- BPB (bits-per-byte) metric alongside NLL for tokenizer-independent comparison
- Modal inference integration
- Audio recipes (LibriSpeech ASR, EkaCare medical ASR) + tml_v0 renderer
- Inkling reasoning-effort controls and model-keyed rollout defaults
- RL rollout: configurable runner, parse-error policy, termination reward policy
- JWKS-based token exchange (get_access_token.py)