B 线探针 · H3 对白协议重渲(4 镜明牌判定)

左 = 0910 三臂批 LXV 生产对照(prompt 含「Off-screen narration, characters do not speak」免责尾缀,L1 时代遗产); 右 = 本探针(删免责尾缀 + <d>[ZH]台词</d>(S#,画内角色,h3_lock_source,唇形窗) 对白标签 + 旁白剥离音频窗)。 同 seed 990202 / 同首尾锚 / 同分辨率时长(FastH3 5步 vs LXV 9步,0910 盲测质量平、最快)。 agent 自评 ≠ 终审。
判读:0.25× 慢放看右侧说话角色嘴部是否随台词音节开合、口型是否给到正确的角色(S1/S2 编号见每镜「驱动对白」行); 切 1× 听台词-口型时轴是否对齐。左侧为对照——预期左侧嘴部闭/随机、右侧开口构音(Phase 0→ARMA 已证的方向)。 判定口径:✅通过=说话方+时轴双对;🔧需修=开口但对错人/错时轴;❌否决=嘴不动或画面崩。

s011_p1

FastH3 5步 · T8 AudioLock · seed 990202 · H3对白协议 + 旁白剥离音频 渲染PASS
原 LXV 镜(39 批生产对照 · 无对白协议)
v1 探针镜(说话人=算法原判)
v1 驱动对白:「阁下可是黑头前辈」(螳螂武士, 47.04-49.43s)
🔍 验收关联:A 线把「阁下可是黑头前辈」判给螳螂(pyannote 归甲虫簇、f0=122 落甲虫带,但 s012 甲虫答「正是」Arm A 实锤 → 问句结构判螳螂,c=0.55,HITL 首选项)。本镜画面只有螳螂——若判错说话人,口型错位会在此镜放大。
探针 prompt 对白段(点击展开)
Dialogue: the following in-frame characters speak, mouths visibly articulating in exact sync with the locked source audio: [ZH] 阁下可是黑头前辈(S1,画内螳螂武士,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 0.17-2.56 秒). Speaker numbering: no off-screen narrator drives any mouth in this shot; S1 = the green praying mantis warrior (thin limbs, white bandage wraps on forearms). Keep each character's appearance exactly as described above; only their mouths move with the dialogue.
判定(s011_p1):

s012_p1

FastH3 5步 · T8 AudioLock · seed 990202 · H3对白协议 + 旁白剥离音频 渲染PASS
原 LXV 镜(39 批生产对照 · 无对白协议)
v1 探针镜(说话人=算法原判)
v1 驱动对白:「辈」(螳螂武士, 49.43-50.44s) / 「正是」(独角仙武士, 50.44-51.10s)
🔍 验收锚(A 线):「正是」= 独角仙(Arm A 画内实锤)。本镜同时验证问句尾「辈」(螳螂 S1) + 答「正是」(甲虫 S2) 的双说话人切换。
探针 prompt 对白段(点击展开)
Dialogue: the following in-frame characters speak, mouths visibly articulating in exact sync with the locked source audio: [ZH] 辈(S1,画内螳螂武士,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 0.00-1.01 秒) [ZH] 正是(S2,画内独角仙武士,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 1.01-1.67 秒). Speaker numbering: no off-screen narrator drives any mouth in this shot; S1 = the green praying mantis warrior (thin limbs, white bandage wraps on forearms); S2 = the rhino beetle warrior (glossy reddish-brown armored shell, large double-pronged horn, beige bandages on forearms). Keep each character's appearance exactly as described above; only their mouths move with the dialogue.
判定(s012_p1):

s017_p1

FastH3 5步 · T8 AudioLock · seed 990202 · H3对白协议 + 旁白剥离音频 渲染PASS
原 LXV 镜(39 批生产对照 · 无对白协议)
v1 探针镜(说话人=算法原判)
v1 驱动对白:「阿毛你不吃了吗」(独角仙武士, 63.75-65.01s) / 「我吃饱了看拳」(毛毛虫小孩, 65.01-67.33s)
🔍 双说话人镜(最难):甲虫「阿毛你不吃了吗」(S1) + 孩子「我吃饱了看拳」(S2),7.2s 长窗,前 3.6s 无对白。
探针 prompt 对白段(点击展开)
Dialogue: the following in-frame characters speak, mouths visibly articulating in exact sync with the locked source audio: [ZH] 阿毛你不吃了吗(S1,画内独角仙武士,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 3.62-4.88 秒) [ZH] 我吃饱了看拳(S2,画内毛毛虫小孩,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 4.88-7.20 秒). Speaker numbering: no off-screen narrator drives any mouth in this shot; S1 = the rhino beetle warrior (glossy reddish-brown armored shell, large double-pronged horn, beige bandages on forearms); S2 = the small orange-yellow fuzzy caterpillar child. Keep each character's appearance exactly as described above; only their mouths move with the dialogue.
判定(s017_p1):

s023_p1

FastH3 5步 · T8 AudioLock · seed 990202 · H3对白协议 + 旁白剥离音频 渲染PASS
原 LXV 镜(39 批生产对照 · 无对白协议)
v1 探针镜(说话人=算法原判)
v2 重渲(说话人修复 · 螳螂「前辈够了没有」+甲虫「阿毛」双标签)
v1 驱动对白:「前辈够了没有阿毛」(独角仙武士, 77.75-80.43s)
v2 驱动对白(说话人修复):「前辈够了没有」(螳螂武士) / 「阿毛」(独角仙武士)
🔍 0911 Kai reject 定谳+倒查:v1 单标签「甲虫全句」错误——镜头口型实锤(78.0/78.5s 帧字幕「前辈/够够没有」+螳螂张嘴、甲虫闭嘴啃番茄)=「前辈够了没有」是螳螂说的,「阿毛」(79.9s 甲虫张嘴)才是甲虫。根因:pyannote 把螳螂/甲虫并成混簇 S04(两角色 f0 122/129Hz 声学碰撞不可分)+纯音频轴无镜头口型信号。v2 已按镜头口型仲裁修复(螳螂+甲虫双标签)重渲,本页第三列对比。音色 f0 实测:本句 129.6Hz 落甲虫带但属两角色重叠区,f0 单维度在本片不可判。
探针 prompt 对白段(点击展开)
Dialogue: the following in-frame characters speak, mouths visibly articulating in exact sync with the locked source audio: [ZH] 前辈够了没有阿毛(S1,画内独角仙武士,h3_lock_source,贴合台词语气的自然口吻; 唇形对齐窗约 0.58-3.26 秒). Speaker numbering: no off-screen narrator drives any mouth in this shot; S1 = the rhino beetle warrior (glossy reddish-brown armored shell, large double-pronged horn, beige bandages on forearms). Keep each character's appearance exactly as described above; only their mouths move with the dialogue.
判定(s023_p1 v2):