Phase 3 — Production

Week 10 — Reliability

Allocate end-to-end deadlines, classify retryable errors, use backoff with jitter, and design safe fallback and abstention behavior.

Week 11 — Observability & Cost

Trace model calls, retrieval stages, tool actions, validation, token use, and latency. Build a cost model tied to workload volume and quality.

Week 12 — Safety & Operations

Model threats across inputs, context, tools, outputs, and stored data. Add least privilege, auditability, redaction, and operational runbooks.

Week 13 — Integration & Review

Run the complete system against the frozen task suite. Perform failure injection, compare with the baseline, simplify weak components, and document the architecture and operating envelope.

Final artifact

A deployable system with:

• explicit component contracts;
• a versioned evaluation suite;
• end-to-end traces;
• bounded latency and cost;
• failure and recovery paths;
• a decision log explaining the final design.

Day 90 reflection

Write three things: what the system does reliably, where it remains uncertain, and which experiment would reduce the most important uncertainty next.