Skip to content

Latest commit

 

History

History
10 lines (7 loc) · 468 Bytes

File metadata and controls

10 lines (7 loc) · 468 Bytes

Model Ladder

This repository assumes a three-model ladder:

  • Flagship: highest quality candidate used for best cognition performance.
  • Fallback: lower-cost model intended to preserve doctrine and bounded behavior.
  • Eval: fast experimental model used for rapid benchmark iteration.

Promotion rule

No model becomes canonical unless it improves reasoning and repair without unacceptable regression in calibration, compression retention, or prompt-independence.