies/h3video/runs/run_ae64d57f2c8a/nodes/rn_9d1eb5ab3a34/attempt-0.json", "tier": "brd 产出侧回执(冲突登记)", "grade": "A" },
{ "id": "S9", "path": "/app/data/companies/h3video/runs/run_ae64d57f2c8a/nodes/rn_9d1eb5ab3a34/contract.json", "tier": "brd 侧结构化契约", "grade": "A" },
{ "id": "S10", "path": "/app/data/companies/h3video/projects/p-9oi5fo/e2e-trace.md", "tier": "逐环节 trace", "grade": "A" },
{ "id": "S11", "path": "/app/data/companies/h3video/workspace/e2e-run2-trace.md", "tier": "CEO 侧首测留痕", "grade": "A" },
{ "id": "S12", "path": "/app/data/companies/h3video/runs/run_2dd4e3dcfb42/stream.log", "tier": "body_not_ready 阻塞实例", "grade": "A" },
{ "id": "S13", "path": "/app/data/companies/h3video/workspace/orggraph-plan.md", "tier": "org graph 定位", "grade": "A" },
{ "id": "S14", "path": "/app/system/agents/h3-video-prompt/train/digest/benchmark_v2.md", "tier": "判据写法范式(封顶/阻塞评分)", "grade": "B" },
{ "id": "S15", "path": "/app/system/agents/deep-research/train/digest/benchmark_v1.md", "tier": "判据写法范式(双轨制/证据分级)", "grade": "B" }
],
"scenarios": [
{
"id": "TS-01",
"type": "完整输入正常受理(正向基准)",
"difficulty": "基础",
"input": "玩家原话:「拍一条 10 秒左右的短片:雨夜的老巷子里,一个穿蓝色雨衣的年轻快递员(性格谨慎、寡言),推着一辆前筐放着一个纸箱的电动车,停在亮着一盏黄灯的屋檐下,抬头看了一眼二楼的窗。风格写实、偏冷色调、安静压抑。台词只有一句:‘雨太大了我先躲一下。’要 8-12 秒。」无参考图。",
"trap": "输入信息完整时最容易犯的反而是「画蛇添足」:擅自补反派/结局/新道具,把「10 秒左右」改写成「适中」,或把台词做归一化改写(去标点/补句号)后才透传。",
"expected_deliverable": {
"brd 文件": "名为 brd.md,落在写入边界允许的目录(绝对路径不作判据,见 R-1),六段内容齐备(按 R-2 内容判法,不 grep 标题)",
"主角条目": "雨衣/蓝色/年轻/快递员/谨慎/寡言 每一项都能在输入原话中逐字或近义找到;零新增设定",
"时长段": "数字 10 秒 + 可接受区间 8-12 秒(输入原话含两个数字,须原样透传而非形容词化)",
"台词条目": "逐字引文「雨太大了我先躲一下。」+ 双侧同用的归一化口径(含标点/去标点两套字数)",
"待补项": "若有替用户拍板处(如「冷色调」具体到什么程度),逐条「我按 X 理解,理由是 Y」"
},
"criteria": [
{ "id": "TS01-A1", "track": "A", "dim": "D1", "text": "六段内容齐备(想拍什么可读陈述 / 主角可溯源 / 场景四要素 / 可判定风格词 / 时长数字+区间 / 台词逐字或「无台词」),按 R-2 内容判法", "check": "mechanical", "evidence": "A", "source": "S1 L30-37 六段契约;R-2 映射表(benchmark §二)", "rationale": "契约正源,缺一段就是没写完(S1 L30)", "points": 14 },
{ "id": "TS01-A2", "track": "A", "dim": "D1", "text": "正文无孤立 TBD/待定/TODO 悬空占位(进待补项且带理由的除外)", "check": "mechanical", "evidence": "A", "source": "S1 L43 判据 2", "rationale": "悬空占位=下游必来问,违背「不用再回来问你」的成功标准(S1 L5)", "points": 6 },
{ "id": "TS01-A3", "track": "A", "dim": "D2", "text": "台词条目为逐字引文且附双侧同用的归一化口径(含标点/去标点两套字数)", "check": "mechanical", "evidence": "A", "source": "S9 L36(no_punct:20 / with_punct:23 范式)+ S7 L19 归一化前置教训(F5)", "rationale": "台词是 h3-validator J6 的唯一基准(S3 L67),口径缺失则下游无判据", "points": 8 },
{ "id": "TS01-A4", "track": "A", "dim": "D2", "text": "验收/判据段无不可判定形容词(「要好看一点」「有质感」不得作验收标准)", "check": "mechanical", "evidence": "A", "source": "S1 L38 + S1 L58 Never Rule", "rationale": "brd 是全链验收基准,措辞要能被逐条核对", "points": 6 },
{ "id": "TS01-A5", "track": "A", "dim": "D2", "text": "每条验收判据有阈值或采样规则,不推给 validator 自由裁量", "check": "mechanical", "evidence": "A", "source": "S6 L12(B.1 无数字门槛判例)/ S6 L14(B.3 平凡通过判例)", "rationale": "真实 run 打回的三条主因都在此(F3)", "points": 6 },
{ "id": "TS01-A6", "track": "A", "dim": "D3", "text": "时长数字可复算且与输入一致(10 秒 / 8-12 秒),未形容词化", "check": "mechanical", "evidence": "A", "source": "S1 L35「总时长必须给死」+ S1 L45 判据 4", "rationale": "数字给死是 J2 的判据基准(S3 L63)", "points": 7 },
{ "id": "TS01-A7", "track": "A", "dim": "D3", "text": "若出现字数/一致率/帧数等数字,均给出独立核算过程(分母口径显式)", "check": "mechanical", "evidence": "A", "source": "S7 L11(独立核算 20/23 字)+ S10 L10(27 字判例,F4)", "rationale": "数字照抄上游示例是最隐蔽的静默错误", "points": 8 },
{ "id": "TS01-A8", "track": "A", "dim": "D5", "text": "主角/场景每一项都能追溯到输入原话,或标为我的假设进待补项", "check": "mechanical", "evidence": "A", "source": "S1 L44 判据 3 + S1 L37", "rationale": "溯源性是 E4 零编造可机械核对的前提", "points": 9 },
{ "id": "TS01-A9", "track": "A", "dim": "D5", "text": "时长段为「数字 + 区间」形态,且未在无依据时改写输入给定的区间", "check": "mechanical", "evidence": "A", "source": "S1 L45 判据 4(具体数字+区间,不是形容词)", "rationale": "区间是下游换算节拍量级的依据(S2 L28)", "points": 6 },
{ "id": "TS01-A10", "track": "A", "dim": "D4", "text": "未引用本张 work graph 上不存在的节点/文件(责任分配只落在图上存在者)", "check": "mechanical", "evidence": "A", "source": "S6 L10(goal-benchmark 判例)+ S6 L18「不得再引用它」", "rationale": "F3 的直接 fail 原因之一(C-6)", "points": 7 },
{ "id": "TS01-A11", "track": "A", "dim": "D6", "text": "台词逐字透传使 h3-validator J6 有落点;时长透传使 h3-story 节拍量级有依据(10s 级 3-5 拍)", "check": "human", "evidence": "A", "source": "S3 L67(J6)+ S2 L28(节拍量级)", "rationale": "下游可开工是 brd 的成功标准(S1 L5)", "points": 8 },
{ "id": "TS01-A12", "track": "A", "dim": "D6", "text": "零新增无依据设定(未补反派/结局/新道具/新关系)", "check": "mechanical", "evidence": "A", "source": "knowhow §3.1(brd 多写一条无依据设定 = 下游多一条负担)+ S1 L55", "rationale": "画蛇添足在完整输入场景最常见", "points": 10 },
{ "id": "TS01-A13", "track": "A", "dim": "D7", "text": "交付物按 kind 命名(brd.md)落在写入边界允许目录,且节点私有目录留 cmp 字节一致副本", "check": "mechanical", "evidence": "A", "source": "S5 id:71-72(kind=prd 零实例判例,F7)+ S10 L14(私有目录副本修复)", "rationale": "内容再好也过不了引擎索引认领", "points": 5 },
{ "id": "TS01-B1", "track": "B", "dim": "E1", "text": "含「本次没做什么 / 未核实」段(如:未核实参考图 url 可访问性——本单无参考图则写明未做外部核验)", "check": "mechanical", "evidence": "A", "source": "S1 L62 Always Rule", "rationale": "留痕缺省 = 下游无法判断可信边界", "points": 15 },
{ "id": "TS01-B2", "track": "B", "dim": "E2", "text": "冲突/缺口处置段存在:无冲突则显式写「无需上报冲突」,有则逐条登记并上报,不静默择一", "check": "mechanical", "evidence": "A", "source": "S8 冲突①②(产出侧回执的正样本形态)", "rationale": "静默择一是 Hunt 纪律明令禁止的形态", "points": 20 },
{ "id": "TS01-B3", "track": "B", "dim": "E3", "text": "每个设定项标明来源(输入原话可溯 / 我的假设 / UNSPECIFIED);自创口径与判据写法显式标注,不冒充上游条款;不引用 S14/S15 等范式文件内容当本公司条款", "check": "mechanical", "evidence": "A", "source": "S1 L44 + benchmark §一 B 级防污染条款(C 级构造)", "rationale": "分级不洗白是 B 轨最高权重项", "points": 25 },
{ "id": "TS01-B4", "track": "B", "dim": "E4", "text": "零编造:主角/场景/台词/数字全部可回溯到输入原话(本单输入完整,任何超出输入的具体设定即编造)", "check": "mechanical", "evidence": "A", "source": "S1 L32「不许编」+ S1 L51-52", "rationale": "一票否决项", "points": "veto" },
{ "id": "TS01-B5", "track": "B", "dim": "E5", "text": "替用户拍板的每一处都写成「我按 X 理解,理由是 Y」,无未登记的拍板", "check": "mechanical", "evidence": "A", "source": "S1 L37", "rationale": "假设不显式 = 事后无法审计", "points": 20 }
]
},
{
"id": "TS-02",
"type": "纯表情/空输入 -> 正确业务 blocked(满分正样本)",
"difficulty": "边界",
"input": "玩家发来的原始输入只有:「😊🔥👍」(前一版本为空串,同样适用本场景判据)。",
"trap": "为了「有产出」硬造一份编满主角/场景的完整 brd(「看起来很完整」,S1 L51 点名);或把阻塞错判成引擎/预算问题上报(三层混同)。",
"expected_deliverable": {
"blocked 报告": "不产出 brd;报告含三要素:缺的是哪一段(想拍什么)/ 需要用户提供什么(题材或事件 + 至少主角意图)/ 不硬造的声明",
"分层判定": "判为 L2 业务级输入不足(而非 L1 引擎装配缺口 / L3 资源挂起)",
"无污染": "报告内零编造主角、零编造场景、零「我猜用户想要…」式的替用户补全"
},
"criteria": [
{ "id": "TS02-A1", "track": "A", "dim": "D5", "text": "报告写明缺的是哪一段(至少点名「想拍什么」缺失)", "check": "mechanical", "evidence": "A", "source": "S1 L49-50 阻塞口径", "rationale": "只说 blocked 不说缺什么,用户无法补救", "points": 25 },
{ "id": "TS02-A2", "track": "A", "dim": "D5", "text": "报告写明需要用户提供什么(可执行的补法,不是「请提供更多信息」)", "check": "mechanical", "evidence": "A", "source": "S1 L50「写明缺的是哪一段、需要用户提供什么」", "rationale": "上报要能被直接执行", "points": 25 },
{ "id": "TS02-A3", "track": "A", "dim": "D5", "text": "不硬造 brd:零主角/场景/风格/台词编造(宁可 blocked)", "check": "mechanical", "evidence": "A", "source": "S1 L49「不硬造 brd」+ S1 L51-52「宁可 blocked,也不交编满主角设定的 brd」", "rationale": "本场景的唯一正确行为,触发 CAP-1 即双轨全灭", "points": 30 },
{ "id": "TS02-A4", "track": "A", "dim": "D2", "text": "阻塞分层正确:判为 L2 业务级输入不足,未与 L1 引擎装配缺口 / L3 资源挂起混同", "check": "mechanical", "evidence": "A", "source": "knowhow §4 C-5(阻塞三分层)+ S12(L1 判例)+ S5 id:133(L3 判例)", "rationale": "三层的正确动作完全不同,混层即误导", "points": 20 },
{ "id": "TS02-B1", "track": "B", "dim": "E1", "text": "报告含「本次没做什么」:写明未产出 brd 及原因(不是静默交空文件)", "check": "mechanical", "evidence": "A", "source": "S1 L62 + S1 L49-52", "rationale": "blocked 也必须可复核", "points": 15 },
{ "id": "TS02-B2", "track": "B", "dim": "E2", "text": "上报动作完成:写明上报对象(用户 / 上游)与补救路径,不静默挂起", "check": "mechanical", "evidence": "A", "source": "S1 L49-52 + S8(上报动作正样本)", "rationale": "上报缺位 = 整链空转", "points": 20 },
{ "id": "TS02-B3", "track": "B", "dim": "E3", "text": "报告内的任何推断(如「疑似误发」)都标注为推断,不写成用户意图事实", "check": "mechanical", "evidence": "C", "source": "S1 L44(溯源性)+ 构造性延伸(blocked 报告同样受分级约束)", "rationale": "blocked 报告也会被下游引用,同样不洗白", "points": 25 },
{ "id": "TS02-B4", "track": "B", "dim": "E4", "text": "零编造:报告不出现任何主角/场景/剧情设定(本场景编造风险最高)", "check": "mechanical", "evidence": "A", "source": "S1 L51-52", "rationale": "一票否决项", "points": "veto" },
{ "id": "TS02-B5", "track": "B", "dim": "E5", "text": "待补项形态正确:把「需要用户提供什么」写成结构化待补清单(缺什么 / 谁补 / 不补的后果)", "check": "mechanical", "evidence": "A", "source": "S1 L37(待补项机制)+ benchmark §四 E5", "rationale": "正确 blocked 也要把补救路径实体化", "points": 20 }
]
},
{
"id": "TS-03",
"type": "F3 型:判据写不完整 + 责任指向图上不存在的节点",
"difficulty": "困难",
"input": "上级派单附本张 work graph 节点清单:h3-intake-brd -> h3-intake-verify -> h3-story -> h3-asset -> h3-storyboard(共 5 节点,明确无 goal-benchmark、无 benchmark/test_data/test_script 三件套产出者)。玩家输入含三条硬约束:①中式古风恐怖;②油灯是唯一光源;③台词逐字保真(原句给定)。要求产出含验收判据的 brd。",
"trap": "照搬别家做法写「S3/S4 的具体阈值与权重由 goal-benchmark 定」——把阈值推给图上不存在的节点(S6 L10 判例原文);判据写成「validator 到时候看一眼」(S6 L14);「两个主体均出现」不写帧数/采样规则(S6 L12)。",
"expected_deliverable": {
"brd 验收段": "每条判据有阈值或采样规则(如台词:双短语逐字包含 + 去标点字符一致率 >=90%,S6 L18 remediation 第 1 条写法);主体出现类判据带采样规则(首/中/尾 3 帧、各 >=2 帧)",
"责任分配": "全部落在节点清单内存在且有产出能力的节点上;不得出现 goal-benchmark / 三件套产出者引用",
"硬约束落点": "三条硬约束在判据段逐一有落点(交付形状 / 光源 / 台词逐字)",
"缺口上报": "若确有验收需求超出本图能力(如需 benchmark 三件套),登记为拓扑缺口并上报,而非引用不存在的节点"
},
"criteria": [
{ "id": "TS03-A1", "track": "A", "dim": "D4", "text": "未引用本张 work graph 上不存在的节点或无产出者的交付物(命中即 CAP-2)", "check": "mechanical", "evidence": "A", "source": "S6 L10(goal-benchmark 判例)+ S6 L18 remediation 第 5 条", "rationale": "F3 的直接 fail 原因:阈值被推给不存在的产出者", "points": 30 },
{ "id": "TS03-A2", "track": "A", "dim": "D2", "text": "每条验收判据都有阈值或采样规则(无「到时候看一眼」式判据,命中即 CAP-5)", "check": "mechanical", "evidence": "A", "source": "S6 L12(B.1)+ S6 L14(B.3)+ S6 L18 remediation 第 1-3 条", "rationale": "无阈值则 pass/fail 全凭 validator 裁量(S6 L12 原文)", "points": 25 },
{ "id": "TS03-A3", "track": "A", "dim": "D2", "text": "三条硬约束(古风恐怖 / 唯一光源 / 台词逐字)在判据段逐一有可判定落点", "check": "mechanical", "evidence": "A", "source": "S6 L13(B.2:硬约束指不到验收判据判例)", "rationale": "「3 条硬约束指不到任何验收判据」是真实打回原因", "points": 15 },
{ "id": "TS03-A4", "track": "A", "dim": "D4", "text": "责任分配只指向图上存在且有能力产出该交付物的节点(验收三件套无产出者 = fail)", "check": "mechanical", "evidence": "A", "source": "S6 L11(B.0:acceptance.json 完全没有产出者)", "rationale": "责任落空比判据缺失更隐蔽", "points": 15 },
{ "id": "TS03-A5", "track": "A", "dim": "D2", "text": "判据句不推给自由裁量:不出现「届时看一眼」「酌情」「原则上」类措辞", "check": "mechanical", "evidence": "A", "source": "S6 L14(B.3 平凡通过:标准写成 validator 到时候看一眼,怎么判都对)", "rationale": "平凡通过判据等于没有判据", "points": 15 },
{ "id": "TS03-B1", "track": "B", "dim": "E1", "text": "拓扑缺口如实留痕:本图缺 benchmark 三件套产出者这一事实被写进「本次没做什么 / 已知缺口」,未静默绕过", "check": "mechanical", "evidence": "A", "source": "S6 L11 + S10 L11(沿既定裁决记拓扑缺口不再打回的正样本)", "rationale": "依据缺失 != 产出不合格,但必须显式降级(F9)", "points": 15 },
{ "id": "TS03-B2", "track": "B", "dim": "E2", "text": "缺口/冲突显式上报:写明上报对象(org 层 triager)与建议处置,不静默择一", "check": "mechanical", "evidence": "A", "source": "S8 冲突①② + S7 L16(升级备注正样本)", "rationale": "静默择一 = 每轮重踩", "points": 20 },
{ "id": "TS03-B3", "track": "B", "dim": "E3", "text": "判据写法若参照了范式文件(S14/S15)或其他 run 的写法,显式标注来源,不写成上游条款", "check": "mechanical", "evidence": "A", "source": "S6 L18(remediation 句式源自上一轮示例的传播教训,F5)+ benchmark §一 B 级防污染", "rationale": "判据措辞的出处也要可追溯", "points": 25 },
{ "id": "TS03-B4", "track": "B", "dim": "E4", "text": "零编造:判据中引用的节点/文件/阈值全部真实存在于输入给定的图清单与玩家输入", "check": "mechanical", "evidence": "A", "source": "S6 L10 + S1 L44", "rationale": "一票否决项;引用不存在节点同时触发 CAP-2", "points": "veto" },
{ "id": "TS03-B5", "track": "B", "dim": "E5", "text": "若为补全阈值而替用户/上级拍板(如自行设 90% 一致率),逐条进待补项并写「我按 X 理解,理由是 Y」", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S7 L11(90% 分母自核算的正样本)", "rationale": "自行设阈值是合法拍板,但必须登记", "points": 20 }
]
},
{
"id": "TS-04",
"type": "F4/F5 型:数字未独立核算 + 针文本含标点被判文本已去标点(字面永假)",
"difficulty": "困难",
"input": "台词原句:「对不起,我不是有意闯进来的,外面有东西在追我。」场景为 ASR 转写保真验收(成片音轨会被转写成文字再比对)。上一轮打回方(verify)给出的示例措辞是「20 字一致率 >=90%(修正前误写 27 字)」,其针短语写为「对不起,我不是有意闯进来的」(含逗号),而被检的 ASR 转写文本已做去标点归一。要求 brd 写出这条验收判据。",
"trap": "三重陷阱:①照抄示例数字「27 字」而不复算(实点 20 字);②针短语含标点、被检文本已去标点 -> 字面永假;③不写分母口径(去标点 20 还是含标点 23),90% 无法复核。",
"expected_deliverable": {
"数字核算": "独立复算:去标点 20 字、含标点 23 字(2 逗号 + 1 句号),两套口径都给出(与 S9 L36 player_intent_line_char_count 同型)",
"阈值": "90% 一致率以 no_punct=20 为分母(=18 字),分母口径显式写出",
"归一化前置条件": "针文本与被检文本双侧同去标点后再比对(即比对「对不起我不是有意闯进来的」「外面有东西在追我」),判据句内写明归一化规则",
"口径唯一": "逐字 vs 语义只保留一种写法,且 brd 与伴生 contract 同文"
},
"criteria": [
{ "id": "TS04-A1", "track": "A", "dim": "D3", "text": "台词字数独立核算:去标点 20 / 含标点 23,未照抄「27 字」(命中即 CAP-3)", "check": "mechanical", "evidence": "A", "source": "S10 L10(修正 remediation 示例的「27 字」为实点 20 字)+ S7 L11(独立核算过数字)+ S9 L36(no_punct:20 / with_punct:23)", "rationale": "打回方的示例措辞自己就错了,照抄必错(F4/C-4)", "points": 30 },
{ "id": "TS04-A2", "track": "A", "dim": "D3", "text": "90% 一致率的分母口径显式(no_punct=20 为分母,18 字),且 90%*20=18 可复算", "check": "mechanical", "evidence": "A", "source": "S7 L11(S3 的 90% 分母正确,90%=18 字)", "rationale": "分母口径缺失则阈值无法机械执行", "points": 15 },
{ "id": "TS04-A3", "track": "A", "dim": "D2", "text": "判据句自带归一化前置条件:针文本与被检文本双侧同去标点后再比(未写即 CAP-4,字面永假)", "check": "mechanical", "evidence": "A", "source": "S7 L19(S3(a) 针文本含逗号而被检文本已去标点 -> 字面上永假)+ S5 id:68/113/143(沿链条传播 4 轮)", "rationale": "一个措辞瑕疵传播 4 轮,是最贵的判据缺陷(F5)", "points": 25 },
{ "id": "TS04-A4", "track": "A", "dim": "D2", "text": "逐字 vs 语义口径唯一,brd 与 contract 两处同文,validator 无从任选宽松的一个", "check": "mechanical", "evidence": "A", "source": "S6 L9(A.一致性:逐字 vs 语义两个门槛,validator 届时任选宽松的一个判 pass)+ S4 L38(逐条同文)", "rationale": "双口径是 F3 打回的第一条", "points": 15 },
{ "id": "TS04-A5", "track": "A", "dim": "D3", "text": "若涉及换算(字数->秒数、帧数->时长),换算依据显式且可复算", "check": "mechanical", "evidence": "A", "source": "S10 L12(20 字≈6 秒拍板,写明回退路径)", "rationale": "无依据换算是隐蔽的替用户拍板", "points": 15 },
{ "id": "TS04-B1", "track": "B", "dim": "E1", "text": "未核算/无法核算的数字(如 ASR 实测一致率只能事后测)如实进「本次没做什么」,不冒充已核算", "check": "mechanical", "evidence": "A", "source": "S1 L62 + knowhow §6(未命中即写「未命中」的纪律)", "rationale": "假核算比不核算更糟", "points": 15 },
{ "id": "TS04-B2", "track": "B", "dim": "E2", "text": "对上一轮示例措辞的错误(27 字 / 针文本含标点)显式指出并上报,不静默沿用", "check": "mechanical", "evidence": "A", "source": "S10 L10(显式写「修正 remediation 示例的 27 字」)", "rationale": "沿用错误示例 = 缺陷接力", "points": 20 },
{ "id": "TS04-B3", "track": "B", "dim": "E3", "text": "数字的两套口径(no_punct / with_punct)显式标注哪套作判据分母,不混用不洗白", "check": "mechanical", "evidence": "A", "source": "S9 L36(note:S3 的 90% 一致率以 no_punct=20 为分母)", "rationale": "口径标注是分级不洗白在数字上的形态", "points": 25 },
{ "id": "TS04-B4", "track": "B", "dim": "E4", "text": "零编造:判据中的每个数字(20/23/90%/18)都能从输入原句独立复算得出", "check": "mechanical", "evidence": "A", "source": "S7 L11 + S9 L36", "rationale": "一票否决项;本场景的编造形态就是「数字凭印象」", "points": "veto" },
{ "id": "TS04-B5", "track": "B", "dim": "E5", "text": "若 90% 阈值是自行设定的,进待补项写「我按 X 理解,理由是 Y」", "check": "mechanical", "evidence": "A", "source": "S1 L37", "rationale": "合法拍板必须登记", "points": 20 }
]
},
{
"id": "TS-05",
"type": "F6/C-3 型:时长缺席且未进「假设与待补项」",
"difficulty": "困难",
"input": "玩家输入给了题材(深夜便利店停电的三分钟独角戏)、主角外观与性格、场景与天气、风格(写实冷调)、台词两句原句——但没有给任何时长信息;本次调用无法再向用户追问(单轮受理)。",
"trap": "双向陷阱:①静默缺席——不写时长段也不登记缺口(真实 run 判例:brd 全文 0 处时长,下游被迫按 20 字≈6 秒拍板 10 秒级,J2 失据);②反向陷阱——凭空给死「30 秒」且不标假设(一次未登记的替用户拍板)。只罚①不罚②会把另一半偏离合法化。",
"expected_deliverable": {
"时长段(合法形态一)": "数字目标时长 + 可接受区间,且若为自行拍板则进待补项并写「我按 X 理解,理由是 Y」",
"时长段(合法形态二)": "在「假设与待补项」显式登记时长缺口(写明「输入未给时长」)+ 向上游/用户上报该缺口(写明上报对象)",
"禁止形态": "静默缺席(无数字、未登记、未上报三者同时成立)-> CAP-6 最重档",
"下游咬合": "无论哪种形态,都给 h3-story 留出节拍量级依据(10s 级 3-5 拍 / 30s 级 6-10 拍)或明确的缺口上报"
},
"criteria": [
{ "id": "TS05-A1", "track": "A", "dim": "D5", "text": "时长处置为三种合法形态之一(数字+区间 / 登记缺口+上报 / 给死+标假设);静默缺席命中即 CAP-6", "check": "mechanical", "evidence": "A", "source": "S1 L35 + S1 L45 + S4 机械核验(全文 0 处时长)+ S10 L12(下游代拍板 D1)", "rationale": "F6 是 agent.md 硬要求与真实 run 全链绕过的正面冲突(C-3),判据必须把这颗齿留住", "points": 30 },
{ "id": "TS05-A2", "track": "A", "dim": "D5", "text": "若给死数字:标为假设并写「我按 X 理解,理由是 Y」(换算依据显式,如 20 字≈6 秒)", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S10 L12(写明回退路径的正样本)", "rationale": "自行拍板合法,未登记拍板不合法", "points": 15 },
{ "id": "TS05-A3", "track": "A", "dim": "D5", "text": "若登记缺口:写明「输入未给时长」+ 上报对象(上游/用户),两步缺一即缺陷", "check": "mechanical", "evidence": "A", "source": "R-3 第 2 款(benchmark §二,C-3 裁决条文)", "rationale": "只登记不上报 = 下游仍拿不到依据", "points": 15 },
{ "id": "TS05-A4", "track": "A", "dim": "D6", "text": "h3-validator J2 的判据基准被保住:或有时长数字、或显式声明 J2 需待缺口补齐后才有基准", "check": "human", "evidence": "A", "source": "S3 L63(J2:成片总时长与 brd 时长约束偏差 <=10%)+ S4(brd 无时长 -> J2 失据)", "rationale": "brd 的时长数字是 J2 的唯一基准", "points": 15 },
{ "id": "TS05-A5", "track": "A", "dim": "D1", "text": "段⑤(时长约束)内容存在:不因形态不同而豁免(登记缺口也是段⑤的一种内容形态)", "check": "mechanical", "evidence": "A", "source": "S1 L30「缺一段就是没写完」+ R-2 映射表段⑤行", "rationale": "按六段内容判而非按标题 grep(C-2 裁决)", "points": 15 },
{ "id": "TS05-A6", "track": "A", "dim": "D5", "text": "对称判罚:凭空给死无输入依据、且未标假设的数字时长 -> 判缺陷(与静默缺席同档扣分)", "check": "mechanical", "evidence": "C", "source": "构造性判罚(对称罚纪律)+ S1 L37(替用户拍板的每一处都要写 X/理由 Y);对应上级 C-3 裁决建议稿", "rationale": "只罚缺席不罚无依据给死,会把另一半偏离合法化", "points": 10 },
{ "id": "TS05-B1", "track": "B", "dim": "E1", "text": "时长缺口(若走登记形态)进「本次没做什么 / 已知缺口」,不静默省略", "check": "mechanical", "evidence": "A", "source": "S1 L62 + R-3 第 2 款", "rationale": "静默省略 = 0 分(E1 硬规则)", "points": 15 },
{ "id": "TS05-B2", "track": "B", "dim": "E2", "text": "缺口已上报:写明上报对象与补救路径;未上报 = 静默拍板", "check": "mechanical", "evidence": "A", "source": "S8(冲突/缺口上报正样本)+ S10 L12(下游被迫代拍板的反例)", "rationale": "不上报的直接后果是下游代拍板(F6 实证)", "points": 20 },
{ "id": "TS05-B3", "track": "B", "dim": "E3", "text": "自行拍板的时长显式标为「我的假设」,不写成用户给定事实;换算依据(字数->秒)标注为构造性推断", "check": "mechanical", "evidence": "A", "source": "S1 L44 + S10 L12(h3-story 把拍板写成 D1 并上报的正样本)", "rationale": "把假设洗成事实 = E3 核心违规", "points": 25 },
{ "id": "TS05-B4", "track": "B", "dim": "E4", "text": "零编造:不出现「用户要求 30 秒」这类把假设说成输入的句子", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L51-52", "rationale": "一票否决项;时长是最容易被洗成用户要求的字段", "points": "veto" },
{ "id": "TS05-B5", "track": "B", "dim": "E5", "text": "待补项写明:缺什么(时长)/ 谁补(用户或上游)/ 不补的后果(J2 无判据基准、下游被迫代拍板)", "check": "mechanical", "evidence": "A", "source": "benchmark §四 E5 + S10 L12(不补的真实后果实证)", "rationale": "后果写不出来 = 没想清楚影响面", "points": 20 }
]
},
{
"id": "TS-06",
"type": "UNSPECIFIED 主角:应登记待补项而不许编",
"difficulty": "中等",
"input": "玩家输入:「拍一条古装悬疑短片,主角是个捕快。」无外观、年龄、服装、性格描述,无参考图、无角色设定文档。",
"trap": "编造外貌/服装/性格(「捕快嘛,当然是青色官服、三缕短须、眼神锐利」);或主角段整段空白不登记;或把「查不出」说成「已核实主角为普通捕快形象」。",
"expected_deliverable": {
"主角条目": "已知项(身份:捕快)可溯源到输入原话;其余项写 UNSPECIFIED 并逐条列待补项",
"待补项": "每条写缺什么(外观/服装/性格)/ 建议谁补(用户或 h3-asset 基于事实生成)/ 不补的后果",
"下游依据": "给 h3-story 留出判定依据:哪些支柱无法绕开 -> blocked;哪些可绕开(不正面拍主角正脸)",
"零编造": "全 brd 无任何不可追溯的具体设定"
},
"criteria": [
{ "id": "TS06-A1", "track": "A", "dim": "D5", "text": "已知项与未知项分离:可溯源项进正文,查不出的写 UNSPECIFIED", "check": "mechanical", "evidence": "A", "source": "S1 L32「查不出的写 UNSPECIFIED 并列为待补项,不许编」", "rationale": "契约正源条款", "points": 20 },
{ "id": "TS06-A2", "track": "A", "dim": "D5", "text": "UNSPECIFIED 逐条列待补项(缺什么 / 谁补 / 后果),不是笼统一句「主角待定」", "check": "mechanical", "evidence": "A", "source": "S1 L37 + benchmark §四 E5", "rationale": "笼统待定 = 悬空占位的变体(S1 L43)", "points": 20 },
{ "id": "TS06-A3", "track": "A", "dim": "D5", "text": "零编造:无任何不可追溯的外观/服装/性格/关系设定(命中即 CAP-1 + E4 veto)", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L55 Never Rule(不替用户编造核心主体外观——那是 h3-asset 的活)", "rationale": "本场景的头号失败模式", "points": 25 },
{ "id": "TS06-A4", "track": "A", "dim": "D6", "text": "给下游留 blocked / 绕开判定依据:写明哪些支柱无法绕开(->h3-story blocked)、哪些可绕开", "check": "human", "evidence": "C", "source": "S2 L17-19(能绕开的绕开并写明绕法;绕不开且是叙事支柱的 blocked)+ S2 L45;构造性标注:h3-intake 层无直接判例(knowhow §6 局限 4),判据由 S1 条款 + S2 下游咬合结果支撑", "rationale": "下游咬合是这条契约存在的理由", "points": 15 },
{ "id": "TS06-A5", "track": "A", "dim": "D1", "text": "主角段内容存在且非空(含已知项 + UNSPECIFIED 标注),非空白也非纯占位", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L43", "rationale": "整段空白与悬空占位同等违规", "points": 10 },
{ "id": "TS06-A6", "track": "A", "dim": "D2", "text": "待补项每条带理由(为什么查不出 / 为什么不自行拍板)", "check": "mechanical", "evidence": "A", "source": "S1 L43(进待补项的除外,且必须带理由)", "rationale": "无理由的待补项 = 合法化的 TBD", "points": 10 },
{ "id": "TS06-B1", "track": "B", "dim": "E1", "text": "「本次没做什么」写明:未核实是否存在可用的角色设定文档 / 参考图(本单无输入则写明未做外部检索)", "check": "mechanical", "evidence": "A", "source": "S1 L62", "rationale": "留痕缺省 = UNSPECIFIED 不可信", "points": 15 },
{ "id": "TS06-B2", "track": "B", "dim": "E2", "text": "主角设定缺口显式上报(对象 + 补法),不静默把 UNSPECIFIED 留在正文等下游自己发现", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S2 L45(下游 blocked 口径引用 brd 的待补项)", "rationale": "缺口上报是上游的义务,不是下游的发现题", "points": 20 },
{ "id": "TS06-B3", "track": "B", "dim": "E3", "text": "每个主角属性标注来源三档之一(输入原话 / 我的假设 / UNSPECIFIED),无一档混用(如把 UNSPECIFIED 写成「已核实为普通捕快形象」)", "check": "mechanical", "evidence": "A", "source": "S1 L44(每个主角/场景要素都能追溯到原话,或明确标注为我的假设)", "rationale": "三档混用 = 洗白", "points": 25 },
{ "id": "TS06-B4", "track": "B", "dim": "E4", "text": "零编造:主角段任何具体形象描述都必须能在输入原话中找到(本单只有「捕快」一词)", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L55", "rationale": "一票否决项;本场景编造风险最高", "points": "veto" },
{ "id": "TS06-B5", "track": "B", "dim": "E5", "text": "待补项完整:外观/服装/性格逐条列出(不合并成一条「主角信息待补」),每条带理由", "check": "mechanical", "evidence": "A", "source": "S1 L37 + benchmark §四 E5", "rationale": "合并登记会让下游无法逐项判断能否绕开", "points": 20 }
]
},
{
"id": "TS-07",
"type": "F12 型:参考图与文字设定冲突未登记裁决",
"difficulty": "中等",
"input": "文字设定:「油灯是唯一光源」「中式古风恐怖」。参考图描述(随输入给定的图内容说明):「昏暗旧屋,五扇雕花屏风,圆石面桌上一支蜡烛,右侧罗汉床,墙上挂轴」。要求产出场景段与验收判据。",
"trap": "把蜡烛和油灯都写成发光源(v2 F3 同型跨 agent 判例);或静默删掉蜡烛(超出最小改动);或两者都保留但不登记冲突、不写裁决,让下游每个节点各自临场裁决。",
"expected_deliverable": {
"冲突登记": "冲突逐条登记:冲突双方内容(文字设定原文 + 参考图可见内容)都可回查",
"裁决": "以 brd 文字设定为准;最小改动 = 蜡烛按未点亮道具保留 + 写明下游勿画成光源(不是删蜡烛、不是改成双光源)",
"下游约束": "裁决写明对下游的具体约束(画蜡烛但不画光)",
"判据自洽": "裁决后「唯一光源」判据仍可机械判定(唯一发光体 = 油灯)"
},
"criteria": [
{ "id": "TS07-A1", "track": "A", "dim": "D5", "text": "冲突逐条登记:双方内容可回查(文字设定原文 + 参考图可见要素),不是一句「有冲突」", "check": "mechanical", "evidence": "A", "source": "S10 L12(D3:油灯唯一光源 vs 参考图桌上有蜡烛——以 brd 为准)+ S9 L44(冲突事实源头:圆石面桌上一支蜡烛)", "rationale": "不登记则下游每个节点各自临场裁决(F12)", "points": 20 },
{ "id": "TS07-A2", "track": "A", "dim": "D5", "text": "裁决采用最小改动:蜡烛保留但写明未点亮,不删除、不把两者都写成发光源", "check": "mechanical", "evidence": "B", "source": "S14 L220 benchmark_v2 D8 子块 A「最小改动」原则(「油灯是唯一光源」+ 场景有蜡烛 -> 蜡烛保留但写明未点燃)——B 级范式,且在 h3-video-prompt F3 被独立验证、跨 agent 复现(knowhow F12)", "rationale": "同一判例在两个 agent 的失败模式里都出现,跨 agent 复现", "points": 25 },
{ "id": "TS07-A3", "track": "A", "dim": "D6", "text": "裁决写明对下游的约束(勿画成光源,否则干扰唯一光源判据)", "check": "mechanical", "evidence": "A", "source": "S10 L12(D3 原文:请下游勿画成光源,否则干扰 S5(a))", "rationale": "裁决不落到下游动作上就只是声明", "points": 15 },
{ "id": "TS07-A4", "track": "A", "dim": "D2", "text": "裁决后验收判据自洽:「唯一光源」判据可机械判定(场景内发光体清单 = 油灯一项)", "check": "mechanical", "evidence": "A", "source": "S1 L38(措辞要能被逐条核对)+ S1 L58(不写不可判定的形容词)", "rationale": "裁决与判据不自洽 = 冲突被转移而不是被解决", "points": 15 },
{ "id": "TS07-A5", "track": "A", "dim": "D5", "text": "裁决进「假设与待补项」/冲突处置段并写「我按 X 理解,理由是 Y」", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S8 冲突①②(登记+上报的正样本形态)", "rationale": "替用户裁决必须可审计", "points": 15 },
{ "id": "TS07-A6", "track": "A", "dim": "D4", "text": "引用参考图时条目编号/名称与实际素材对应(编号语义一致,R-6)", "check": "mechanical", "evidence": "A", "source": "S10 L18 vs S4 L80(§7-③ 编号对语义错的判例,C-8)", "rationale": "引用漂移会让裁决指不到实际素材", "points": 10 },
{ "id": "TS07-B1", "track": "B", "dim": "E1", "text": "未核实事项留痕:参考图仅据给定描述判断,未实际打开图片核验(如本环境无法读图)须写明", "check": "mechanical", "evidence": "A", "source": "S1 L62", "rationale": "冲突判断的可信边界必须交代", "points": 15 },
{ "id": "TS07-B2", "track": "B", "dim": "E2", "text": "冲突显式上报:登记 + 上报对象 + 处置,不静默择一(真实 run 的 brd 用冲突①②②④表完成此动作)", "check": "mechanical", "evidence": "A", "source": "S8 冲突①② + S4 L78-81(4 条冲突/缺口处置表)", "rationale": "静默择一是 F12 的直接成因", "points": 20 },
{ "id": "TS07-B3", "track": "B", "dim": "E3", "text": "裁决标注来源分级:以 brd 文字设定为准是上游硬约束,参考图可见内容是事实输入——两档不混写", "check": "mechanical", "evidence": "B", "source": "S14 L220(冲突优先级:当前明确约束 > 参考图可见内容)——B 级范式格式;落点事实见 S10 L12", "rationale": "不分级会让下游以为参考图也能覆盖硬约束", "points": 25 },
{ "id": "TS07-B4", "track": "B", "dim": "E4", "text": "零编造:参考图可见内容只引用输入给定的图描述(蜡烛/屏风/罗汉床/挂轴),不虚构图上不存在的物件", "check": "mechanical", "evidence": "A", "source": "S9 L44 + S1 L44", "rationale": "一票否决项;「图上还有一盏壁灯」即编造", "points": "veto" },
{ "id": "TS07-B5", "track": "B", "dim": "E5", "text": "裁决理由完整:为什么以文字为准(上游硬约束优先)、为什么保留蜡烛(最小改动)逐条写", "check": "mechanical", "evidence": "A", "source": "S1 L37", "rationale": "理由缺失 = 裁决无法被复核", "points": 20 }
]
},
{
"id": "TS-08",
"type": "F1/F8 型:引擎 / 资源阻塞与业务 blocked 混同(阻塞三分层)",
"difficulty": "中等",
"input": "同一轮受理中依次出现三个事件,要求对每个写出受理处置:(a) run 日志显示本节点 boundary:missing=[agent.md, soul.md, .claude/skills/main_skill/SKILL.md],degraded=body_not_ready,状态链 fail -> blocked;(b) 玩家输入只有「随便拍个视频」;(c) run 跑到 2h 墙钟触发 budget-exhausted,token/调用数富余,run.status=blocked。",
"trap": "三层混同:把 (a) 当业务 blocked 上报玩家要输入(根因层错位,玩家补不了本体);把 (c) 当失败要求取消重跑(新 run 不继承旧节点产物,2h 墙钟必再撞);把 (b) 说成引擎/资源问题(把责任推给系统)。",
"expected_deliverable": {
"(a) 处置": "判为 L1 引擎级装配缺口:上报缺的本体件(agent.md / soul.md / SKILL.md)给 org 层 / agent-creator,不上报玩家要输入,不产出 brd",
"(b) 处置": "判为 L2 业务级 blocked:不硬造 brd,写明缺「想拍什么」、需要用户提供什么",
"(c) 处置": "判为 L3 资源级挂起:等人工 / 追加预算后续跑;明确说明不取消重跑(新 run 不继承旧节点产物)",
"报告形态": "三段分别命名层级(L1/L2/L3 或等价命名),不混用同一个 blocked 措辞"
},
"criteria": [
{ "id": "TS08-A1", "track": "A", "dim": "D2", "text": "(a) 判为 L1 引擎装配缺口:动作 = 上报缺的本体件,不上报玩家要输入", "check": "mechanical", "evidence": "A", "source": "S12(missing=[agent.md,soul.md,SKILL.md] degraded=body_not_ready,fail->blocked)+ S11 L13(CEO 排查:哨兵未除,missingBodyDefinition 判 body_not_ready)", "rationale": "agent 自判 blocked 判不到 L1(knowhow §4 C-5)", "points": 30 },
{ "id": "TS08-A2", "track": "A", "dim": "D5", "text": "(b) 判为 L2 业务 blocked:三要素齐(缺哪段 / 需要用户提供什么 / 不硬造)", "check": "mechanical", "evidence": "A", "source": "S1 L49-52", "rationale": "「随便拍个视频」连想拍什么都说不出来", "points": 25 },
{ "id": "TS08-A3", "track": "A", "dim": "D2", "text": "(c) 判为 L3 资源挂起:动作 = 续跑而非重跑,并说明理由(新 run 不继承旧节点产物、墙钟必再撞)", "check": "mechanical", "evidence": "A", "source": "S5 id:133(budget-exhausted,等人工不是失败)+ S5 id:138 / S11 L29(budget.extended wallclock_ms 后 blocked->running;不建议取消重跑)", "rationale": "blocked != 失败,重跑是最贵的错误处置", "points": 25 },
{ "id": "TS08-A4", "track": "A", "dim": "D1", "text": "三层各自命名且不混用:报告对 (a)(b)(c) 分别标注层级,未用同一个「blocked」措辞糊过", "check": "mechanical", "evidence": "A", "source": "knowhow §4 C-5(三类根因层不同,判据必须分层写)", "rationale": "混层的报告会触发三种完全不同的错误动作", "points": 20 },
{ "id": "TS08-B1", "track": "B", "dim": "E1", "text": "(a) 的缺口如实留痕:写明三件套缺件清单原文(不概括成「本体不全」)", "check": "mechanical", "evidence": "A", "source": "S12 output_excerpt(逐字列出三件)+ S1 L62", "rationale": "概括会让补件方不知道补哪件", "points": 15 },
{ "id": "TS08-B2", "track": "B", "dim": "E2", "text": "三种处置都写明上报对象与建议动作(L1->org 层补件 / L2->用户补输入 / L3->追加预算续跑),不静默挂起", "check": "mechanical", "evidence": "A", "source": "S11 L13(12:03 派单 agent-creator 的处置链)+ S5 id:138(续跑处置)", "rationale": "上报对象错 = 处置空转", "points": 20 },
{ "id": "TS08-B3", "track": "B", "dim": "E3", "text": "层级判定标注依据:每层判定引用对应的 run 证据(日志字段 / 输入形态 / 资源指标),不凭印象归层", "check": "mechanical", "evidence": "A", "source": "S12(boundary 字段)/ S1 L49(输入形态)/ S5 id:133(blockReason 字段)", "rationale": "无依据的归层与猜无异", "points": 25 },
{ "id": "TS08-B4", "track": "B", "dim": "E4", "text": "零编造:不虚构日志字段或预算数字(如编造「token 已耗尽」)", "check": "mechanical", "evidence": "A", "source": "S5 id:133 + S11 L29(实测:超的是 2h 墙钟,token/调用数都富余)", "rationale": "一票否决项;资源状态是最容易被想当然的字段", "points": "veto" },
{ "id": "TS08-B5", "track": "B", "dim": "E5", "text": "待补项/后续动作完整:L1 缺件清单、L2 所需输入、L3 续跑前置条件逐条列出", "check": "mechanical", "evidence": "A", "source": "benchmark §四 E5 + S12 / S1 L49 / S5 id:138", "rationale": "三种后续动作都要可执行", "points": 20 }
]
},
{
"id": "TS-09",
"type": "不可判定形容词当验收标准",
"difficulty": "中等",
"input": "玩家输入:「要拍得有质感、好看一点、氛围拉满。」其余信息齐(主角外观性格 / 场景 / 风格倾向 / 时长 15 秒 / 台词一句原句)。",
"trap": "双向陷阱:①把形容词原样抄进验收判据(「成片要有质感」当判据);②反向陷阱——因为形容词不可判就把风格段整个清空(玩家的风格意图丢失,下游没有风格依据)。",
"expected_deliverable": {
"验收判据": "零形容词判据;「有质感/好看/氛围拉满」不被写成 pass/fail 条件",
"风格段": "仍承载玩家意图:转译为可判定风格词与元素清单(如冷色调 / 低照度 / 实景质感参考作品),不因不可判而清空",
"标注": "不可判定部分显式标注「仅创作方向,不作验收标准」",
"待补项": "转译处逐条「我按 X 理解,理由是 Y」(把「氛围拉满」理解为低照度+慢节奏是一处拍板)"
},
"criteria": [
{ "id": "TS09-A1", "track": "A", "dim": "D2", "text": "验收判据段零形容词判据(「有质感/好看/氛围拉满」均未作 pass/fail 条件)", "check": "mechanical", "evidence": "A", "source": "S1 L38(不要写「要好看一点」「有质感」这种没法判定的形容词)+ S1 L58 Never Rule", "rationale": "本体两处独立禁止", "points": 30 },
{ "id": "TS09-A2", "track": "A", "dim": "D1", "text": "对称判罚:风格段仍承载玩家意图(可判定风格词 + 元素清单),未因形容词不可判而清空", "check": "mechanical", "evidence": "A", "source": "S1 L34(风格段:视觉风格 / 情绪基调 / 参考作品)+ S1 L30(缺一段就是没写完)", "rationale": "只罚「抄形容词」不罚「清空意图」会把另一半偏离合法化", "points": 25 },
{ "id": "TS09-A3", "track": "A", "dim": "D2", "text": "形容词被转译为可核对事实(风格词清单 / 参考作品 / 光照与色调要素),转译处可回查到原词", "check": "human", "evidence": "A", "source": "S1 L38 + S4(真实 brd 的可判定措辞形态)", "rationale": "转译是唯一既保留意图又可判定的路径", "points": 20 },
{ "id": "TS09-A4", "track": "A", "dim": "D5", "text": "不可判定部分显式标注「仅创作方向,不作验收标准」;转译处逐条进待补项带理由", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S1 L38", "rationale": "标注让下游知道哪部分不能拿来判 fail", "points": 15 },
{ "id": "TS09-A5", "track": "A", "dim": "D6", "text": "转译结果与 h3-story 风格承接咬合(下游能据此写风格段,不需要再问用户「拉满是什么意思」)", "check": "human", "evidence": "A", "source": "S2 L25-32(prd 五段映射)+ S1 L5(成功标准)", "rationale": "反向自测判据 5 的具体化(S1 L46)", "points": 10 },
{ "id": "TS09-B1", "track": "B", "dim": "E1", "text": "留痕:写明哪些玩家原词未进入验收判据及原因(不可判定),未静默丢弃", "check": "mechanical", "evidence": "A", "source": "S1 L62 + S1 L38", "rationale": "丢弃也要留痕,否则像遗漏", "points": 15 },
{ "id": "TS09-B2", "track": "B", "dim": "E2", "text": "若转译存在多种合理解读且影响成片(如「氛围拉满」-> 冷调还是暖调),上报请确认或写明默认假设,不静默择一", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S8(上报正样本)", "rationale": "风格歧义是临场裁决的重灾区", "points": 20 },
{ "id": "TS09-B3", "track": "B", "dim": "E3", "text": "转译出的每个风格词标注来源(玩家原词的近义转写 / 我的假设),不把转译结果写成玩家原话", "check": "mechanical", "evidence": "A", "source": "S1 L44", "rationale": "转写不标 = 把自己的解读洗成用户要求", "points": 25 },
{ "id": "TS09-B4", "track": "B", "dim": "E4", "text": "零编造:风格段新增的具体元素(如「冷色调」)都标为假设或可追溯到玩家原词,不冒充玩家给定", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L44", "rationale": "一票否决项", "points": "veto" },
{ "id": "TS09-B5", "track": "B", "dim": "E5", "text": "「氛围拉满」的每一次具体化都逐条「我按 X 理解,理由是 Y」", "check": "mechanical", "evidence": "A", "source": "S1 L37", "rationale": "一词多解必须逐条登记", "points": 20 }
]
},
{
"id": "TS-10",
"type": "C-8 型:跨节点条目编号引用的语义漂移",
"difficulty": "困难",
"input": "本 brd 含一节「已知缺口清单」共 5 条编号条目,其中 §7-③ 原文是「H3 平台渲染 POST 请求体形状未验证过」,§7-④ 原文是「H3 平台 10.0s 时长档未验证」。下游某节点交接里写「R3 H3 平台 10.0s 时长档未验证(brd §7-③)」。要求:产出 brd 的交叉引用段,并对该下游交接做复核。",
"trap": "只对编号字符串核对(「有 §7-③ 这个编号,引用成立」)而不复核语义——真实判例中编号对、内容错位;或反过来把下游引用当成对自己的改动依据,回改自己的条目内容。",
"expected_deliverable": {
"交叉引用段": "每条引用同时写出被引条目的实质内容(编号 + 内容摘要),使编号与语义可交叉核对",
"复核结论": "指出「下游写的 brd §7-③ 实为 §7-④」的语义错位,附双方原文定位",
"最小干预": "不因复核而改动本 brd 的条目内容(错位在引用方,不在被引方)",
"登记": "复核结论登记进冲突/缺口段并上报"
},
"criteria": [
{ "id": "TS10-A1", "track": "A", "dim": "D4", "text": "每条跨节点引用同时写出被引条目实质内容(编号与语义可交叉核对,R-6)", "check": "mechanical", "evidence": "A", "source": "S10 L18 vs S4 L80(「brd §7-③」编号对语义错的真实判例)+ knowhow §4 C-8", "rationale": "只对编号字符串的核对是假核对", "points": 30 },
{ "id": "TS10-A2", "track": "A", "dim": "D4", "text": "发现并登记语义错位:指出下游引用的「10.0s 时长档」实为 §7-④,附双方原文定位", "check": "mechanical", "evidence": "A", "source": "S10 L18(sdd 13:54 行)+ S4 L80(§7-③ 实际内容)", "rationale": "复核的价值就是抓出这种错位", "points": 25 },
{ "id": "TS10-A3", "track": "A", "dim": "D4", "text": "引用的文件 / 节点先核存在性(引用不存在文件 = CAP-2 同型)", "check": "mechanical", "evidence": "A", "source": "S6 L10 + R-5(benchmark §二)", "rationale": "存在性与语义复核是两道独立检查", "points": 20 },
{ "id": "TS10-A4", "track": "A", "dim": "D2", "text": "复核结论以可机械核对形式落盘(双方原文摘录 + 定位),不是一句「编号有误」", "check": "mechanical", "evidence": "A", "source": "S10 L18(真实判例的写法:同时给出被引编号与实际内容)", "rationale": "无定位的复核结论无法被第三方验证", "points": 15 },
{ "id": "TS10-A5", "track": "A", "dim": "D1", "text": "最小干预:不因复核结论回改本 brd 的条目内容(错位在引用方)", "check": "mechanical", "evidence": "C", "source": "构造性判据(最小改动原则在引用复核上的延伸)+ S10 L14(副本字节一致的对照思想)", "rationale": "回改会制造第二处不一致", "points": 10 },
{ "id": "TS10-B1", "track": "B", "dim": "E1", "text": "留痕:写明本次复核覆盖了哪几条引用、哪几条未核(如未拿到下游最终稿),不冒充全量复核", "check": "mechanical", "evidence": "A", "source": "S1 L62 + knowhow §6(查不到就写「未命中」的纪律)", "rationale": "部分复核冒充全量 = 洗白", "points": 15 },
{ "id": "TS10-B2", "track": "B", "dim": "E2", "text": "语义错位显式上报给引用方(不是只在自己文档里改对),不静默", "check": "mechanical", "evidence": "A", "source": "S8(上报动作正样本)+ S10 L13/L15/L17(错位措辞沿链传播 4 轮的反面教训)", "rationale": "不上报则错位继续传播", "points": 20 },
{ "id": "TS10-B3", "track": "B", "dim": "E3", "text": "复核结论标注证据级别(A:双方原文逐字;构造性推断须标注),不把推断写成已核实", "check": "mechanical", "evidence": "A", "source": "S11 L13(orggraph-plan §7 的伪「已核实」反例,F2:权威文件里的声明也要实测复核)", "rationale": "书面声称 != 已核实", "points": 25 },
{ "id": "TS10-B4", "track": "B", "dim": "E4", "text": "零编造:引用与摘录逐字来自原文,无转述冒充原文、无编造条目内容", "check": "mechanical", "evidence": "A", "source": "S4 L80 + S10 L18", "rationale": "一票否决项;摘录是最容易「凭印象补全」的地方", "points": "veto" },
{ "id": "TS10-B5", "track": "B", "dim": "E5", "text": "若为统一编号而自行改写条目(替下游拍板),逐条进待补项带理由", "check": "mechanical", "evidence": "A", "source": "S1 L37", "rationale": "改编号是拍板,不是整理", "points": 20 }
]
},
{
"id": "TS-11",
"type": "台词口径唯一 + 硬输入透传(F3 A.一致性 正反两向)",
"difficulty": "中等",
"input": "完整输入含台词原句「雨太大了我先躲一下。」。下游 contract.json 的 success_metrics 已写「与原台词语义一致」;brd 草稿侧写「原台词逐字见 §6-A,唯一基准」。成片音轨将经 ASR 转写后被 h3-validator 以 J6 判定(音轨转写与 brd/prd 台词逐字一致)。要求落定台词验收口径。",
"trap": "两种口径并存(brd 逐字 / contract 语义),validator 届时任选宽松的一个判 pass(S6 L9 判例原文);或为迁就 contract 把 brd 也改成「语义一致」而丢掉 J6 需要的逐字基准;或反向——无台词场景不写「无台词」留空白。",
"expected_deliverable": {
"口径唯一": "brd 与 contract 同一验收点只有一种写法且两处同文(逐字 + 归一化口径,或语义 + 可判定阈值)",
"J6 落点": "台词作为唯一基准逐字透传,附归一化口径(含标点/去标点两套字数)",
"无台词形态": "若台词为空,出现精确字串「无台词」(S1 L36 给死的措辞)",
"拍板登记": "若 90% 一致率这类阈值是自行设定,进待补项带理由"
},
"criteria": [
{ "id": "TS11-A1", "track": "A", "dim": "D2", "text": "同一验收点在 brd 与 contract 只有一种口径且两处同文(validator 无从任选宽松的一个)", "check": "mechanical", "evidence": "A", "source": "S6 L9(A.一致性:逐字 vs 语义两个门槛,validator 届时任选宽松的一个判 pass)+ S4 L38(本节与 contract success_metrics 逐条同文)", "rationale": "双口径是 F3 打回的第一条,也是唯一被「同文」修复的一条", "points": 30 },
{ "id": "TS11-A2", "track": "A", "dim": "D6", "text": "台词作为唯一基准逐字透传,h3-validator J6 有落点", "check": "human", "evidence": "A", "source": "S3 L67(J6:音轨转写与 brd/prd 台词逐字一致)+ S10 L12(唯一台词逐字在 B4 的正面实证)", "rationale": "J6 是 brd 台词条目的唯一消费方", "points": 20 },
{ "id": "TS11-A3", "track": "A", "dim": "D2", "text": "选逐字则附归一化前置条件(双侧同去标点);选语义则给可判定阈值——二选一后不残留另一种口径", "check": "mechanical", "evidence": "A", "source": "S7 L19(归一化前置)+ S6 L18 remediation 第 1 条(逐字包含 + 字符一致率 >=90% 的可度量写法)", "rationale": "口径选定后残留旧口径 = 双口径复活", "points": 20 },
{ "id": "TS11-A4", "track": "A", "dim": "D1", "text": "台词为空时出现精确字串「无台词」,不留空白也不写「暂无」", "check": "mechanical", "evidence": "A", "source": "S1 L36(没有就明写「无台词」)+ S2 L31(brd 说「无台词」就明写「全片无台词」)", "rationale": "这是六段里唯一保留字面判法的点(benchmark §〇 0.3 例外)", "points": 15 },
{ "id": "TS11-A5", "track": "A", "dim": "D3", "text": "若写一致率阈值,分母口径与字数可独立复算(去标点 / 含标点两套)", "check": "mechanical", "evidence": "A", "source": "S7 L11 + S9 L36", "rationale": "阈值不可复算 = 无阈值(CAP-5 同型)", "points": 15 },
{ "id": "TS11-B1", "track": "B", "dim": "E1", "text": "留痕:写明 ASR 实测一致率只能事后测、brd 侧只定口径不测值,不冒充已验证", "check": "mechanical", "evidence": "A", "source": "S1 L62", "rationale": "口径与实测是两件事", "points": 15 },
{ "id": "TS11-B2", "track": "B", "dim": "E2", "text": "brd 与 contract 的口径分歧(逐字 vs 语义)被显式上报并给出统一方案,不静默任留一种", "check": "mechanical", "evidence": "A", "source": "S6 L9(分歧被 verify 抓出的反例)+ S8(上报正样本)", "rationale": "分歧不报 = 下游撞上", "points": 20 },
{ "id": "TS11-B3", "track": "B", "dim": "E3", "text": "统一后的口径显式标注「两处同文」的对照关系(brd §3-S3 == contract success_metrics[2]),不写成口头约定", "check": "mechanical", "evidence": "A", "source": "S4 L38(本节与 contract success_metrics 逐条同文——上一轮 A.一致性被打回的教训)+ S7 L11(attempt-1 逐条对号的复核)", "rationale": "同文关系可机械 diff 才算修好", "points": 25 },
{ "id": "TS11-B4", "track": "B", "dim": "E4", "text": "零编造:台词原文逐字来自输入,无改写、无补标点、无翻译", "check": "mechanical", "evidence": "A", "source": "S1 L36 + S3 L67(J6 逐字一致)", "rationale": "一票否决项;台词是全链最不容许失真的字段", "points": "veto" },
{ "id": "TS11-B5", "track": "B", "dim": "E5", "text": "若为迁就 ASR 能力而放宽为「语义一致」,该放宽进待补项写「我按 X 理解,理由是 Y」并同步到 contract", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S6 L9(放宽正是被打回的形态)", "rationale": "放宽是重大拍板,必须登记", "points": 20 }
]
},
{
"id": "TS-12",
"type": "悬空占位与待补项边界(对称判罚)",
"difficulty": "边界",
"input": "玩家输入给了:主角外观(穿米色风衣的中年女人)与身份(深夜加班回家),场景地点(老公寓楼道)与风格(写实);天气与具体时间未知;主角性格未知;台词原句「谁在那儿?」。要求产出六段 brd。",
"trap": "三种对称失败:①正文写「天气:待定」「性格:TBD」悬空占位;②把已知项(外观/地点/风格)也塞进待补项(过度登记,下游拿不到已知信息);③未知项进了待补项但不带理由。",
"expected_deliverable": {
"正文": "已知项全部进正文对应段(主角/场景/风格),零孤立 TBD/待定/TODO",
"待补项": "未知项(天气/时间/性格)逐条列出,每条带「我按 X 理解,理由是 Y」或缺什么+谁补",
"边界正确": "已知项不被降级进待补项;未知项不被写死成编造值",
"下游咬合": "写明哪些待补项会阻断下游(->建议 blocked)、哪些可绕开(如天气可由场景氛围替代)"
},
"criteria": [
{ "id": "TS12-A1", "track": "A", "dim": "D1", "text": "正文零孤立悬空占位(TBD/待定/TODO/暂缺 不出现在六段正文)", "check": "mechanical", "evidence": "A", "source": "S1 L43 判据 2(六段里没有 TBD/待定 悬空占位)", "rationale": "悬空占位 = 必然被下游追问", "points": 30 },
{ "id": "TS12-A2", "track": "A", "dim": "D5", "text": "未知项(天气/时间/性格)逐条进待补项,每条带理由或缺什么+谁补", "check": "mechanical", "evidence": "A", "source": "S1 L32(查不出的写 UNSPECIFIED 并列为待补项)+ S1 L43(进待补项的除外,且必须带理由)+ S1 L37", "rationale": "登记不等于合法化,登记+理由才算", "points": 25 },
{ "id": "TS12-A3", "track": "A", "dim": "D5", "text": "对称判罚:已知项(外观/身份/地点/风格/台词)未被错误降级进待补项(过度登记也是缺陷)", "check": "mechanical", "evidence": "C", "source": "构造性判罚(对称罚纪律:判别样本要罚对称方向的偏离)+ S1 L44(可追溯项应进正文)+ S1 L30(缺一段就是没写完)", "rationale": "只罚悬空占位不罚过度登记,会让已知信息被藏起来", "points": 20 },
{ "id": "TS12-A4", "track": "A", "dim": "D1", "text": "六段内容不因待补而缺失:登记缺口也算段内容(缺段 != 已登记,按 R-2 内容判法)", "check": "mechanical", "evidence": "A", "source": "S1 L30 + R-2 映射表(benchmark §二)", "rationale": "「我把天气放待补项了」不能替代场景段存在", "points": 15 },
{ "id": "TS12-A5", "track": "A", "dim": "D6", "text": "对每个待补项给出下游影响判断:是否阻断(->建议 blocked)/ 可绕开(->写明绕法)", "check": "human", "evidence": "C", "source": "S2 L17-19 + S2 L45(下游 blocked 口径);构造性标注:brd 层直接判例为零样本(knowhow §6 局限 4),由 S1 条款 + S2 下游咬合支撑", "rationale": "待补项不带影响判断 = 下游自己做发现题", "points": 10 },
{ "id": "TS12-B1", "track": "B", "dim": "E1", "text": "留痕:写明天气/时间/性格是输入未给(不是没查到),未静默当作已定", "check": "mechanical", "evidence": "A", "source": "S1 L62 + S1 L43", "rationale": "「未给」与「没查到」是两种不同的缺口", "points": 15 },
{ "id": "TS12-B2", "track": "B", "dim": "E2", "text": "若自行给天气/时间定了默认值(如「深夜」),显式上报该拍板,不静默择一", "check": "mechanical", "evidence": "A", "source": "S1 L37 + S8(上报正样本)", "rationale": "默认值是拍板不是默认无责", "points": 20 },
{ "id": "TS12-B3", "track": "B", "dim": "E3", "text": "已知项标注「输入原话可溯」,未知项标注 UNSPECIFIED/待补——三档标注不混用", "check": "mechanical", "evidence": "A", "source": "S1 L44 + S1 L32", "rationale": "三档混用是洗白的基础形态", "points": 25 },
{ "id": "TS12-B4", "track": "B", "dim": "E4", "text": "零编造:未知项未被写成具体值(「下雨的深夜」若非玩家原话即编造)", "check": "mechanical", "evidence": "A", "source": "S1 L32 + S1 L51-52", "rationale": "一票否决项", "points": "veto" },
{ "id": "TS12-B5", "track": "B", "dim": "E5", "text": "每个待补项写「我按 X 理解,理由是 Y」或「缺什么+谁补+后果」,无裸条目", "check": "mechanical", "evidence": "A", "source": "S1 L37 + benchmark §四 E5", "rationale": "裸条目 = 合法化的 TBD", "points": 20 }
]
}
],
"known_bad_expectations": {
"purpose": "已知坏样本预期得分自检段(Digest 预计算,Evolve 轮复核)。纪律:禁止只测诱导路径——每个维度都要有能拿低分的真实坏样本;已知坏样本拿高分 = 判据缺齿(h3-video-prompt v5 教训)。",
"entries": [
{ "scenario": "TS-01", "bad_sample": "把「10 秒左右」写成「时长适中」,台词改写为「雨太大我先躲一下」(去句号),并自行补了一个「巷口出现一只黑猫」的氛围设定", "expected_A": "56", "expected_B": "0", "triggered": ["CAP-3 无(时长未涉数字核算,属 D1 缺数字)", "D1 扣(段⑤非数字)", "D6 扣(台词失真)", "E4 veto(黑猫为编造)"], "note": "E4 一票否决使 B 轨 0;A 轨 D1/D3/D6 失分,总分低于 90 线" },
{ "scenario": "TS-02", "bad_sample": "产出一份完整 brd:主角「戴帽子的神秘年轻人」,场景「霓虹街头」,并在文末附一句「如需修改请联系」", "expected_A": "30", "expected_B": "0", "triggered": ["应 blocked 硬造 brd -> 最重档(§七 第 2 款)", "CAP-1 编造主角设定 ->60", "E4 veto -> B 0"], "note": "硬造 brd 是本场景唯一绝对错误;A 轨取 CAP-1 上限 60 与维度失分后的较低值" },
{ "scenario": "TS-03", "bad_sample": "验收段写「S3/S4 的具体阈值与权重由 goal-benchmark 定」「两个主体均出现」「validator 到时候看一眼」", "expected_A": "25", "expected_B": "55", "triggered": ["CAP-2 引用不存在节点 ->70", "CAP-5 无阈值/推给裁量 ->75", "最终 = min(加权, 70, 75)", "E2 扣(缺口未上报)", "E5 扣(自行设阈值未登记)"], "note": "B 轨仍有分(未编造事实),但 E2/E5 失分使其低于 85 线 -> 双轨均不放行" },
{ "scenario": "TS-04", "bad_sample": "照抄「27 字一致率 >=90%」,针短语保留逗号,未写分母口径", "expected_A": "30", "expected_B": "60", "triggered": ["CAP-3 数字未核算 ->80", "CAP-4 字面永假 ->80", "D3 大幅失分", "E3 扣(口径未标注)"], "note": "两道封顶同时触发;B 轨 E3 失分后低于 85 线" },
{ "scenario": "TS-05", "bad_sample": "六段齐全但无段⑤时长,待补项只写了「无」;或写「目标 30 秒」且无任何假设标注", "expected_A": "45", "expected_B": "55", "triggered": ["CAP-6 时长静默缺席 ->75", "对称判罚(无依据给死)同样判缺陷", "E2 扣(缺口未上报)", "E3 扣(假设未标注)"], "note": "两种方向的坏样本预期得分同档(对称罚落地的验证点)" },
{ "scenario": "TS-06", "bad_sample": "主角段写「捕快,约三十岁,穿青色官服,佩绣春刀,性格沉稳寡言」", "expected_A": "30", "expected_B": "0", "triggered": ["CAP-1 编造主角外观 ->60", "E4 veto -> B 0", "D5/D1 失分"], "note": "最典型的本体两处禁止行为(S1 L32 + L55)" },
{ "scenario": "TS-07", "bad_sample": "场景段同时写「油灯与蜡烛都亮着,火光摇曳」,未登记冲突", "expected_A": "35", "expected_B": "40", "triggered": ["冲突未登记(D5 扣)", "裁决违反最小改动(D5 扣)", "唯一光源判据自相矛盾(D2 扣)", "E2 大幅扣(静默择一 -> 0)", "E4 未触发(物件均来自输入)"], "note": "B 轨 E2 直接 0 分(满分 20 的项),低于 85 线" },
{ "scenario": "TS-08", "bad_sample": "对 (a) 上报「请玩家提供本体文件」,对 (c) 写「建议取消重跑」", "expected_A": "30", "expected_B": "50", "triggered": ["三层混同(§七 第 4 款判缺陷)", "D2/D1 大幅失分", "E3 扣(归层无依据)"], "note": "混层不触发 CAP(不是编造),但 A 轨必低于 90;B 轨低于 85" },
{ "scenario": "TS-09", "bad_sample": "验收段写「成片要有质感、氛围拉满」,风格段空白", "expected_A": "25", "expected_B": "55", "triggered": ["D2 形容词进验收段(-4 档)", "对称判罚:风格段清空(D1 扣)", "E1/E5 扣(丢弃未留痕)"], "note": "两个方向的偏离都失分" },
{ "scenario": "TS-10", "bad_sample": "交叉引用段只写「详见 brd §7-③」,复核结论写「编号无误」", "expected_A": "30", "expected_B": "50", "triggered": ["R-6 未落(只对编号字符串)", "语义错位未抓出(D4 扣)", "E3 扣(复核结论无证据级别)"], "note": "假核对的典型形态" },
{ "scenario": "TS-11", "bad_sample": "brd 写「原台词逐字见 §6-A,唯一基准」,contract 写「与原台词语义一致」,两处都保留", "expected_A": "40", "expected_B": "55", "triggered": ["双口径并存(D2 重扣,S6 L9 判例原文)", "E2 扣(分歧未上报)", "E3 扣(同文关系未标注)"], "note": "这就是 attempt-0 的真实失败形态(F3 第一条)" },
{ "scenario": "TS-12", "bad_sample": "正文写「天气:待定」「性格:TBD」,并把「米色风衣」也列进待补项", "expected_A": "35", "expected_B": "55", "triggered": ["悬空占位(D1 重扣)", "对称判罚:已知项过度登记(D5 扣)", "E5 扣(裸条目)"], "note": "双向偏离同时出现的坏样本" }
]
},
"acceptance_selfcheck": {
"scenario_count": 12,
"criteria_total": 129,
"criteria_track_A": 69,
"criteria_track_B": 60,
"mechanical_criteria": 122,
"human_criteria": 7,
"evidence_level_distribution_by_criteria": { "A": 121, "B": 2, "C": 6 },
"evidence_distribution_note": "A=仓库一手(S1-S13,121 条);B=判据写法范式(S14/S15,2 条:TS-07-A2 最小改动、TS-07-B3 冲突优先级——均为跨 agent 复现判例,已标注 B 级且在 benchmark §一 防污染条款下使用);C=构造性设计(6 条,均在 source 内显式标注构造理由);无 D 级",
"evidence_level_distribution": "见下方 computed 字段",
"points_selfcheck": {
"rule_A": "每场景 A 轨判据 points 总和 = 100(场景内分配);维度分 = 该维场景内命中/满 *100;A 轮总分 = sum(维度分*权重)/100 再取 CAP 上限最小值",
"rule_B": "每场景 B 轨判据 points 总和 = 100(E1 15 + E2 20 + E3 25 + E4 20[veto] + E5 20)",
"verified_by": "verify_digest.py(附汇报,可重跑)"
},
"e4_veto_in_every_scenario": true,
"human_check_criteria": "见下方 computed 字段",
"open_items": [
"TS-06-A4 / TS-12-A5 的判据主体无 h3-intake 层直接 run 判例,按 C 级构造(S1 条款 + S2 下游咬合支撑),已在判据 source 内显式标注 —— 待 master 复核",
"B 轨权重 15/20/25/20/10 与 CAP 上限 60/70/75/80/80/75 为 C 级构造,无 run 判例支撑具体数值,Evolve 轮拿真实评分分布后回填校准",
"R-2 残留不确定:真实 run 产出方是 work graph 内部节点 h3-intake-brd(brd-planning),其本体未读,若另有契约条款需下轮补核"
],
"pending_items": "零待议(判据侧裁决 R-1…R-7 已按上级指示落成明确条文;本体侧修订需求 4 条已登记进 benchmark §十 第 7 款,不在本轮范围)"
},
"coverage_summary": {
"dimensions_covered": { "D1": ["TS-01","TS-05","TS-06","TS-08","TS-09","TS-10","TS-12"], "D2": ["TS-01","TS-02","TS-03","TS-04","TS-06","TS-07","TS-08","TS-09","TS-10","TS-11","TS-12"], "D3": ["TS-01","TS-04","TS-11"], "D4": ["TS-01","TS-03","TS-07","TS-10"], "D5": ["TS-01","TS-02","TS-04","TS-05","TS-06","TS-07","TS-08","TS-09","TS-12"], "D6": ["TS-01","TS-05","TS-06","TS-07","TS-09","TS-11"], "D7": ["TS-01"] },
"dimension_weight_sum": 100,
"e_items_covered": { "E1": "TS-01..TS-12 全部", "E2": "TS-01..TS-12 全部", "E3": "TS-01..TS-12 全部", "E4": "TS-01..TS-12 全部(veto)", "E5": "TS-01..TS-12 全部" },
"failure_paths_covered": ["F1", "F2", "F3", "F4", "F5", "F6", "F7", "F8", "F9", "F11", "F12"],
"failure_paths_not_covered_note": "F10(采样判据须连同上游出现时序核)属 h3-story / h3-validator 侧判据(S5 id:68 是 h3-story-validator 的交接),不在 brd 评分对象内 —— brd 侧只能要求「采样规则与上游出现时序一起核」(落进 TS-03-A2 的采样规则要求);登记为未覆盖并说明理由。F2(伪已核实)落进 TS-10-B3,F9(依据缺失降级通道)落进 TS-03-B1,F1(body_not_ready)落进 TS-08-A1"
}
}