Day 2

Day 2 evolves the loop in stage order, measures model choices, and finishes with one unattended challenge episode.

Morning

The morning is the merge train: evolve the loop in stage order through reviewed PRs.

1. Starter: CLAUDE.md safari: hunt for one convention worth adopting.
2. Stage 1: dispatcher: configure yesterday’s agent loop into the dispatcher and run it live.
3. Stage 2: intake: turn natural-language contacts into structured reports.
4. Stage 3: reconciliation: insert deterministic matching above incident CRUD.
5. Stage 4: replanning: handle contradictions, corrections, redirects, and closure.

Afternoon

The afternoon proves the loop: compare models, bound the agent, and run the finale.

1. Model comparisons: compare each role’s model with one alternative on frozen fixtures and repeated game runs.
2. Presentation and finale: publish your solution page, show what made the difference, and run the challenge map.

Merge train

Each stage is a reviewed scenario-sized PR, and nobody merges their own - a different member reviews and merges each. Merge in order because later mechanics depend on earlier contracts, then stop feature work for the integration buffer before the finale.