Strengthen chapter review quality gates

- chapter_review.yml: mandatory PROSE EVIDENCE section (3-5 verbatim quotes)
- CHARACTER VOICE AUDIT with per-constraint checks and violation quoting
- Score anchors (95-100/85-94/70-84/<70) to break rubber-stamp clustering
- MUST-FIX items require ORIGINAL quote + PROBLEM + FIX rewrite
- Adjudication: prose_evidence(35) specificity(30) voice_audit(20) coverage(15)
- chapter_roundtable.yml: require quoted evidence, add character_profiles check

templates/chapter_review.yml

@@ -1,7 +1,8 @@
 name: chapter_review
 description: >
   Independent chapter review. Agent reads the chapter and produces a structured
-  critique in their editorial domain with explicit must-fix vs optional categorization.
+  critique with mandatory verbatim prose quotes, score-anchored verdict, and
+  explicit character voice violation checks.
 debug: true
 rag_exclude: [business_plan, charter]
 model: power
@@ -18,7 +19,7 @@ sections:
 
 steps:
   - type: think
-    max_tokens: 4000
+    max_tokens: 5000
     hint: |
       TARGET AUDIENCE: {genre_audience}
       GENRE: {genre_name}
@@ -29,54 +30,82 @@ steps:
       {chapter_text}
       ---
 
-      Write a detailed editorial review from your perspective.
-      Be specific — cite line numbers or quote passages where relevant.
-      Structure your review using EXACTLY these six labeled sections:
+      Write a detailed editorial review. You MUST quote the actual text to support every
+      claim you make. Vague commentary ("the pacing drags") without a quoted passage is
+      not acceptable and will cause this review to fail adjudication.
 
-        1. STRENGTHS TO PRESERVE
-           List elements that are working well and must survive the editorial process unchanged.
-           Be specific: quote or paraphrase the passage. Vague praise ("the writing is good")
-           is not useful here.
-           ALSO CHECK VOICE SIGNATURES:
-           CHARACTER VOICE PROFILES:
+      Structure your review using EXACTLY these seven labeled sections:
+
+        1. PROSE EVIDENCE
+           Quote 3-5 passages verbatim from this chapter (use quotation marks and indicate
+           approximate position: "early", "mid", "late").
+           For each quote, give one sentence of inline commentary: what it does well or
+           what it fails to do. This section is MANDATORY -- do not summarize or paraphrase.
+           A review with zero verbatim quotes will automatically score below 70.
+
+        2. CHARACTER VOICE AUDIT
            {character_profiles}
-           If the above block is non-empty: verify each named character in this chapter
-           speaks to their profile. Can you identify each character's dialogue without
-           speaker tags? Note YES or NO for each main character.
-           If the block is empty: check RAG/project context for "## Voice Signatures"
-           and note whether voice differentiation is present.
+           If the above block is non-empty: for EACH named character who speaks in this
+           chapter, check their voice against their profile. Quote one line of their
+           dialogue and state YES or NO for each constraint:
+             - Does the character use their signature vocabulary / verbal tics?
+             - Do they avoid any explicitly forbidden speech patterns?
+             - Is their emotional register consistent with their arc position?
+           If a violation is found, quote the offending line and state exactly what rule
+           is broken (e.g. "Dorian uses contraction 'don't' -- profile forbids contractions
+           unless in extreme pain or physical exhaustion. This scene does not qualify.").
+           If the block is empty: check RAG context for "## Voice Signatures" and report
+           what you find.
 
-        2. MUST-FIX — CONTINUITY
+        3. STRENGTHS TO PRESERVE
+           List 2-4 elements working well that must survive the editorial process unchanged.
+           Each item must include a verbatim quote or specific scene reference.
+           Generic praise ("the atmosphere is strong") without evidence is not allowed here.
+
+        4. MUST-FIX -- CONTINUITY
            List factual errors, POV breaks, timeline inconsistencies, or world-rule violations.
-           These are non-negotiable — they cannot be left as "optional." For each item, state:
-           - What the error is
-           - What the correction should be
+           These are non-negotiable. For each item state:
+           - ORIGINAL: quote the offending passage verbatim
+           - PROBLEM: what rule or established fact it violates
+           - FIX: the concrete correction (rewrite the sentence if necessary)
 
-        3. MUST-FIX — CLARITY
+        5. MUST-FIX -- CLARITY
            List passages where meaning is genuinely obscured, transitions are dropped, or
            threads are left dangling in a way that blocks reader comprehension. For each item:
-           - Quote or reference the passage
-           - State the concrete fix
+           - ORIGINAL: quote the offending passage verbatim
+           - PROBLEM: why it blocks comprehension
+           - FIX: the concrete correction
 
-        4. OPTIONAL SUGGESTIONS
+        6. OPTIONAL SUGGESTIONS
            Craft improvements that would strengthen the chapter but are NOT required for it to
-           pass. Label each suggestion clearly as optional. Do not inflate this section — only
-           include suggestions with a clear upside and low risk of voice damage.
-           ⚠️ Do NOT suggest adding thematic contrast metaphors, making dialogue more elegant,
-           or smoothing sentence variation — these are voice PRESERVATION issues, not fixes.
+           pass. Label each suggestion clearly as optional. Include the relevant quote.
+           Do not inflate this section -- only include suggestions with a clear upside and low
+           risk of voice damage.
+           Do NOT suggest adding thematic contrast metaphors, making dialogue more elegant,
+           or smoothing sentence variation -- these are voice PRESERVATION issues, not fixes.
 
-        5. FORBIDDEN CHANGES / NON-GOALS
-           List things that might appear to be problems but should NOT be changed — intentional
-           voice choices, genre conventions, structural decisions that are working. This section
-           protects the draft from over-editing.
-           ALWAYS list here: verbal tics, repeated phrases, and "imperfect" speech that is
+        7. FORBIDDEN CHANGES / NON-GOALS
+           List things that might appear to be problems but should NOT be changed -- intentional
+           voice choices, genre conventions, structural decisions that are working.
+           ALWAYS include: verbal tics, repeated phrases, and "imperfect" speech that is
            clearly a character signature, not an error.
 
-        6. VERDICT
+        8. VERDICT
            One of: PASS / REVISE
+           SCORE ANCHORS -- be honest, scores that cluster at 90-96 without strong evidence
+           signal a rubber-stamp review and will be rejected by the adjudicator:
+             95-100: Zero MUST-FIX items. PROSE EVIDENCE quotes show above-average craft.
+                     Voice audit finds no violations.
+             85-94:  1-2 minor MUST-FIX items. At least one PROSE EVIDENCE quote shows
+                     a clear weakness with explanation.
+             70-84:  Clear structural or voice problems. Every MUST-FIX item is quoted
+                     verbatim with a rewrite suggestion.
+             Below 70: Multiple serious violations. Every failing passage is quoted and
+                        corrected. Automatic REVISE verdict.
+
+           State your score and a one-sentence justification citing your evidence.
            Use PASS if only optional improvements remain.
            Use REVISE if any MUST-FIX items are present.
-           Reserve judgment on full rewrites — that decision belongs to the roundtable.
 
   - type: document
     filename: "Chapter_{chapter_number}_review_{review_letter}"
@@ -89,12 +118,16 @@ adjudication:
   pass_threshold: 60
   deliverable_type: editorial_review
   criteria:
-    specificity:
-      weight: 40
-      description: "Review cites specific passages, scenes, or structural elements in all sections — not vague general commentary"
-    actionability:
+    prose_evidence:
       weight: 35
-      description: "Each MUST-FIX item includes a concrete, implementable correction; OPTIONAL items are clearly labeled"
+      description: "Review contains 3+ verbatim quoted passages with inline commentary. Zero quotes = automatic fail. Paraphrases do not count as quotes."
+    specificity:
+      weight: 30
+      description: "Every MUST-FIX item includes the original passage quoted verbatim AND a concrete rewrite or correction. Generic commentary without quotes scores zero here."
+    voice_audit:
+      weight: 20
+      description: "CHARACTER VOICE AUDIT section checks each speaking character against their profile constraints. Each violation is quoted and the broken rule named explicitly."
     coverage:
-      weight: 25
-      description: "All six sections are present; STRENGTHS and FORBIDDEN sections are populated, not left empty"
+      weight: 15
+      description: "All eight sections present and populated. STRENGTHS section includes evidence. FORBIDDEN section is not left empty."
+