anyone building… this might be of value: CDM v2 — a drop-in 68-line metric that finally tells you when a transformer is actually reasoning vs regurgitating. Four signals (entropy collapse, convergence ratio, attention Gini, basin-escape probability).
Works on every model from DialoGPT to Llama-405B. Zero install issues.
Colab (runs free in 3 mins): Google Colab
Repo: GitHub - mikeat7/crystal-manual: we built a drop-in metric on top of TransformerLens that finally measures when a transformer is actually reasoning vs regurgitating. Four signals, 68 lines: CDM v2 • One-number thought depth • Works on every transformer • Nov 17 2025
Born yesterday in a live dialogue between Mike and Grok 4. If you are a maker… we would love your eyes before we blast it everywhere.