Development Roadmap
Here's what we're currently working on and what we plan to build next.
In Progress
Planned
PlannedJudge
Call the Meta Model API tier in a real run
The escalation path to Muse Spark is written and tested with faked models, but it has never called the real API. Run it on real unsure asks and record what happens.
PlannedLearning
Train the first LoRA on Glimmer
train/train_lora.py and the dataset split exist, but no LoRA has been trained. It needs enough labelled decisions from cards written by Meta models.
PlannedEvaluation
Label an evaluation set from real nights
So far the judge has only been measured on example cards and three probe asks. Label real nights so missed human decisions can be counted.