Skip to content

The maturity ladder

The big picture: There are four maturity levels, from a throwaway prototype to a harness you measure like any other system. Each one has an entry condition, a mechanism, and an exit symptom, so you can tell where you are and what, if anything, should change.

Why it matters: “Start simple, add process as you grow” is not actionable on its own. Without entry and exit criteria you cannot tell whether you are under-built or over-built.

Yes, but: the levels are not a progress bar. Staying at Level 0 or 1 is correct for a project that has not been hurt yet.

Level You are here when Mechanism: Claude Code Mechanism: Blueprint Leave when
0 · Prototype The goal is “does it work at all”; one person; no users Plan mode and the “interview me, write SPEC.md, start a fresh session” pattern (best practices); /clear between tasks (best practices); checkpoints and /rewind. A CLAUDE.md of 30 lines or fewer with build and test commands only None needed. You do not need Blueprint here Something exists you would be sad to lose, or a second session touches the same files
1 · Controlled Changes must be traceable and recoverable A CLAUDE.md of 200 lines or fewer; a permission deny list for force-push, merge, and secrets, and hooks for rules a deny list cannot express Issue with acceptance criteria, then branch or worktree (scripts/wt), then MR on a green pipeline (/mr); tests in the same commit; changelog fragments; the fast-paths table in CLAUDE.md; the .claude/settings.json deny list and hooks A defect ships that a rule already forbade, or review by the author stops finding things
2 · Hardened Correctness claims need evidence someone else can check Hooks and /goal to define “done”; sub-agents and the bundled /code-review for fresh-context review. Available locally; Blueprint wires /code-review as a trial on the Pure bugfix row (see CLAUDE.md, #103) A fresh-context reviewer (completeness-check); gate self-tests that run in the same job as the gate; stack-conditional schema fuzzing; an off-by-default complexity check (ci/sonar-complexity.yml), plus an off-by-default Sonar rule-key check (ci/sonar-exclusions.yml), with no static-analysis scan job shipped; the gate ledger You cannot say whether the last harness change helped or only cost tokens
3 · Measured Harness changes are themselves experiments claude plugin eval suites with a pinned model and a no-plugin baseline A behavioral eval suite for the agents and skills, with a no-plugin baseline and a weekly scheduled run, ships in 0.8 (#51). Planned, not shipped: escape rate (#85), plugin packaging (#100). The ledger’s per-gate yield and the /kaizen audit are the measurement input Level 3 builds on; they do not make Blueprint Level 3. Promote a rule to Blueprint only with yield evidence from at least two projects; demote 0%-yield gates Steady state: run the audit on a cadence

Nightly mutation testing, a domain-invariant suite, release evidence by category, and an external scanner with a new-code quality gate are also Level 2 mechanisms. Blueprint ships a reference mutation gate (off by default); a project supplies the rest itself.

Level 3 also means tracking cost per merged MR next to escape rate. Neither is measured by Blueprint yet; #85 is the issue that adds both.

  1. Entry is by symptom, not ambition. You move up because something hurt, not because the next level sounds more professional. This matches Blueprint’s own origin: its rules were written after the incidents that made them necessary.
  2. Every level must be removable without silently weakening the level below. scripts/check-gate-selftest-parity.sh already enforces the related property for gates: every CI gate must prove it can still fail in its own job, or record why it cannot. Rules files do not have the same property, so deleting one is a manual audit.

As a template, Blueprint is Level 2. It ships the Level 1 workflow and some Level 2 evidence mechanisms: a fresh-context reviewer, gate self-tests, the gate ledger, and stack-conditional fuzzing. A static-analysis gate you add yourself. These ship as templates and CI includes that you enable yourself; /kickoff asks which add-ins you want and recommends none (fuzzing, for one, stays inert until ci/python.yml is included and you wire the API-boot hook).

As a harness for harnesses, it has no Level 3 mechanism yet. Escape rate, an eval suite for the harness itself, and plugin packaging are open issues, linked in the table. Until they land, Blueprint does not claim Level 3 for itself.