Heady Resonance Studio

Instructions

1. Define the Resonance Model

resonance_model:
  response:
    id: uuid
    source: headybuddy-core | heady-coder | heady-analyze | other
    prompt: the input that generated the response
    content: the response text
    persona: active persona at generation time
    context: relevant conversation history

  quality_dimensions:
    - name: accuracy
      description: factual correctness and relevance to the prompt
      scale: 1-5
      weight: 0.25

    - name: tone_alignment
      description: match between response tone and active persona
      scale: 1-5
      weight: 0.20

    - name: completeness
      description: covers all aspects of the user's need
      scale: 1-5
      weight: 0.20

    - name: conciseness
      description: appropriate length without unnecessary padding
      scale: 1-5
      weight: 0.15

    - name: helpfulness
      description: actionable and useful to the user
      scale: 1-5
      weight: 0.15

    - name: safety
      description: free from harmful, biased, or inappropriate content
      scale: binary (pass/fail)
      weight: gate (fail blocks everything)

  composite_score: weighted sum of dimensions (0.0-1.0)
  threshold: { acceptable: 0.7, good: 0.85, excellent: 0.95 }

2. Build the Evaluation Pipeline

Three evaluation modes using Heady tools:

Automated Evaluation (heady-critique):

1. Response generated by any Heady AI surface
2. mcp_Heady_heady_critique(code="{response}", criteria="accuracy, tone, completeness, conciseness, helpfulness, safety")
3. heady-critique returns per-dimension scores + rationale
4. Scores logged to heady-metrics, response+scores stored in HeadyMemory
5. If composite_score < 0.7: flag for human review

Comparative Evaluation (heady-battle):

1. Generate two response variants (different models, prompts, or personas)
2. mcp_Heady_heady_battle(action="evaluate", participants=["variant_a", "variant_b"], criteria="quality dimensions")
3. heady-battle returns winner + per-dimension comparison
4. Winning patterns fed to heady_soul(action="learn")
5. Results stored in HeadyMemory for pattern analysis

Human Feedback Loop:

1. User provides feedback (thumbs up/down, text correction, style preference)
2. Feedback logged to HeadyMemory with response context
3. mcp_Heady_heady_vinci(data="{feedback + response}", action="learn") → pattern extraction
4. Patterns stored in heady-patterns as style refinements
5. heady_soul(content="{feedback context}", action="learn") → long-term adaptation

3. Design Persona Calibration

Tune persona behavior through the Resonance Studio:

persona_calibration:
  target_persona: persona being calibrated
  calibration_set:
    - prompt: representative user input
      expected_tone: how this persona should sound
      expected_approach: how this persona should handle this type of request
      boundary: what this persona should NOT do

  calibration_process:
    1. Generate responses to calibration set using target persona
    2. Evaluate each response via heady-critique against expected behavior
    3. Compare via heady-battle against baseline persona responses
    4. Identify gaps between expected and actual behavior
    5. Generate pattern adjustments via heady-vinci
    6. Store refined patterns in heady-patterns under persona namespace
    7. Re-evaluate with calibration set to confirm improvement

  calibration_dimensions:
    vocabulary: words and phrases characteristic of this persona
    sentence_structure: length, complexity, rhythm
    emotional_register: warmth, formality, enthusiasm level
    domain_expertise: how deeply persona engages with specific topics
    boundaries: topics persona redirects or declines

  tracking:
    metric: persona_alignment_score via heady-metrics
    trend: heady-vinci tracks calibration effectiveness over time
    alert: heady-observer fires if alignment drops below 0.8

4. Build the Style Template System

Managed through heady-patterns:

style_templates:
  storage: heady-patterns library
  format:
    name: template-name
    type: tone | format | domain | persona-overlay
    rules:
      - dimension: vocabulary
        guidance: "Use active voice, avoid jargon unless user is technical"
      - dimension: length
        guidance: "Responses under 3 sentences for simple questions"
      - dimension: structure
        guidance: "Lead with the answer, then provide context"

  built_in_templates:
    - name: concise-technical
      type: tone
      rules: [direct, code-first, minimal explanation, assume expertise]

    - name: warm-supportive
      type: tone
      rules: [encouraging, explanatory, empathetic, celebrate progress]

    - name: nonprofit-professional
      type: domain
      rules: [mission-centered language, data-backed claims, funder-aware framing]

    - name: report-format
      type: format
      rules: [structured sections, bullet points, executive summary first]

  custom_templates:
    creation: user defines via Resonance Studio UI on HeadyWeb
    learning: heady-vinci extracts templates from user's preferred responses
    sharing: templates can be shared within HeadyConnection workspace
    versioning: heady-traces tracks template changes over time

  application:
    1. Active persona selects applicable templates
    2. Templates merged with persona calibration data
    3. Combined guidance shapes response generation
    4. heady-critique validates output against applied templates

5. Implement the Feedback Learning Loop

Continuous improvement cycle powered by heady-vinci and heady-soul:

learning_loop:
  collection:
    implicit:
      - user edits Buddy's response → preference signal
      - user asks follow-up for clarification → completeness signal
      - user ignores response → negative engagement signal
      - user shares/reuses response → positive quality signal
      sources: heady-observer tracks engagement signals

    explicit:
      - thumbs up/down on responses
      - text corrections with "I prefer this instead"
      - style preferences ("be more concise", "explain more")
      - persona feedback ("too formal", "not helpful enough")
      sources: HeadyBuddy UI, HeadyWeb, heady-buddy-portal

  processing:
    1. Feedback events aggregated in heady-metrics
    2. mcp_Heady_heady_vinci(data="{feedback batch}", action="recognize") → pattern identification
    3. Patterns categorized: style preference, accuracy issue, persona misalignment, knowledge gap
    4. Style patterns → update heady-patterns templates
    5. Accuracy issues → flag for content review
    6. Persona misalignment → trigger persona recalibration
    7. Knowledge gaps → flag for training data update
    8. All outcomes → heady_soul(action="learn") for long-term improvement

  measurement:
    quality_trend: heady-metrics tracks composite_score over time
    satisfaction: heady-metrics tracks explicit feedback ratio (positive/total)
    engagement: heady-metrics tracks response engagement rate
    calibration_drift: heady-observer alerts when persona alignment degrades
    learning_velocity: how quickly quality improves after feedback

6. Design the Resonance Studio Dashboard

HeadyWeb interface for response quality management:

Studio workflows:

• Quick Tune: Select a response → edit to preferred version → one-click feedback to heady-soul
• Batch Evaluate: Upload prompt set → generate responses → heady-critique scores all → review outliers
• A/B Test: Define two variants → heady-battle evaluates → winner patterns learned automatically
• Persona Workshop: Select persona → run calibration set → review gaps → apply pattern updates

Output Format

When designing Resonance Studio features, produce:

1. Resonance model with quality dimensions, scoring, and thresholds
2. Evaluation pipeline with automated, comparative, and human feedback modes
3. Persona calibration workflow with dimension tracking
4. Style template system using heady-patterns
5. Learning loop with feedback collection, processing, and measurement
6. Dashboard design with studio workflows

Tips

• heady-critique evaluates, heady-battle compares, heady-soul learns — use all three in the quality pipeline
• Implicit feedback is more honest than explicit — user edits and follow-ups reveal true preferences better than thumbs up/down
• Persona calibration is ongoing — language preferences drift over time; schedule periodic recalibration
• heady-patterns is the style library — store every approved pattern; the richer the library, the more consistent the output
• Quality is multidimensional — a response can be accurate but unhelpful, or concise but incomplete; measure each dimension
• Safety is a gate, not a score — safety failures block everything regardless of other quality dimensions