Skip to content

Latest commit

 

History

History
27 lines (22 loc) · 1.1 KB

File metadata and controls

27 lines (22 loc) · 1.1 KB

DARPA-Style Program Control

This package is not treated as a chat toy. It is treated as a controlled cognition program.

Governing principles

  • 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.

Program lanes

  1. Base-model lane: select and baseline candidate brains.
  2. Cognition-shaping lane: improve planning, critique, repair, compression, and calibration.
  3. Evaluation lane: run benchmark suites, paraphrase tests, and retention checks.
  4. Promotion lane: compare against incumbent and reject regressions.
  5. Deployment lane: export only promoted candidates to GGUF and verify local-serving behavior.

Required evidence before saying a candidate is stronger

  • benchmark run artifact
  • scored benchmark artifact
  • promotion decision artifact
  • quantization retention artifact
  • experiment record

Current truth

This repository is an execution-first alpha lab. It is not a frontier cognition engine yet.