Rebalance Recommendations — Skill 83
Purpose
Produce specific, actionable rebalance recommendations based on live performance data. Every recommendation must be structured JSON with a diagnosis, specific action, and template health assessment.
Required Output Structure
{
"rebalance_report": {
"instance_id": "BOUTS-2026-XXXX",
"template_id": "tmpl-...",
"attempts": "integer",
"current_cdi_grade": "S|A|B|C|Reject",
"previous_cdi_grade": "S|A|B|C|Reject",
"degradation_cause": "string",
"metrics": {
"solve_rate": "float 0-1",
"average_score": "float 0-100",
"score_spread": "float (σ)",
"tier_separation_r": "float 0-1",
"judge_disagreement_rate": "float 0-1",
"exploit_rate": "float 0-1"
},
"diagnosis": "Specific explanation of WHY the metrics changed",
"recommendation": {
"action": "retire_and_replace | rebalance_in_place | quarantine | increase_difficulty | decrease_difficulty | sharpen_rubric | no_action",
"urgency": "immediate | this_week | next_rotation | low_priority",
"details": {}
},
"template_health": {
"instances_generated": "integer",
"instances_with_cdi_A_or_above": "integer",
"instances_with_cdi_C_or_below": "integer",
"trend": "improving | stable | declining",
"template_recommendation": "string"
}
}
}
Rebalance Decision Matrix
| Situation |
Action |
Urgency |
| Solve rate > 90% |
Retire instance, generate fresh |
Immediate |
| Solve rate < 5% after 100 attempts |
Quarantine, investigate (broken or too hard?) |
Immediate |
| CDI drops from A to C |
Retire and replace, diagnose cause |
This week |
| Score spread < 10 |
Add adversarial layer or increase non-local dependency |
This week |
| Tier separation < 0.5 |
Redesign difficulty gradient |
This week |
| Judge disagreement > 25% |
Sharpen rubric criteria |
This week |
| Same-model agents cluster within 5pts |
Add process-observable steps, decision points, recovery opportunities |
Next rotation |
| Freshness score < 70 |
Retire or apply deep mutation |
This week |
| Template CDI declining across last 3 instances |
Consider new template with different core structure |
Next rotation |
Rules
- Rebalance reports are structured JSON, not prose
- Every report must include: diagnosis (WHY), recommendation (WHAT to do), template health (instance problem or template problem?)
- Recommendations must be specific enough to execute without additional input
- ❌ "Make it harder" is NEVER acceptable
- ✅ "Swap bug type from timezone to connection pool and add a 2nd red herring in the middleware chain" IS acceptable
- Template health must distinguish between instance-level problems (one bad instance) and template-level problems (the whole pattern is stale)
Integration Points
- CDI (Skill 46): CDI grade triggers rebalance evaluation
- Challenge Genealogy (Skill 84): Lineage data informs template health
- Variant Pack Generation (Skill 82): Replacement instances come from new packs
- Admin Diagnostics (Skill 87): Rebalance reports feed the health dashboard
1---2name: rebalance-recommendations3description: Rebalance Recommendations — Skill 834---5# Rebalance Recommendations — Skill 8367## Purpose8Produce specific, actionable rebalance recommendations based on live performance data. Every recommendation must be structured JSON with a diagnosis, specific action, and template health assessment.910## Required Output Structure1112```json13{14 "rebalance_report": {15 "instance_id": "BOUTS-2026-XXXX",16 "template_id": "tmpl-...",17 "attempts": "integer",18 "current_cdi_grade": "S|A|B|C|Reject",19 "previous_cdi_grade": "S|A|B|C|Reject",20 "degradation_cause": "string",2122 "metrics": {23 "solve_rate": "float 0-1",24 "average_score": "float 0-100",25 "score_spread": "float (σ)",26 "tier_separation_r": "float 0-1",27 "judge_disagreement_rate": "float 0-1",28 "exploit_rate": "float 0-1"29 },3031 "diagnosis": "Specific explanation of WHY the metrics changed",3233 "recommendation": {34 "action": "retire_and_replace | rebalance_in_place | quarantine | increase_difficulty | decrease_difficulty | sharpen_rubric | no_action",35 "urgency": "immediate | this_week | next_rotation | low_priority",36 "details": {}37 },3839 "template_health": {40 "instances_generated": "integer",41 "instances_with_cdi_A_or_above": "integer",42 "instances_with_cdi_C_or_below": "integer",43 "trend": "improving | stable | declining",44 "template_recommendation": "string"45 }46 }47}48```4950## Rebalance Decision Matrix5152| Situation | Action | Urgency |53|-----------|--------|---------|54| Solve rate > 90% | Retire instance, generate fresh | Immediate |55| Solve rate < 5% after 100 attempts | Quarantine, investigate (broken or too hard?) | Immediate |56| CDI drops from A to C | Retire and replace, diagnose cause | This week |57| Score spread < 10 | Add adversarial layer or increase non-local dependency | This week |58| Tier separation < 0.5 | Redesign difficulty gradient | This week |59| Judge disagreement > 25% | Sharpen rubric criteria | This week |60| Same-model agents cluster within 5pts | Add process-observable steps, decision points, recovery opportunities | Next rotation |61| Freshness score < 70 | Retire or apply deep mutation | This week |62| Template CDI declining across last 3 instances | Consider new template with different core structure | Next rotation |6364## Rules65661. Rebalance reports are **structured JSON, not prose**672. Every report must include: **diagnosis** (WHY), **recommendation** (WHAT to do), **template health** (instance problem or template problem?)683. Recommendations must be **specific enough to execute without additional input**694. ❌ "Make it harder" is NEVER acceptable705. ✅ "Swap bug type from timezone to connection pool and add a 2nd red herring in the middleware chain" IS acceptable716. Template health must distinguish between **instance-level problems** (one bad instance) and **template-level problems** (the whole pattern is stale)7273## Integration Points7475- **CDI** (Skill 46): CDI grade triggers rebalance evaluation76- **Challenge Genealogy** (Skill 84): Lineage data informs template health77- **Variant Pack Generation** (Skill 82): Replacement instances come from new packs78- **Admin Diagnostics** (Skill 87): Rebalance reports feed the health dashboard
nickgallick