jmclark-lab rips out the AI engine and bolts on an audit trail
A permanent, no-going-back switch to a new AI model - paired with a record of exactly what wrote every answer.
jmclark-lab retired the fork's old AI model and moved wholesale to Sakana AI's Fugu, a newer large language model. No toggle, no phased rollout - a clean, permanent swap, routed through a single internal layer so the rest of the system no longer cares which model sits underneath.
Three things landed together:
- A hard model swap. Every answer now comes from Fugu; the old engine is gone entirely.
- A provenance stamp. Each assistant message is tagged with the exact model that produced it - success, cancellation, or error alike - so any past answer can be traced to its source long after the fact.
- A boot-time guardrail. The system now refuses to start without the model's access key, turning a silent misconfiguration into an obvious, immediate failure.
So what Anyone who has to answer "which model wrote this?" months later - GCs, compliance leads, litigation teams - should note the provenance work as much as the engine swap.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?