#Model & Cost Optimization (WI-10)

#Context

Retroactive challenge document for AI model routing and cost optimization work. This covers task-based model selection (Gemini-3-flash promoted to default), token optimization strategies, DeepSeek adapter integration, post-generation patching, and cost tracking. All work is complete — this doc exists for rollout tracking consistency.

#Current State

• All 5 challenges implemented and verified
• Gemini-3-flash is default model for most extraction tasks
• Token usage reduced ~63% via STYLE_RULES deduplication
• Cost tracking operational with per-model budgets and daily caps

#Challenges

#Challenge 10.1: Model Routing

Requirement: Task-based tier selection for AI model usage. Acceptance Criteria:

• Model routing config maps task types to model tiers
• Gemini-3-flash promoted to default for standard extraction
• Higher-tier models reserved for validation and complex reasoning
• Fallback chain when primary model is unavailable Evaluation: PASS

#Challenge 10.2: Token Optimization

Requirement: Reduce token consumption without sacrificing output quality. Acceptance Criteria:

• STYLE_RULES deduplication achieves ~63% token reduction
• Micro-batch amortization spreads system prompt cost across multiple facts
• Thinking budgets tiered by task complexity (0/1024/4096 tokens)
• Evergreen title dedup cap prevents redundant regeneration Evaluation: PASS

#Challenge 10.3: DeepSeek Adapter

Requirement: Integrate DeepSeek models for specific task types. Acceptance Criteria:

• Free-text JSON bypass for DeepSeek's non-standard JSON output
• v4 adapter handles DeepSeek-specific response format
• Graceful fallback to primary model on adapter failure Evaluation: PASS

#Challenge 10.4: Post-Generation Patching

Requirement: Automated quality fixes applied after AI generation. Acceptance Criteria:

• Passive voice detector and rewriter (36 patterns)
• Generic reveal family matcher (21 families)
• Textbook register rewriter for overly formal language
• Patches applied before drift coordinator validation Evaluation: PASS

#Challenge 10.5: Cost Tracking

Requirement: Per-model budget enforcement and spend monitoring. Acceptance Criteria:

• Per-model budget caps (configurable per task type)
• Daily spend caps with automatic throttling
• Reasoning token tracking (separate from completion tokens)
• Cost attribution by task type for reporting Evaluation: PASS

#Implementation Notes

• Model routing config in packages/ai/ — task → model tier mapping
• Token optimization strategies are composable and can be toggled per task
• DeepSeek adapter in packages/ai/ — handles v4 protocol differences
• Post-generation patching runs as a pipeline stage between generation and drift validation
• Cost tracking integrates with the admin dashboard (Challenge 6.9) for visibility