docs(jimeng): 添加即梦文生视频提示词评估框架和完整技能文档

- 新增 evaluation-framework.md 用于系统性评估技能产出质量
- 新增 jimeng-prompt-text2image SKILL.md 完整技能定义文档
- 新增 jimeng-prompt-text2video SKILL.md 完整技能定义文档
- 添加 video-core-methodology.md 视频核心构建方法论
- 添加 video-validation-checklist.md 视频验证清单
- 添加 video-writing-rules.md 视频专用写作规则文档

skills/jimeng-skills/jimeng-prompt-text2image/SKILL.md

@@ -99,8 +99,6 @@ Each component is optional. Select and weight components based on the scenario:
 3. **Aspect ratio drives composition** — a prompt written for 16:9 will compose differently than the same prompt at 9:16. Always confirm the intended aspect ratio
 4. **Style keywords can dominate** — strong style keywords (cyberpunk, 水墨, Pixar) can override other elements. Place them carefully in the prompt order
 5. **Prompt language** — 即梦 works best with Chinese prompts. If the user provides English, translate and adapt rather than directly using it
-6. **⚠️ Never write hex color values into prompts** — The model treats `#ff2121`, `#9d2933` etc. as literal text to render in the image, NOT as color instructions. Always use Chinese color names only: \"大红\", \"胭脂\", \"桃红\" — never \"大红 #ff2121\" or any form with hex codes
-7. **Color library is for YOUR reference only** — Load `color-library/chinese-traditional.md` to FIND the right color name to write into the prompt. Do NOT copy hex codes into the output prompt text
 6. **⚠️ Never write hex color values into prompts** — The model treats `#ff2121`, `#9d2933` etc. as literal text to render in the image, NOT as color instructions. Always use Chinese color names only: "大红", "胭脂", "桃红" — never "大红 #ff2121" or any form with hex codes
 7. **Color library is for YOUR reference only** — Load `color-library/chinese-traditional.md` to FIND the right color name to write into the prompt. Do NOT copy hex codes into the output prompt text
 

skills/jimeng-skills/jimeng-prompt-text2video/SKILL.md

@@ -92,11 +92,15 @@ The key difference from text-to-image prompting:
 → Load `rules/video-core-methodology.md` for component-by-component build guide + presentation format
 
 ### Step 4: Apply video writing rules
-→ Load `rules/video-writing-rules.md` for all 7 rules (motion, camera, duration, light, etc.)
+→ Load `rules/video-writing-rules.md` for all 10 rules (motion, camera, duration, light, model-matching, differentiation, action layers)
 
 ### Step 5: Validate
 → Load `rules/video-validation-checklist.md` and run through all checks
 
+### Step 6 (User assessment): Evaluate prompt output
+When the user provides an existing prompt output and asks for evaluation ("评估""检测命中率""看看优化方向"):
+→ Load `references/evaluation-framework.md` for the systematic 6-dimension assessment methodology
+
 ## Gotchas
 
 1. **Action must be explicit** — 即梦 cannot infer motion from static description. "一个人在街上" = static shot
@@ -107,9 +111,11 @@ The key difference from text-to-image prompting:
 6. **First-time web authorization** — some models require browser auth before first use
 7. **Character consistency not guaranteed** — same face across segments not reliable
 8. **Camera + complex motion = risk** — prioritize one over the other
-9. **Model version matters** — 视频3.0 Pro uses natural language, Seedance 2.0 needs structured camera direction. 智能多帧 requires uploading frames + per-frame duration settings
-10. **⚠️ Never write hex color codes in video prompts** — Same as image prompts: `#RRGGBB` values get rendered as text in the video frames, NOT as color instructions. Use Chinese color names only
-10. **⚠️ No hex color values** — Same rule as image prompts: never write `#ff2121` into video prompts. Use Chinese color names only
+9. **Model-prompt mismatch is the #1 output killer (Rule 8)** — long descriptive prose → 视频3.0 Pro; structured short sentences → Seedance 2.0. Mismatching causes detail loss and poor generation. **This is the single most common error in real usage.**
+10. **Action layer count by model (Rule 10)** — Seedance 2.0 can't handle 4+ action layers reliably. Cap at 3. 视频3.0 Pro can handle 4-5.
+11. **Multi-scheme differentiation is non-optional (Rule 9)** — When providing 2+ alternatives, verify visual distinction. At least 2 dimensions must differ (运镜/动作密度/景别/节奏/场景). Similar schemes waste user's time.
+12. **Light must change, not just exist (Rule 5)** — "洒入" is a static snapshot. Use "缓缓流动""逐渐变亮""光影移动" to encode time passage.
+13. **⚠️ Never write hex color codes in video prompts** — Same as image prompts: `#RRGGBB` values get rendered as text in the video frames, NOT as color instructions. Use Chinese color names only
 
 ## Available Resources
 
@@ -117,11 +123,12 @@ The key difference from text-to-image prompting:
 |----------|-------------|--------------|
 | `rules/video-category-table.md` | 12 video scenario categories | Step 1 |
 | `rules/video-core-methodology.md` | Component-by-component build guide | Step 3 |
-| `rules/video-writing-rules.md` | 7 video-specific writing rules | Step 4 |
-| `rules/video-validation-checklist.md` | 6-point pre-submission checklist | Step 5 |
+| `rules/video-writing-rules.md` | 10 video-specific writing rules | Step 4 |
+| `rules/video-validation-checklist.md` | 基础+进阶+多方案校验清单 | Step 5 |
 | `video-words/motion.md` | Motion vocabulary library | When writing motion descriptions |
 | `video-words/scene-style.md` | Scene/style/vocabulary library | When describing scenes |
 | `camera-basic.md` | 7 basic camera movements | Basic camera work needed |
 | `camera-advanced.md` | 7 compound moves + emotion mapping | Advanced camera work |
 | `references/jimeng-video-3.0-guide.md` | 视频3.0/3.0 Pro: 8 dimensions | User mentions 视频3.0 |
-| `references/smart-multi-frame-guide.md` | 智能多帧: 多图一镜到底 | User mentions 智能多帧/多帧 |
+| `references/smart-multi-frame-guide.md` | 智能多帧: 多图一镜到底 | User mentions 智能多帧/多帧 |
+| `references/evaluation-framework.md` | 6维评估框架: 词库命中率/规则遵从度/公式完整性/模型匹配/区分度/校验 | User asks to assess/evaluate skill output |

skills/jimeng-skills/jimeng-prompt-text2video/references/evaluation-framework.md

@@ -0,0 +1,168 @@
+# 即梦文生视频提示词评估框架
+
+> 用于系统性地评估技能产出的质量——检查 prompt 对技能本身规则/词库/方法的遵从程度。
+> 当用户说"评估一下这个输出""检测命中率""看看有哪些优化方向"时加载。
+
+## 评估维度总览
+
+| # | 维度 | 检查内容 | 权重 |
+|---|------|----------|------|
+| 1 | 词库命中率 | prompt 中使用的动作/场景/风格/运镜词汇与技能词库的匹配度 | 20% |
+| 2 | 规则遵从度 | 10 条 writing rules 的逐条通过率 | 30% |
+| 3 | 公式完整性 | prompt 公式 `主体+动作+场景+镜头+风格+情绪+照明` 各组件的覆盖度 | 20% |
+| 4 | 模型-风格匹配 | prompt 写作风格与推荐模型是否匹配 (Rule 8) | 15% |
+| 5 | 多方案区分度 | 如有多个方案，视觉差异是否显著 (Rule 9) | 10% |
+| 6 | 校验完成度 | 是否通过 validation checklist | 5% |
+
+---
+
+## Step 1: 词库命中率分析
+
+### 方法
+将用户提供的 prompt 拆解为词汇片段，与技能词库逐条对照：
+
+| 技能词库 | 加载文件 | 比对方法 |
+|----------|----------|----------|
+| 动作词库 | `references/video-words/motion.md` | 逐动作词检查命中 |
+| 场景/风格词库 | `references/video-words/scene-style.md` | 逐场景/风格词检查命中 |
+| 运镜词库 | `references/camera-basic.md` | 逐运镜术语检查命中 |
+
+### 输出格式
+```
+**动作词库命中率: ~XX%**
+| 技能词库条目 | prompt 中命中 | 命中? |
+|---|---|---|
+| `轻抚` | "用指尖轻柔地抚摸" | ✅ |
+| `低头` | 未使用 | ❌ |
+
+**场景/风格词库命中率: ~XX%**
+**运镜词库命中率: ~XX%**
+```
+
+**判定参考**:
+- ≥80%: 优秀（充分利用了技能词库）
+- 60-79%: 良好（有提升空间）
+- <60%: 需要加强（提示词可能太笼统，未调用词库资源）
+
+---
+
+## Step 2: 规则遵从度逐条检查
+
+加载 `rules/video-writing-rules.md`（10 条规则），逐一检查：
+
+| 规则 | 主题 | 检查方式 |
+|------|------|----------|
+| Rule 1 | Motion 明确具体 | prompt 中每个动作是否使用了具体动词 |
+| Rule 2 | 运镜讲故事 | 运镜选择是否有情感意图 |
+| Rule 3 | 单主动作 | 是否只有一个主要动作线索 |
+| Rule 4 | 时长匹配 | 动作复杂度是否与推荐时长一致 |
+| Rule 5 | 光影时序变化 | 光描述是否包含 temporal verb |
+| Rule 6 | 运镜在后段 | 运镜描述是否在 prompt 后半段 |
+| Rule 7 | 8维控制(3.0 Pro) | 覆盖了几个维度 |
+| Rule 8 | 模型-风格匹配 | prompt 风格是否匹配推荐模型 |
+| Rule 9 | 方案区分度 | 多方案是否有 2+ 维度差异 |
+| Rule 10 | 动作层数上限 | 层数是否在模型容限内 |
+
+### 输出格式
+```
+| 规则 | 准守度 | 说明 |
+|------|--------|------|
+| Rule 1: Motion 明确 | ✅ 优 | 抚摸/微笑/发丝——每个都是具体动词 |
+| Rule 5: 光影时序 | ⚠️ 可优化 | "光线洒入"偏静态，建议加"逐渐变亮" |
+| Rule 8: 模型匹配 | ❌ 不匹配 | 长散文→推荐 视频3.0 Pro，但实际推了 Seedance 2.0 |
+```
+
+---
+
+## Step 3: 公式完整性检查
+
+使用目标模型的公式模板逐组件检查：
+
+| 公式组件 | 视频3.0 Pro 版 | Seedance 2.0 版 |
+|----------|----------------|-----------------|
+| 公式 | `主体+动作+场景+镜头+风格+情绪+照明` | `主体+动作/运动+空间背景/光影/风格+镜头调度/音效` |
+| 主体 | ✅ 一位气质温婉的年轻母亲 | ✅ 同 |
+| 动作 | ✅ 抚摸、微笑 | ✅ 同 |
+| 场景 | ✅ 卧室窗边 | ✅ 同 |
+| 镜头 | ✅ 中景推近至半身特写 | ✅ 同 |
+| 风格 | ⚠️ 建议加入更具体的风格词 | ✅ 同 |
+| 情绪 | ✅ 隐含在动作中 | — (2.0公式可选) |
+| 照明 | ⚠️ 缺少时序变化 | ⚠️ 光影偏静态 |
+
+---
+
+## Step 4: 方案区分度检查（多方案时）
+
+加载 `rules/video-writing-rules.md` 的 Rule 9 区分度检查表：
+
+| 检查项 | 方案A vs 方案B |
+|--------|---------------|
+| 运镜方案不同？ | ❌ 两者都是推镜 |
+| 动作密度不同？ | ⚠️ 接近 |
+| 景别不同？ | ❌ 都是中景-特写 |
+| 时长/节奏不同？ | ❌ 都是10s慢节奏 |
+| 场景视角不同？ | ❌ 相同场景 |
+| **结论** | ⚠️ 区分度低，可能浪费用户时间 |
+
+---
+
+## Step 5: 综合评分与优先级
+
+### 评分公式
+```
+总分 = 词库命中率×20% + 规则遵从度×30% + 公式完整性×20% + 模型匹配×15% + 方案区分度×10% + 校验完成度×5%
+```
+
+### 评分基准
+| 分数 | 评级 | 含义 |
+|------|------|------|
+| 90-100 | 🥇 卓越 | 可直接交付 |
+| 75-89 | 🥇 优秀 | 少量微调即可 |
+| 60-74 | ✅ 可用 | 需要一些修改 |
+| 40-59 | ⚠️ 需重写 | 2+ 维度有严重问题 |
+| <40 | ❌ 不可用 | 多个维度不合格 |
+
+### 优先级分级
+| 优先级 | 象征 | 处理建议 |
+|--------|------|----------|
+| 🔴 最高优先 | 影响成片质量的核心问题 | 必须修复，否则产出不合格 |
+| 🟡 高优先级 | 显著影响但不致命 | 建议修复，收益明显 |
+| 🟢 中优先级 | 细节优化 | 有时间再优化 |
+| 🔵 低优先级 | 流程完善 | 用于迭代技能本身 |
+
+---
+
+## 完整评估输出模板
+
+```markdown
+## 评估报告
+
+### 一、词库命中率分析
+...
+
+### 二、规则遵从度逐条检查
+...
+
+### 三、公式完整性
+...
+
+### 四、多方案区分度（如有）
+...
+
+### 五、综合评分
+| 维度 | 得分 |
+|------|------|
+| 词库命中率 | XX/100 |
+| 规则遵从度 | XX/100 |
+| 公式完整性 | XX/100 |
+| 模型-风格匹配 | XX/100 |
+| 方案区分度 | XX/100 |
+| 校验完成度 | XX/100 |
+| **总分** | **XX/100** |
+
+### 六、优先优化建议
+🔴 最高优先: ...
+🟡 高优先级: ...
+🟢 中优先级: ...
+🔵 低优先级: ...
+```

skills/jimeng-skills/jimeng-prompt-text2video/rules/video-core-methodology.md

@@ -6,17 +6,36 @@
 2. **Action/Motion** (动作/运动): **THE critical video component**. Describe what moves, how, at what speed. Use motion vocabulary from `word-library.md`. Be explicit — "缓缓转身，长发随风飘动" not "动了一下"
 3. **Scene** (场景): Where + when. Provides context for the motion
 4. **Camera Movement** (运镜): How the camera moves. Use `camera-library.md`. Pick 1-2 movements max
-5. **Lighting/Atmosphere** (光影氛围): In video, describe how light changes — flickering, shifting, dimming
+5. **Lighting/Atmosphere** (光影氛围): In video, describe **how light changes** over time — flickering, shifting, dimming. NOT just its starting state. Use temporal verbs (参考 Rule 5 和 temporal light vocabulary).
 6. **Duration/Pacing** (时长/节奏): How long + how fast. Seedance 2.0 supports 4-15 seconds
 7. **Style/Quality** (风格/画质): Video-specific quality terms like "电影级画质""高帧率流畅""4K视频"
 
+### CRITICAL: Choose model BEFORE writing
+
+The writing style must match the target model (Rule 8). Decide your model first, then write the prompt in that model's native language.
+
+| If target model is | Write in this style |
+|---|---|
+| **视频3.0 / 3.0 Pro** | Natural language prose — descriptive paragraphs, sensory detail, soft scene setting. The model extracts structure from free text. |
+| **Seedance 2.0** | Structured segments — clear subject→action→scene→camera breakdown. Use shorter sentences, explicit camera directions, separate motion from atmosphere. |
+
+**Before building, first decide which model, then load the corresponding guide.**
+
 **At each component step**, also apply these rules:
 - Motion first, scene second — place action description before or immediately after the subject
 - One primary movement — complex multi-subject interactions + camera work tend to fail
 - Chinese camera terms > English — "镜头缓缓推进" > "dolly in slowly"
 - Match action to duration — 5s = one clear action; 10s = action with development; 15s = multi-stage progression
 - Describe how light changes — unlike images, video light can evolve: "夕阳逐渐西沉""烛光摇曳闪烁"
+- Action layer count — Seedance 2.0 ≤ 3 layers; 视频3.0 Pro ≤ 5 layers (参考 Rule 10)
 
 ## Present the result
 
 Always present: complete prompt in code block + model variant + suggested duration + camera explanation + optional variations.
+
+### Multi-scheme outputs (≥ 2 alternatives)
+
+When presenting multiple schemes, verify differentiation (Rule 9) and include a **one-line "看片体验" tagline** for each:
+- ✅ "情感聚焦型 — 推镜逼近母子互动，突出情绪细节"
+- ✅ "氛围沉浸型 — 固定镜头搭配环境呼吸感，宁静治愈"
+- ✅ "情绪释放型 — 亲吻+抬头+微笑，5秒短而有力"

skills/jimeng-skills/jimeng-prompt-text2video/rules/video-validation-checklist.md

@@ -1,8 +1,26 @@
 # Video Validation Checklist
 
+## 基础校验（必须全部通过）
 - [ ] Motion is explicit and specific (concrete action verb, not "有动作")
 - [ ] Camera movement clearly described (1-2 movements, Chinese terms)
 - [ ] Action complexity matches suggested duration
 - [ ] Only one primary subject performing one primary action
 - [ ] Light/atmosphere includes temporal change where appropriate
 - [ ] Model variant + duration recommendation included
+
+## 进阶校验（方案输出前必查）
+- [ ] Prompt style matches recommended model's strength (Rule 8)
+      — 长描述散文 → 视频3.0 Pro；结构化短句 → Seedance 2.0
+- [ ] Action layer count within model tolerance (Rule 10)
+      — Seedance 2.0 ≤ 3 layers; 视频3.0 Pro ≤ 5 layers
+- [ ] Light description contains a temporal verb (Rule 5)
+      — "洒入"❌ → "缓缓流动""逐渐变亮"✅
+- [ ] No hex color codes — Chinese color names only
+- [ ] Duration fits action complexity (Rule 4)
+
+## 多方案区分度校验（输出 ≥ 2 方案时）
+- [ ] At least 2 dimensions differ between schemes (Rule 9):
+      — 运镜 / 动作密度 / 景别 / 节奏 / 场景视角
+- [ ] Each scheme has a distinct one-sentence "看片体验" tagline
+      — e.g. "情感聚焦型" vs "氛围沉浸型" vs "情绪释放型"
+- [ ] No two schemes would look similar in actual output

skills/jimeng-skills/jimeng-prompt-text2video/rules/video-writing-rules.md

@@ -23,9 +23,21 @@ Seedance 2.0 works best with a single clear action. For 15s clips, 2-3 stage pro
 | 7-10s | Action with development ("从远处走来→经过镜头→渐行渐远") |
 | 11-15s | 2-3 stage progression ("推门→环顾→走向窗边") |
 
-### Rule 5: Light can (and should) change
+### Rule 5: Light can (and should) change — with temporal language
 
-"夕阳缓缓沉入海平面""烛光在微风中摇曳""霓虹灯渐次亮起"
+Describe how light **transforms over time** — not just its starting state.
+
+✅ "晨光透过薄纱窗帘逐渐变亮，光影在母亲面庞上缓缓移动"
+✅ "夕阳缓缓沉入海平面，光线从金黄渐变为橘红"
+✅ "霓虹灯渐次亮起，招牌的蓝光在水洼中跳动"
+✅ "烛光在微风中摇曳，阴影在墙壁上轻轻舞动"
+✗ "清晨的柔和光线透过窗帘洒入"（静态描述 → 缺少变化感）
+✗ "暖色台灯散发着柔和的光晕"（静态 → 光没有在变）
+
+**Temporal light vocabulary:**
+`光线逐渐变亮` · `光线渐暗` · `光影流动` · `光斑在地面缓缓移动` · `阴影拉长` · `晨光渐浓` · `暮色渐沉` · `烛光摇曳` · `灯光渐次亮起` · `阳光角度缓慢偏移` · `穿透云层的光束缓慢移动` · `水面反光跳动闪烁`
+
+**Rule of thumb**: If you can replace the light description with a photo caption, it's static. A video light description should contain a verb that implies time. "洒入" is still a snapshot — "缓缓流动""逐渐变亮" carries the time dimension.
 
 ### Rule 6: Camera in the latter half
 
@@ -43,3 +55,65 @@ When using 视频3.0/3.0 Pro, leverage these 8 dimensions from `references/jimen
 6. **Prompt控制美感** — 人物外形+画面美感
 7. **切镜能力** — 镜头切换和转场
 8. **创意特效** — 首尾帧类型/玩法/I+T2V特效
+
+---
+
+### Rule 8: Match prompt style to model version — 致命误区
+
+Each model excels at different writing styles. **Mismatching will degrade output quality severely.**
+
+| Prompt 风格特征 | 推荐模型 | 原因 |
+|---|---|---|
+| 大段自然语言、细腻场景铺陈、氛围描写、多感官描述 | **视频3.0 / 3.0 Pro** | 官方公式 = 自然语言自由书写。长文本理解力强，善于从散文中提炼场景要素 |
+| 结构化短句、分段清晰、动词集中、运镜指令明确 | **Seedance 2.0 系列** | 官方公式 = 分解式 `主体 + 动作/运动 + 空间背景/光影/风格 + 镜头调度/音效`。对结构化指令的解析更准 |
+| 极简动作 + 氛围主导 | **视频3.0 Pro** | 氛围理解力强，能补充细节 |
+| 需要抽帧/慢动作/首尾帧特效 | **视频3.0 Pro** | 支持创意特效维度 |
+| 需要原生音频/多模态参考 | **Seedance 2.0** | 原生音视频联合生成 |
+| 图生视频 | **Seedance 2.0** (I2V/V2V) 或 **视频3.0 Pro** (I+T2V) | 根据是否需要图像外的控制力选择 |
+
+**Wrong assignment example（来自实际案例）**:
+- ❌ 长描述自然语言 prompt → 推荐 Seedance 2.0 → Seedance 对散文式描述的语义解析不如 3.0 Pro，细节易丢失
+- ✅ 长描述自然语言 prompt → 推荐 视频3.0 Pro → 发挥其自然语言理解优势
+- ✅ 如果坚持用 Seedance 2.0 → 将长描述重写为结构化短句版本
+
+**Rule of thumb**:
+> 如果你的 prompt 是一段散文 → 选 **视频3.0 Pro**
+> 如果你的 prompt 是分段清单 → 选 **Seedance 2.0**
+> 如果你的 prompt 两者都可以 → **推荐优先视频3.0 Pro**（容错率更高）
+
+### Rule 9: Multi-scheme differentiation — 方案区分度
+
+When providing multiple alternatives for the same scene, ensure each scheme delivers a **visually distinct** viewing experience. Similar schemes waste the user's choice.
+
+✅ **Good differentiation:**
+- 方案A: 慢镜头推近 + 手部微动作 → 情感聚焦型
+- 方案B: 固定镜头 + 环境呼吸感（窗帘飘动/光影流动）→ 氛围沉浸型
+- 方案C: 亲吻 + 抬头 + 全景 → 情绪释放型
+
+❌ **Bad differentiation (real example):**
+- 方案A: 慢镜头 + 推近 + 抚摸 + 花瓣颤动 → 温情
+- 方案B: 慢镜头 + 极缓推近 + 窗帘飘动 + 呼吸 → 氛围
+→ 两者核心视觉差异不显著，实际成片难区分
+
+**Differentiation checklist for each new scheme:**
+1. ❓ 运镜方案不同？（推 vs 固定 vs 环绕 vs 摇）
+2. ❓ 动作/情绪密度不同？（高能量 vs 低能量）
+3. ❓ 景别/构图不同？（特写 vs 中景 vs 全景）
+4. ❓ 节奏/时长不同？（5s 紧凑 vs 10s 舒展）
+5. ❓ 场景/视角不同？（室内 vs 室外？主观 vs 客观？）
+6. ❓ 如果以上 5 条都相似 → **合并 or 重新构思** — 用户不需要两个同质的选项
+
+**At minimum, differentiate on 2+ dimensions.**
+
+### Rule 10: Action layer count — by model tolerance
+
+Different models have different tolerance for multi-layer motion.
+
+| Layer Count | Definition | 视频3.0 Pro | Seedance 2.0 |
+|---|---|---|---|
+| **1 layer** | 一个主要动作（如"一个人行走"） | ✅ 安全 | ✅ 安全 |
+| **2-3 layers** | 主体动作 + 1-2 环境/自然动态（如"行走+树叶飘落"） | ✅ 安全 | ✅ 安全 |
+| **4-5 layers** | 主体动作 + 微表情 + 环境动态 + 次对象动作（如"抚摸+微笑+发丝+花瓣+手指颤动"） | ✅ 可行 | ⚠️ risk — 可能出现部分动态丢失或 artifacts |
+| **6+ layers** | 多主体交互 + 多层环境动态 | ⚠️ risk | ❌ 不建议 — distortion 风险高 |
+
+**Rule**: For Seedance 2.0, cap action layers at 3 (one primary + up to 2 secondary/environmental). For 视频3.0 Pro, 4-5 layers is acceptable — use the 5th layer for micro-expressions or subtle environmental detail.