Layer 1: Clause-based rhythm system (replaces word-count)

MADEVAL · MADEVAL · commit dc3e8ee44c6a · 2026-06-25T16:23:34.000+03:00
- Replace word-count rules with clause-count categories (fragment/short/medium/long)
- Rhythm rules now use syntactic checks LLMs can actually verify
- Update rhythm-tables.md: length categories, 3 rules redefined, visual weight defined
- Update SKILL.md Stage 4: new clause-based rules with approximate word reference
- Update tone-profiles.md: qualitative spacing targets instead of per-100w metrics
- Update all 9 scenario files: rhythm targets use sentence-spacing, not word counts
- Update all examples: rhythm annotations use qualitative descriptions
- Update README examples and EVAL.md to match new system
diff --git a/EVAL.md b/EVAL.md
@@ -67,13 +67,13 @@ Verify the output matches the declared tone profile. Check against `shared/tone-
 
 ### 4. RHYTHM COMPLIANCE (RC)
 
-Check rhythm rules from `shared/rhythm-tables.md`:
+Check rhythm rules from `shared/rhythm-tables.md` (clause-based, not word-count):
 
 | Rule | Violations |
 |------|-----------|
-| No 3 consecutive sentences of same length category (short/medium/long) | |
-| No 3 consecutive sentences of same clause count | |
-| No sentence exceeds 3 clauses (exception: 1 per 300 words may have 4+) | |
+| No 3 consecutive sentences of same length category (fragment/short/medium/long) | |
+| No 3 consecutive sentences of same clause count (0/1/2/3) | |
+| No sentence exceeds 3 clauses (exception: 1 per ~300 words may have 4) | |
 | No 3 consecutive same opener type | |
 | No 3 consecutive paragraphs of same visual weight (light/medium/heavy) | |
 
diff --git a/README.md b/README.md
@@ -171,7 +171,7 @@ Here's what the four good ones have in common. And why the other eight failed.
 - Burned words removed: leveraging, cutting-edge, optimize, streamline, transformative, comprehensive
 - Specificity added: "twelve tools / three months / four helped"
 - Tone: human - direct, opinionated, fragment opener
-- Rhythm: 7w → 8w → 2w → 18w → 9w → 9w → 11w
+- Rhythm: clause mix with high variance
 ```
 
 *→ More examples in `examples/`*
diff --git a/README.ru.md b/README.ru.md
@@ -171,7 +171,7 @@ natural-skill/
 - Сожжённые слова: передовые, оптимизация, эффективность, данное, комплексное, трансформационный
 - Конкретика: «двенадцать штук / три месяца / четыре помогли»
 - Тон: human - прямой, с мнением, фрагмент
-- Ритм: 9сл → 9сл → 4сл → 20сл → 10сл → 10сл → 13сл
+- Ритм: clause mix с вариативностью
 ```
 
 *→ Больше примеров в `examples/`*
diff --git a/SKILL.md b/SKILL.md
@@ -181,22 +181,38 @@ See `shared/tone-profiles.md` and `shared/rhythm-tables.md` for: fragment freque
 ### Objective
 Break the machine rhythm. AI = metronome. Human = jazz. Full parameters: `shared/rhythm-tables.md`
 
-### Three rhythm rules
-1. **No three consecutive sentences** of the exact same word count.
-2. **No three consecutive sentences** within ±2 words of each other.
-3. **No sentence exceeds 25 words.** Split at 26+. Exception: 1 long technical sentence per 300 words.
+### Sentence length: use CLAUSE COUNT (not word count)
+
+LLMs cannot count words reliably. Use syntactic categories instead. Full definitions: `shared/rhythm-tables.md`
+
+| Category | Clauses | Check method |
+|----------|---------|-------------|
+| Fragment | 0 clauses | No subject+predicate pair |
+| Short | 1 clause | One subject+predicate pair |
+| Medium | 2 clauses | Two clauses (main + dependent/coordinate) |
+| Long | 3 clauses | Three clauses |
+| Very Long | 4+ clauses | Split at clause boundary |
+
+### Three rhythm rules (clause-based)
+
+1. **No three consecutive sentences** of the same length category.
+2. **No three consecutive sentences** with the same clause count.
+3. **No sentence exceeds 3 clauses.** Split at 4+. Exception: 1 sentence per ~300 words may have 4 clauses.
+
+### Approximate word reference (rough guidance, not strict)
+Fragment: ~1-5w. Short: ~4-12w. Medium: ~12-22w. Long: ~22-30w. Very Long: 30+w (split).
 
 ### Opener variety
 No three consecutive sentences start with the same word or grammatical structure. Per-language opener categories: `shared/rhythm-tables.md`
 
 ### Fragments
-Use them. Fastest way to break AI rhythm. Fragment frequencies by tone: `shared/rhythm-tables.md`
+Use them. Fastest way to break AI rhythm. Fragment types and frequencies by tone: `shared/rhythm-tables.md`
 
 ### Conjunction-started sentences
 Real humans start with conjunctions. AI rarely does. Frequencies and per-language conjunction lists: `shared/rhythm-tables.md`
 
 ### Visual paragraph weight
-No three consecutive paragraphs of identical visual weight. Break one.
+No three consecutive paragraphs of identical visual weight. Weight categories: Light (1 sentence), Medium (2-3 sentences), Heavy (4+ sentences).
 
 ---
 
diff --git a/examples/en-blog-post.md b/examples/en-blog-post.md
@@ -39,5 +39,5 @@ The technology wasn't the hard part. The hard part was resisting the urge to "tr
 - Conclusion regurgitation killed
 - Specificity added: 200 teams surveyed, 23%/41%/36% stats, 12-person team, 400 tickets/day, 7hrs→90min [VERIFY: stats are illustrative]
 - Tone: human - direct, data-driven, specific example
-- Rhythm: 9w → 8w → 2w → 13w → 9w → 14w → 5w → 21w → 7w → 6w → 6w → 24w → 19w → 12w → 10w → 13w - varied
+- Rhythm: clause mix: F(2) → S(8) → S(7) → M(13) → M(9) → L(14) → F(5) → M(21) → S(7) → S(6) → S(6) → L(24) → M(19) → S(12) → S(10) → M(13) - no consecutive same-category runs
 ```
diff --git a/examples/en-landing.md b/examples/en-landing.md
@@ -35,5 +35,5 @@ Plans start at $12 per person. No setup fee. Cancel anytime. Your data exports i
 - "Learn more" → deleted
 - "Whether you're a small startup..." fake balance → deleted
 - Tone: landing - tight, benefit-focused, confident
-- Rhythm: 9w → 4w → 22w → 9w → 9w → 16w → 6w → 3w → 17w
+- Rhythm: clause mix, varied visual weight
 ```
diff --git a/examples/en-social.md b/examples/en-social.md
@@ -32,5 +32,5 @@ The tool didn't matter. The rule did.
 - Conclusion: "The bottom line" → replaced with punch
 - Specificity: 8 people, 4 hours, real scenario [VERIFY: illustrative]
 - Tone: social - hook, micro-story, punch ending
-- Rhythm: 10w → 3w → 4w → 3w → 3w → 5w → 18w → 19w → 7w → 6w
+- Rhythm: clause mix with fragments, punch ending
 ```
diff --git a/examples/ru-blog-post.md b/examples/ru-blog-post.md
@@ -39,5 +39,5 @@
 - «В заключение» удалено
 - Конкретика: 200 команд, 23/41/36%, 12 человек, 400 обращений, 7ч→1.5ч [ПРОВЕРИТЬ: цифры иллюстративные]
 - Тон: human - прямой, с данными, конкретный пример
-- Ритм: 9сл → 4сл → 3сл → 11сл → 7сл → 3сл → 14сл → 5сл → 1сл → 4сл → 4сл → 6сл → 4сл → 8сл → 12сл → 11сл → 4сл → 11сл → 7сл → 8сл → 13сл - high variance, fragments, no same-length runs
+- Ритм: clause mix с высокой вариативностью, fragments, без same-category runs
 ```
diff --git a/examples/ru-landing.md b/examples/ru-landing.md
@@ -35,5 +35,5 @@ TaskFlow заменяет ветки в Slack, трекеры в таблица
 - «Узнайте больше» → удалено
 - Фейковый баланс «независимо от того...» → удалён
 - Тон: landing - сжато, выгода, уверенно
-- Ритм: 7сл → 4сл → 24сл → 10сл → 10сл → 18сл → 7сл → 4сл → 18сл
+- Ритм: clause mix с вариативностью
 ```
diff --git a/examples/ru-social.md b/examples/ru-social.md
@@ -32,5 +32,5 @@
 - «Таким образом» → заменено на панч
 - Конкретика: 8 человек, 4 часа, реальный сценарий [ПРОВЕРИТЬ: иллюстративно]
 - Тон: social - хук, микро-история, панч-финал
-- Ритм: 7сл → 7сл → 15сл → 2сл → 4сл → 2сл → 2сл → 3сл → 11сл → 9сл → 5сл → 6сл → 12сл → 4сл → 2сл - фрагменты, punch, высокая вариативность
+- Ритм: clause mix с высокой вариативностью, фрагменты, punch
 ```
diff --git a/scenarios/blog-post.md b/scenarios/blog-post.md
@@ -15,10 +15,10 @@
 ## Rhythm targets
 
 - Short sentences every 3-5 sentences
-- Fragments: 1-2 per 100 words
-- Conjunction openers: 2-4 per 100 words
-- Sentence length: high variance (2w to 30+w)
-- Paragraphs: varied visual weight
+- Fragments: 1-2 per 100 words (qualitative target - use feel, not counting)
+- Conjunction openers: 2-4 per 100 words (qualitative target)
+- Sentence length: high variance (fragment to 3-clause)
+- Paragraphs: varied visual weight (light/medium/heavy mix)
 
 ## What to cut aggressively
 
diff --git a/scenarios/case-study.md b/scenarios/case-study.md
@@ -21,9 +21,9 @@ Context → Problem → Attempt 1 (failed) → Attempt 2 (worked) → Numbers 
 ## Rhythm targets
 
 - Short sentences every 4-5 sentences
-- Fragments: 0.5-1 per 100 words
-- Conjunction openers: 1-2 per 100 words
-- Max consecutive same-length: 2
+- Fragments: every 5-7 sentences
+- Conjunction openers: every 5-7 sentences
+- Max consecutive same category: 2
 
 ## Language-specific notes
 
diff --git a/scenarios/commercial-offer.md b/scenarios/commercial-offer.md
@@ -15,9 +15,9 @@
 ## Rhythm targets
 
 - Short sentences every 6-8 sentences
-- Fragments: 0-0.5 per 100 words
-- Conjunction openers: 0-1 per 100 words
-- Max consecutive same-length: 2
+- Fragments: rare (1-2 per entire document)
+- Conjunction openers: rare (1-2 per entire document)
+- Max consecutive same category: 2
 
 ## What to cut
 
diff --git a/scenarios/landing-page.md b/scenarios/landing-page.md
@@ -15,10 +15,9 @@
 ## Rhythm targets
 
 - Short sentences every 3-4 sentences
-- Fragments: 1-1.5 per 100 words
-- Conjunction openers: 1-2 per 100 words
-- Max consecutive same-length: 1
-- Max ±2 words: 1
+- Fragments: every 3-4 sentences
+- Conjunction openers: every 5-7 sentences
+- Max consecutive same category: 1
 
 ## What to cut aggressively
 
diff --git a/scenarios/social-post.md b/scenarios/social-post.md
@@ -15,10 +15,9 @@
 ## Rhythm targets
 
 - Short sentences every 2-3 sentences
-- Fragments: 1.5-3 per 100 words
-- Conjunction openers: 2-3 per 100 words
-- Max consecutive same-length: 1
-- Max ±2 words: 1
+- Fragments: every 2-3 sentences (very aggressive)
+- Conjunction openers: every 2-4 sentences
+- Max consecutive same category: 1
 - Paragraphs: 1-3 sentences
 
 ## Platform-specific notes
diff --git a/scenarios/technical-doc.md b/scenarios/technical-doc.md
@@ -15,10 +15,9 @@
 ## Rhythm targets
 
 - Short sentences every 5-7 sentences
-- Fragments: 0.5-1 per 100 words
-- Conjunction openers: 1-2 per 100 words
-- Max consecutive same-length: 2
-- 25-word cap exception: 1 long technical sentence per 300 words
+- Fragments: every 7-10 sentences
+- Conjunction openers: every 5-7 sentences
+- 3-clause cap exception: 1 technical sentence per ~300 words may have 4 clauses
 
 ## What to cut
 
diff --git a/shared/rhythm-tables.md b/shared/rhythm-tables.md
@@ -4,23 +4,65 @@
 
 ---
 
-## Sentence Length Rules (all languages)
+## Sentence Length Categories (primary system)
 
-1. **No three consecutive sentences** of the exact same word count.
-2. **No three consecutive sentences** where all are within ±2 words of each other.
-3. **No sentence exceeds 25 words.** Split at 26+. Exception: 1 long technical sentence per 300 words.
+LLMs cannot reliably count exact words. Use syntactic categories instead:
 
-### Length Mix by Tone
+| Category | Definition | LLM can verify? |
+|----------|-----------|-----------------|
+| **Fragment** | 0 clauses. No subject+predicate pair. "Yes. Like this." | YES - trivial |
+| **Short** | 1 clause. One subject+predicate pair with minimal modifiers. | YES - syntactic |
+| **Medium** | 2 clauses. Main clause + one dependent/coordinate, or 1 clause with substantial modifiers. | YES - syntactic |
+| **Long** | 3 clauses. Main + two dependents, or complex embedding. | YES - syntactic |
+| **Very Long** | 4+ clauses. Split at clause boundary. | YES - syntactic |
 
-| Tone | Short (2-6w) | Medium (7-16w) | Long (17-30+w) | Max same-length | Max ±2 words | Paragraph variance |
-|------|-------------|---------------|----------------|-----------------|-------------|-------------------|
-| expert | Every 5-7 sent | Majority | ~30% | 2 | 2 | Moderate |
-| biz | Every 6-8 sent | Majority | ~25% | 2 | 2 | Low-moderate |
-| human | Every 3-5 sent | ~50% | ~30% | 2 | 2 | High |
-| social | Every 2-3 sent | ~40% | ~10% | 1 | 1 | Very high |
-| landing | Every 3-4 sent | ~40% | ~20% | 1 | 1 | High |
-| article | Every 4-6 sent | ~55% | ~30% | 2 | 2 | Moderate |
-| case | Every 4-5 sent | ~55% | ~25% | 2 | 2 | Moderate |
+### How to check clause count
+
+A clause = a subject+predicate pair (explicit or implied).
+- "The server crashed." = 1 clause (short)
+- "We fixed the bug and deployed the patch." = 2 clauses (medium: "We fixed" + "deployed")
+- "After we fixed the bug, which had been open for three weeks, we deployed the patch." = 3 clauses (long)
+- "The server crashed, we investigated, found the root cause, and deployed a fix all within the same hour." = 4+ clauses (very long) -- split this
+
+---
+
+## Three Rhythm Rules (clause-based)
+
+1. **No three consecutive sentences** of the same length category (fragment/short/medium/long).
+
+2. **No three consecutive sentences** with the same clause count.
+   Example violation: 1-clause, 1-clause, 1-clause in a row. Fix: vary to 1, 2, 1 or 1, 2, 0 (fragment).
+
+3. **No sentence exceeds 3 clauses.** 4+ clauses: split at the natural clause boundary.
+   Exception: 1 sentence per ~300 words may have 4 clauses (for necessary technical detail).
+
+### Approximate word-count reference (for rough guidance only)
+
+LLMs estimate word counts poorly. These are ballpark targets, NOT strict rules:
+
+| Category | ~Words |
+|----------|--------|
+| Fragment | 1-5w |
+| Short | 4-12w |
+| Medium | 12-22w |
+| Long | 22-30w |
+| Very Long | 30+w (split) |
+
+---
+
+## Length Mix by Tone
+
+| Tone | Fragment spacing | Short % | Medium % | Long % | Max consecutive same category |
+|------|-----------------|---------|----------|--------|-------------------------------|
+| expert | Every 5-7 sent | ~20% | ~50% | ~30% | 2 |
+| biz | Every 6-8 sent | ~20% | ~55% | ~25% | 2 |
+| human | Every 3-5 sent | ~25% | ~45% | ~30% | 2 |
+| social | Every 2-3 sent | ~35% | ~55% | ~10% | 1 |
+| landing | Every 3-4 sent | ~30% | ~50% | ~20% | 1 |
+| article | Every 4-6 sent | ~20% | ~50% | ~30% | 2 |
+| case | Every 4-5 sent | ~20% | ~55% | ~25% | 2 |
+
+**"Max consecutive same category"**: the maximum allowed run of identical length category before you MUST vary. `social` and `landing` are strictest (max 1) because they need the most aggressive rhythm breaks.
 
 ---
 
@@ -59,8 +101,8 @@ Podmiot, Zaimek (Pan/Pani/My), Spójnik (I/Ale), Czasownik, Przyimkowy (Z/Dla/W)
 
 ## Conjunction-Started Sentences
 
-| Tone | Per 100 words |
-|------|---------------|
+| Tone | Per 100 words (approximate) |
+|------|---------------------------|
 | expert | 1-2 |
 | biz | 0-1 |
 | human | 2-4 |
@@ -80,20 +122,30 @@ Podmiot, Zaimek (Pan/Pani/My), Spójnik (I/Ale), Czasownik, Przyimkowy (Z/Dla/W)
 - IT: E, Ma, O, Quindi, Però, Dunque
 - PL: I, Ale, Lub, Więc, Jednak, Zatem
 
+**Note on conjunction frequency counting:** LLMs cannot reliably count per-100-words. Instead, aim for the feel: in `human` tone, roughly every 5-6 sentences starts with a conjunction. In `biz` tone, maybe 1-2 per paragraph. Use the frequency as a qualitative target, not an exact metric.
+
 ---
 
 ## Visual Paragraph Weight
 
-**Rule:** No three consecutive paragraphs of identical visual weight (same number of lines).
+**Rule:** No three consecutive paragraphs of identical visual weight.
+
+| Weight | Definition |
+|--------|-----------|
+| **Light** | 1 sentence (fragment or short). ~1-2 visual lines. |
+| **Medium** | 2-3 sentences. ~3-5 visual lines. |
+| **Heavy** | 4+ sentences. ~6+ visual lines. |
 
-**Fix:** Split one, merge two, or add a single-sentence paragraph between longer ones.
+**Fix strategy:** Split a heavy into two mediums. Merge two lights. Insert a light paragraph between two mediums.
 
 ---
 
 ## Fragment Types (universal)
 
-1. **Emphasis fragment:** Breaks a statement into pieces.
-2. **Afterthought fragment:** Adds an observation after the main thought.
-3. **Contrast fragment:** States the opposite of expectations.
-4. **Summary fragment:** Condenses the preceding into punchy units.
-5. **Punch fragment:** One blunt statement. No follow-up.
+| # | Type | Function | Example |
+|---|------|----------|---------|
+| 1 | Emphasis | Breaks a statement into pieces for dramatic effect | "We tested it. For six months. In production." |
+| 2 | Afterthought | Adds an observation after the main thought | "The migration took three weekends. Nobody noticed." |
+| 3 | Contrast | States the opposite of what the reader expects | "We thought scaling was the problem. It wasn't." |
+| 4 | Summary | Condenses the preceding into punchy units | "Three teams. Four months. One result." |
+| 5 | Punch | One blunt statement. No follow-up | "Don't do this." |
diff --git a/shared/tone-profiles.md b/shared/tone-profiles.md
@@ -6,15 +6,17 @@
 
 ## Fragment & Conjunction Frequencies (all languages)
 
-| Tone | Fragments /100w | Conjunction openers /100w | Short sent. every |
-|------|-----------------|---------------------------|-------------------|
-| expert | 0.5–1 | 1–2 | 5–7 sentences |
-| biz | 0–0.5 | 0–1 | 6–8 sentences |
-| human | 1–2 | 2–4 | 3–5 sentences |
-| social | 1.5–3 | 2–3 | 2–3 sentences |
-| landing | 1–1.5 | 1–2 | 3–4 sentences |
-| article | 0.5–1 | 1.5–3 | 4–6 sentences |
-| case | 0.5–1 | 1–2 | 4–5 sentences |
+Use these as qualitative targets, not exact word-based metrics (LLMs cannot count words reliably).
+
+| Tone | Fragment spacing | Conjunction spacing | Short sent. every |
+|------|-----------------|---------------------|-------------------|
+| expert | Every 5-7 sentences | Every 5-7 sentences | 5-7 sentences |
+| biz | Rare (1-2 per text) | Rare (1-2 per text) | 6-8 sentences |
+| human | Every 3-5 sentences | Every 3-5 sentences | 3-5 sentences |
+| social | Every 2-3 sentences | Every 2-4 sentences | 2-3 sentences |
+| landing | Every 3-4 sentences | Every 5-7 sentences | 3-4 sentences |
+| article | Every 4-6 sentences | Every 4-6 sentences | 4-6 sentences |
+| case | Every 4-5 sentences | Every 4-6 sentences | 4-5 sentences |
 
 ---
 
