Voice Fitness Check

Table of Contents

• Analogy direction priority
• Workflow
• Tier ratings
• Worked example
• Guardrails

Related skills: Called by the Intuition Builder in the pipeline (step 5). Reads shared-context/voice-profile.md and shared-context/voices/{section}.md if a target section is specified.

Analogy direction priority

From voice-profile.md section 9:

1. Biology → AI architecture and emergence (immune system, thymus, crypts, DNA/protein, ant colony) — Tier 1.
2. Organizational / institutional → multi-agent systems (departments, courtroom, bucket brigade) — Tier 2.
3. Sports → calibration / prediction / betting (cricket phases, soccer-as-practice, scoreboard) — Tier 3.

Everything else is Tier 4 (acceptable but not a voice signature).

Physics, military, warfare, weapons, combat — these are voice violations. Flag strongly. Propose a replacement in Tier 1–3 direction if one is available.

Workflow

For each of the 5 framings:
- [ ] Step 1: Classify the source domain
- [ ] Step 2: Assign a tier (1-4) based on the priority table
- [ ] Step 3: Flag any Tier 4 or voice-violating framings
- [ ] Step 4: Rank the 5 framings overall (tier + craft quality)
- [ ] Step 5: Recommend first choice

Step 1: Domain classification

Look at the source. Is it:

• A living system, cellular process, evolution, immunology? → biology (Tier 1)
• A human organization, hierarchy, committee, courtroom, workflow? → organizational (Tier 2)
• A sport, game, match, contest, calibration of a prediction? → sports (Tier 3)
• A machine, mechanical system, fluid, physical law, vehicle? → physics (Tier 4)
• A battle, weapon, combat, tactic, strategy (in the military sense)? → voice violation

"Neural network" itself counts as biology — but only if the framing actually uses biological relations (neurons firing, learning as development), not just the name.

Tier ratings

• Tier 1 (biology): preferred, voice signature. Use often.
• Tier 2 (organizational): preferred, voice signature. Use when biology doesn't fit.
• Tier 3 (sports): acceptable for calibration/prediction content specifically. Out of register for pure ML architecture.
• Tier 4 (other): usable if no Tier 1-3 candidate fits, but not a voice signature.
• Violation (physics/military): flag. Do not include without explicit writer override.

Worked example

Topic: Multi-agent AI system.

5 framings (from generate-analogy-set):

1. Everyday: "A potluck dinner where each guest brings one dish" — Tier 4 (household).
2. Physical metaphor: "Distributed load across power-grid nodes" — Tier 4 (physics; voice-adjacent violation).
3. Contrarian: "Not a single super-agent; a parliament of specialized agents" — Tier 2 (organizational).
4. Historical: "The Sumerian scribe running grain accounting — system intelligence exceeding any individual" — Tier 2 (borrowed, attributed).
5. Counterfactual: "Remove inter-agent communication; you're back to single-agent with longer prompts" — neutral (structural, not an analogy in the usual sense).

Ranking:

1. Contrarian (parliament) — Tier 2, strong fit with writer's voice signature for multi-agent.
2. Historical (Sumerian scribe) — Tier 2, attributed borrow, strong.
3. Counterfactual — neutral; always welcome.
4. Everyday (potluck) — Tier 4, usable but not voice-native.
5. Physical (power grid) — Tier 4 with voice-adjacent risk. Suggest replacement: "An immune system's parallel T-cell response" (Tier 1, biology).

Recommendation: First choice is the parliament contrarian. Swap framing 5 for the biological alternative.

Guardrails

1. Tier assignments are not opinions — they come directly from the voice-profile's analogy-direction table.
2. Flag voice violations (physics/military) strongly. Do not merely rank low; call them out.
3. Suggest a replacement for any violation, drawn from Tier 1-3.
4. Ranking combines tier + craft quality. A Tier-2 framing with a weak mapping ranks below a Tier-3 with strong mapping. Use judgement for ties.
5. Counterfactual framings are usually neutral — they're structural, not source-domain-based. Rate on mechanical clarity instead.
6. If the target section has its own voice overlay (voices/{section}.md) with different priorities, apply those on top of global.

Quick reference

• Input: 5 framings + voice-profile.md.
• Output: per-framing tier + overall ranking + flagged violations + replacement suggestions.
• Final step before the Intuition Builder presents to the writer.