Post-train Qwen/Qwen3.8-27B (Apache-2.0, native VLM). Score is paired
displacement against the current champion on a private holdout. Regressions are
never crowned.
- Submit over HTTP to the Cortex gateway (
/challenge/relearn/v1/submissions). - Pay Lium yourself (
LIUM_API_KEY/X-Lium-Api-Key). - Declare what you trained on in
manifest. An undeclared manifest fails the contamination gate — it does not skip it. - The holdout reaches the eval pod only after your submission digest freezes, and only for the length of that run.
- Official public benchmarks are out of bounds; the blocklist is
eval/src/relearn_eval/decontam.py. - Only the public ids on
GET /challenge/relearn/v1/statusare trainable.
Check GET /challenge/relearn/v1/status before submitting: while can_score
is false, submissions answer 503. eval_backend tells you whether a verdict
came from the pinned eval image (lium) or an operator's offline harness
(sim, CI and local only).
Eval image contract and how your artifact is loaded and scored:
EVAL-IMAGE.md.
Control plane: https://github.com/CortexLM/cortex