CRMA claims near-zero LLM forgetting

This is an intriguing continual-learning result, but right now it reads more like an unverified benchmark claim than a finished research release.

• –The headline number is strong: -0.1% average drift versus +351% forgetting for a naive baseline on four sequential domains
• –If the result holds without replay, EWC, or knowledge distillation, it would be highly relevant for long-running domain adaptation and fine-tuning workflows
• –The lack of a paper, repo, or reproducible benchmark package makes independent verification the real story here, not the claimed win itself
• –The repeated Reddit and Hugging Face forum posts suggest early community seeding, but not yet a mature launch or broadly recognized method