This package is not treated as a chat toy. It is treated as a controlled cognition program.
- No silent success claims.
- No promotion without measurable improvement.
- No frontier-class language without benchmark evidence.
- No quantized deployment without retention checks.
- No local-backend claim without a reproducible runbook.
- Base-model lane: select and baseline candidate brains.
- Cognition-shaping lane: improve planning, critique, repair, compression, and calibration.
- Evaluation lane: run benchmark suites, paraphrase tests, and retention checks.
- Promotion lane: compare against incumbent and reject regressions.
- Deployment lane: export only promoted candidates to GGUF and verify local-serving behavior.
- benchmark run artifact
- scored benchmark artifact
- promotion decision artifact
- quantization retention artifact
- experiment record
This repository is an execution-first alpha lab. It is not a frontier cognition engine yet.