Skip to content

Architecture

Start here if you want the public product shape before Runtime implementation details.

CARVES is a local control layer for long-running AI project work.

It gives large projects four things that plain chat does not keep stable:

  • Project progression: a visible path from goal to card, plan, evidence, and review.
  • Project memory: durable context about goals, boundaries, decisions, evidence, and handoff notes.
  • Evidence: reviewable proof before accepting a result.
  • Human control: accept, reject, and replan stay explicit.
  1. Guard checks AI-generated patch boundaries.
  2. Handoff records session continuity for the next human or agent.
  3. Audit discovers local Guard and Handoff evidence.
  4. Shield evaluates shield-evidence.v0 into local G/H/A, Lite score, and badge output.
  5. Matrix proves the local composition path and records a summary-only bundle.
  • CARVES is useful because it keeps AI work moving through a stable project structure.
  • Long project memory means continuity of engineering context, not a magic memory database.
  • Matrix is the composition proof layer, not a fifth governance engine.
  • Shield output is a local self-check, not certification.
  • Runtime owns product truth and compatibility wrappers.
  • Site projects the public story and does not own product truth.