Themis Synthesizer

Synthesis Process

Step 1: Component Score Calculation

Use the merge_scores.py utility or calculate manually:

python3 scripts/merge_scores.py --content-council '<json>' --market-council '<json>' --critic '<json>'

For each component in the final output:

Step 2: Overall Virality Score

Weighted sum of components:

• hook_effectiveness: 25%
• emotional_resonance: 20%
• production_quality: 15%
• trend_alignment: 20%
• shareability: 20%

virality_score = round(
    hook * 0.25 +
    emotion * 0.20 +
    production * 0.15 +
    trend * 0.20 +
    shareability * 0.20
)

Step 3: Tier Assignment

Step 4: Confidence Calculation

Base confidence = minimum of all 6 judges' confidence scores. Apply Critic's overall_confidence_adjustment (capped at -0.20 to +0.10). Clamp to [0.0, 1.0].

Step 5: Distribution Section

Pull from Market Council consensus:

• primary_audiences — from Audience Mapper, enriched by Subject Analyst's classifications
• geographic_reach — from Audience Mapper
• recommended_strategy — merged from Audience Mapper + Trend Analyst timing

Step 5.5: Authenticity Section

Build the authenticity section from the Authenticity Analyst's output:

1. verdict: Copy from Authenticity Analyst (likely_human, likely_ai, mixed, uncertain)
2. confidence: Copy the Authenticity Analyst's confidence score
3. ai_probability: Copy from Authenticity Analyst (0.0-1.0)
4. indicators: Copy the list of indicators with type, signal, direction, and weight
5. statistical_metrics: Copy from forensics data (burstiness, TTR, hedging, entropy, paragraph CV, readability variance, transition frequency, composite probability)
6. caveat: Must always include the mandatory disclaimer about AI detection limitations

Important: The authenticity section is a peer of virality, not nested within it. Authenticity scores do NOT factor into the virality score calculation.

If no Authenticity Analyst output is available, omit the authenticity section entirely (backward compatible).

Step 5.9: Rationale Blurbs

Generate a 1-2 sentence rationale TLDR for each of the three main output sections. These give the user an at-a-glance explanation of why each conclusion was reached.

1. virality.rationale — Explain the virality score by citing the strongest and weakest components. Example: "Scored 72 (strong) primarily due to an exceptional hook and strong trend alignment, held back by moderate shareability."

2. authenticity.rationale — Explain the verdict by citing the most decisive signals. Example: "Classified as likely human based on high sentence-length variance and distinctive voice, despite slightly elevated transition word frequency."

3. distribution.rationale — Explain the audience recommendation by citing the content characteristics that drive it. Example: "Best suited for Dev Twitter and Reddit/HN due to the technical subject matter and tutorial format, with strong LinkedIn crossover potential."

Tone: Direct, specific, no hedging. Cite the actual scores or signals that drove the conclusion.

Step 6: Reasoning Section

Executive Summary

Write a 3-5 sentence summary that:

1. States the overall verdict (virality tier + score)
2. Identifies the single biggest strength
3. Identifies the single biggest weakness
4. Gives one concrete, actionable recommendation

Tone: Professional, direct, actionable. Not marketing-speak. Not academic.

Strengths

Extract the top 3-5 strengths from across all judge outputs. Prioritize:

• Strengths that multiple judges independently noted
• Strengths with high confidence scores
• Strengths relevant to the virality score

Weaknesses

Extract the top 3-5 weaknesses. Include:

• Issues the Critic highlighted
• Low-scoring dimensions with specific explanations
• Missed opportunities judges identified

Improvement Suggestions

Produce 3-5 specific, actionable suggestions:

• Each must target a specific area (hook, emotion, production, trend, audience)
• Each must include a concrete suggestion (not "improve the hook" but "open with the product reveal instead of the logo animation")
• Each must estimate expected_impact (low/medium/high) based on how much it would shift the virality score

Council Disagreements

Include ALL disagreements where judges differed by >20 points on any dimension:

• State both positions with reasoning
• State the resolution (which position the synthesis favored, or why both are preserved)
• These are a feature — they represent genuine uncertainty that the user should know about

Step 7: Metadata

{
  "mode": "full | fast",
  "debate_rounds": 2,
  "total_tokens_used": 0,
  "estimated_cost_usd": 0.00,
  "judges_used": ["hook_analyst", "emotion_analyst", "production_analyst", "authenticity_analyst", "trend_analyst", "subject_analyst", "audience_mapper"],
  "evaluation_timestamp": "ISO 8601"
}

Token estimation: Sum all token counts reported by individual judge tasks and critic/synthesis steps.

Cost estimation:

• Sonnet tokens: input $3/MTok, output $15/MTok
• Opus tokens: input $15/MTok, output $75/MTok
• Haiku tokens: input $0.25/MTok, output $1.25/MTok

Output Quality Checklist

Before producing the final JSON, verify:

• All 5 component scores are present and 0-100
• Virality score matches weighted formula
• Tier matches score range
• Confidence is between 0.0 and 1.0
• At least 1 primary audience in distribution
• Each audience has all 3 platform fit scores
• Executive summary is 3-5 sentences
• At least 3 strengths and 3 weaknesses
• At least 3 improvement suggestions with area/suggestion/impact
• All disagreements >20 points are preserved
• Authenticity section present with verdict, confidence, ai_probability, indicators, metrics, caveat
• Authenticity scores are NOT factored into virality calculation
• All three rationale blurbs present (virality, authenticity, distribution) — each 1-2 sentences
• Metadata has mode, rounds, tokens, cost, judges, timestamp
• authenticity_analyst is included in judges_used list
• No null or missing required fields

Final Output

Produce the complete JSON conforming to the schema in skills/themis-evaluate/references/output-schema.md.

Also produce a human-readable narrative summary suitable for display to the user — this is the executive summary plus key scores and top recommendations formatted for quick reading.