Calibration Persona Suite — Skill 94
Purpose
Design and calibrate against behavioral solver archetypes, not just generic tiers. Real agents have behavioral patterns — a "reckless fast solver" and a "careful planner" might have similar ELO but approach challenges completely differently.
The 8 Calibration Personas
1. The Speedrunner
- Behavior: Reads minimally, codes immediately, submits after first passing test
- Strengths: Fast time-to-solution, sometimes lucky
- Weaknesses: Shallow testing, misses hidden invariants, poor docs
- Expected archetypes: premature_convergence, visible_test_overfitting
- Good challenge response: Visible tests pass quickly (feels like success), adversarial tests destroy the submission
2. The Polished Mediocre
- Behavior: Clean, well-documented code that's architecturally wrong
- Strengths: High code quality, good documentation, follows conventions
- Weaknesses: Wrong approach to core problem, doesn't verify assumptions
- Expected archetypes: false_confidence_hallucination, shallow_decomposition
- Good challenge response: Code LOOKS excellent but fails hidden tests
3. The Tool Spammer
- Behavior: Runs every tool repeatedly, reads everything, massive context
- Strengths: Thorough information gathering
- Weaknesses: Wastes time, doesn't synthesize, incomplete submission
- Expected archetypes: scope_explosion, false_confidence_stop (time)
- Good challenge response: Enough files to tempt exhaustive reading, rewards targeted investigation
4. The Careful Planner
- Behavior: 40% reading/planning, then methodical execution
- Strengths: Good decomposition, systematic approach
- Weaknesses: Slow to pivot if plan was wrong
- Expected archetypes: strategic_myopia (if plan wrong and no pivot)
- Good challenge response: Rewards planning but includes recovery branch forcing adaptation
5. The Exploit Seeker
- Behavior: Looks for shortcuts — read tests? Hardcode? Game judges?
- Strengths: Creative lateral thinking (when honest)
- Weaknesses: Integrity violations, fragile solutions
- Expected archetypes: integrity_degradation
- Good challenge response: Exploit temptations that are detectable, rewards honest flagging
6. The Honest Conservative
- Behavior: Acknowledges uncertainty, flags issues, makes safe choices
- Strengths: High integrity, accurate self-assessment
- Weaknesses: May underperform on objective metrics (too cautious)
- Expected archetypes: false_confidence_stop (stops too early)
- Good challenge response: Integrity bonuses reward this behavior, ambiguity handled better than overconfident agents
7. The Recovery Specialist
- Behavior: Mediocre first attempt, steep iteration trajectory
- Strengths: Excellent error diagnosis, strong recovery
- Weaknesses: Slow start
- Expected archetypes: None (avoids common archetypes)
- Good challenge response: Multiple iterations, recovery branches that reward this pattern
8. The Brute Forcer
- Behavior: Tries many approaches rapidly, no deep analysis
- Strengths: Eventually stumbles on working solution
- Weaknesses: Messy code, fragile solution, no understanding
- Expected archetypes: toolchain_misuse, context_drift
- Good challenge response: Dynamic adversarial tests catch brute-forced solutions
Using Personas in Design
For every challenge, mental simulation: "How would each of the 8 personas approach this?"
| Result |
Interpretation |
Action |
| 3+ personas score identically |
Challenge doesn't discriminate behavioral patterns |
Redesign |
| One persona dominates |
Challenge only tests one approach |
Add elements rewarding other strengths |
| Each persona gets different score matching engineering quality |
Ideal |
Publish |
Using Personas in Calibration
Run calibration with persona-configured agents (not just naive/standard/elite):
- Speedrunner agent (low read time, fast submit)
- Tool Spammer agent (reads everything, high tool calls)
- Careful Planner agent (long planning phase)
- Recovery Specialist agent (mediocre first attempt, strong iteration)
If Speedrunner and Careful Planner score the same → challenge isn't measuring process quality.
Integration Points
- Calibration Packaging (Skill 81): Persona agents extend the 4-tier calibration
- Per-Challenge Failure Taxonomy (Skill 80): Personas predict specific failure patterns
- Anti-Convergence (Skill 72): Persona diversity is an anti-convergence mechanism
- Discrimination Science (Skill 46): Persona spread predicts CDI
1---2name: calibration-persona-suite3description: Calibration Persona Suite — Skill 944---5# Calibration Persona Suite — Skill 9467## Purpose8Design and calibrate against behavioral solver archetypes, not just generic tiers. Real agents have behavioral patterns — a "reckless fast solver" and a "careful planner" might have similar ELO but approach challenges completely differently.910## The 8 Calibration Personas1112### 1. The Speedrunner13- **Behavior**: Reads minimally, codes immediately, submits after first passing test14- **Strengths**: Fast time-to-solution, sometimes lucky15- **Weaknesses**: Shallow testing, misses hidden invariants, poor docs16- **Expected archetypes**: premature_convergence, visible_test_overfitting17- **Good challenge response**: Visible tests pass quickly (feels like success), adversarial tests destroy the submission1819### 2. The Polished Mediocre20- **Behavior**: Clean, well-documented code that's architecturally wrong21- **Strengths**: High code quality, good documentation, follows conventions22- **Weaknesses**: Wrong approach to core problem, doesn't verify assumptions23- **Expected archetypes**: false_confidence_hallucination, shallow_decomposition24- **Good challenge response**: Code LOOKS excellent but fails hidden tests2526### 3. The Tool Spammer27- **Behavior**: Runs every tool repeatedly, reads everything, massive context28- **Strengths**: Thorough information gathering29- **Weaknesses**: Wastes time, doesn't synthesize, incomplete submission30- **Expected archetypes**: scope_explosion, false_confidence_stop (time)31- **Good challenge response**: Enough files to tempt exhaustive reading, rewards targeted investigation3233### 4. The Careful Planner34- **Behavior**: 40% reading/planning, then methodical execution35- **Strengths**: Good decomposition, systematic approach36- **Weaknesses**: Slow to pivot if plan was wrong37- **Expected archetypes**: strategic_myopia (if plan wrong and no pivot)38- **Good challenge response**: Rewards planning but includes recovery branch forcing adaptation3940### 5. The Exploit Seeker41- **Behavior**: Looks for shortcuts — read tests? Hardcode? Game judges?42- **Strengths**: Creative lateral thinking (when honest)43- **Weaknesses**: Integrity violations, fragile solutions44- **Expected archetypes**: integrity_degradation45- **Good challenge response**: Exploit temptations that are detectable, rewards honest flagging4647### 6. The Honest Conservative48- **Behavior**: Acknowledges uncertainty, flags issues, makes safe choices49- **Strengths**: High integrity, accurate self-assessment50- **Weaknesses**: May underperform on objective metrics (too cautious)51- **Expected archetypes**: false_confidence_stop (stops too early)52- **Good challenge response**: Integrity bonuses reward this behavior, ambiguity handled better than overconfident agents5354### 7. The Recovery Specialist55- **Behavior**: Mediocre first attempt, steep iteration trajectory56- **Strengths**: Excellent error diagnosis, strong recovery57- **Weaknesses**: Slow start58- **Expected archetypes**: None (avoids common archetypes)59- **Good challenge response**: Multiple iterations, recovery branches that reward this pattern6061### 8. The Brute Forcer62- **Behavior**: Tries many approaches rapidly, no deep analysis63- **Strengths**: Eventually stumbles on working solution64- **Weaknesses**: Messy code, fragile solution, no understanding65- **Expected archetypes**: toolchain_misuse, context_drift66- **Good challenge response**: Dynamic adversarial tests catch brute-forced solutions6768## Using Personas in Design6970For every challenge, mental simulation: "How would each of the 8 personas approach this?"7172| Result | Interpretation | Action |73|--------|---------------|--------|74| 3+ personas score identically | Challenge doesn't discriminate behavioral patterns | Redesign |75| One persona dominates | Challenge only tests one approach | Add elements rewarding other strengths |76| Each persona gets different score matching engineering quality | **Ideal** | Publish |7778## Using Personas in Calibration7980Run calibration with persona-configured agents (not just naive/standard/elite):81- Speedrunner agent (low read time, fast submit)82- Tool Spammer agent (reads everything, high tool calls)83- Careful Planner agent (long planning phase)84- Recovery Specialist agent (mediocre first attempt, strong iteration)8586If Speedrunner and Careful Planner score the same → challenge isn't measuring process quality.8788## Integration Points8990- **Calibration Packaging** (Skill 81): Persona agents extend the 4-tier calibration91- **Per-Challenge Failure Taxonomy** (Skill 80): Personas predict specific failure patterns92- **Anti-Convergence** (Skill 72): Persona diversity is an anti-convergence mechanism93- **Discrimination Science** (Skill 46): Persona spread predicts CDI