Phase 2 — Systems

Week 5 — Retrieval Pipeline

Build ingestion, normalization, chunking, indexing, querying, reranking, and context assembly as separately testable stages.

Week 6 — Grounded Generation

Connect selected evidence to structured answers. Preserve source identity and design behavior for missing, weak, and conflicting evidence.

Week 7 — Tool Contracts

Create narrow, typed tools. Separate read and write permissions. Validate arguments, sanitize outputs, and verify state changes.

Week 8 — Agent Control Loops

Implement observe, decide, act, and verify. Add step, time, and cost budgets plus explicit stop conditions.

Week 9 — Evaluation Harness

Unify cases, configurations, traces, graders, comparisons, and error review. Make every aggregate result traceable to examples.

Phase artifact

A retrieval-assisted tool-using system with a reproducible evaluation suite, execution traces, and documented permission boundaries.