Carlos's conversation uses a chosen model

[danielnaab]

User Story

As Maya, in order to pick which LLM drives Carlos's conversational form-filling experience and see how different models perform on adaptive interviewing, I want filling variants to be selectable from Settings → Variants, backed by persona-driven evaluation.

Preconditions

• Story 10: Maya chooses how her form was extracted #58 (Story 10 variant picker) merged to main
• Carlos completes complex sections through conversation #9 (Story 9 conversational sections) merged to main

Acceptance Criteria

• New evaluation kind filling-interview scores a variant against scripted personas (completeness of collected fields + conditional routing accuracy), mirroring the assignment-10 TextGrad evaluator pattern
• Variants registered: filling/haiku, filling/sonnet, filling/opus
• Filling tab in Settings → Variants renders all three variants
• Session records variantId at start (new column on form_sessions table)
• <VariantBadge task="filling" ...> rendered on conversation UI
• Persona fixtures committed (5 train + 3 test minimum)
• New catalog suite catalog/experiments/filling-model-comparison/ with _suite.md + one markdown per variant, each with metrics + findings
• catalog/experiments/_roadmap.md updated with shipped status and one-line finding

Success Metrics

• Observable tradeoff across variants (e.g., Opus higher completeness, Haiku faster/cheaper)
• Persona-driven eval runs reproducibly from bun run cli evaluate run <variant>

Notes

• Port evaluator shape from llm-class-2026-winter-cohort/notes/assignment-10/ — persona-driven simulator + completeness scorer
• Filling registry stub exists at src/services/forms/filling/registry.ts (empty); extend it

Definition of Done

• Acceptance criteria met
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable