Roadmap
We don't declare 'done.' It's earned.
A capability moves right only when a live run has reproduced it, the closed loop has captured what it taught, and the numbers behind it are real and sourced. Here's where things honestly stand, with no dates we can't defend.
Proven
reproduced, closed-loop, sourced
- Read-only recall over the knowledge substrates (Claude Code plugin).
- 342 canonical lessons: retrievable, actionable, attributed.
- 109 compiled detect → fix patterns.
- Determinism suite: 994 tests green, byte-compared in CI.
Building
in review / behind the gate
- ~10,700 categorized candidates working through the verify + human-gate pipeline.
- The plan lane: goal sentence → reviewable bring-up plan.
- Co-design cascaded gate hardening (parse → compile → correctness → benchmark).
Planned
not started / earning it
- Public benchmark runs, each with a full provenance stamp.
- The orchestrate lane, generally available with per-step human gates.
- Broader substrate coverage and framework matrices.
Corpus counts are real and update on every promote. Benchmark numbers stay as labeled placeholders until the first reproduced live run.