Revision · Project 7 — Mosaic

Mosaic assembled the boundaries built throughout the book into one operable system.

What you can now build

• A typed CLI or HTTP boundary that rejects malformed work before inference.
• Text, image, audio, and video evidence with provenance and reversible locators.
• Provider-neutral scripted, local, and remote model capabilities.
• Structured output with syntactic, structural, semantic, and evidence validation.
• Narrow tools with effect classes, authorization, receipts, and untrusted observations.
• Bounded agent loops with progress, termination, repair, and reconciliation.
• Durable runs, leases, idempotency, replay, and content-addressed artifacts.
• Frozen evals, calibrated graders, routing, test-time compute, and tuning decisions.
• Backpressured async services with streaming, cancellation, security, telemetry, rollout, and rollback.

The final learning loop

build a useful behavior
  → run it
  → inspect the artifact
  → break one boundary
  → diagnose the first violated invariant
  → change one responsible layer
  → rerun the same tests and evals
  → explain the trade

That loop is the book. Rust and AI are not parallel theory tracks: every project forces the language and system concepts into the open at the moment they become necessary.

Keep the Deep Reference Library nearby when a project exposes a boundary you want to understand in greater depth.