Product direction
Track the post-v0.9 verl-user workflow without turning the separate external-alignment study (#39) into the product roadmap.
Priorities, with no promised dates:
- policy-gradient OPD conformance for one documented profile;
- same-base multi-teacher adapter materialization;
- an optional authenticated remote top-k teacher protocol;
- a second fully measured model-family recipe;
- a community hardware matrix with reproducible system records;
- a stable compatibility-profile and plugin API.
Boundary
This roadmap does not imply PPO, GRPO, Ray, FSDP, Megatron or distributed execution inside miniVERL. Each capability needs a focused design, fail-closed compatibility contract and separately verified evidence before it can become public support.
Product direction
Track the post-v0.9 verl-user workflow without turning the separate external-alignment study (#39) into the product roadmap.
Priorities, with no promised dates:
Boundary
This roadmap does not imply PPO, GRPO, Ray, FSDP, Megatron or distributed execution inside miniVERL. Each capability needs a focused design, fail-closed compatibility contract and separately verified evidence before it can become public support.