Get a GRPO run
One run: status, ledger-true spend across all three meters (reward, GPU, environment), and the outcome card for terminal runs.
Authorizations
Your workspace API key, e.g. sk_sovereign_..., sent as Authorization: Bearer <key>.
Path Parameters
Response
The run.
FAILED means the trainer gave up; STOPPED is an external termination (user stop, wall clock, dead box); OVERBUDGET means the reward budget's kill switch fired.
ACTIVE, COMPLETED, OVERBUDGET, STOPPED, FAILED Hard ceiling on judge/reward spend for this run.
Ledger-true reward spend so far.
Platform GPU-hour spend, billed under its own prefix; never counted against the reward budget.
Environment/sandbox execution spend, also outside the reward budget.
Frozen marked-up GPU rate (rate x count): the live burn rate while ACTIVE. Null for BYO-GPU runs.
Terminal runs, when the training box produced its artifact: {steps, first_half_mean_reward, second_half_mean_reward, stopped_by_tripwire, bakeoff?: {verdict, delta, delta_ci95, prompts, k, mean_sim_fraction, mean_tool_steps}}. Null while ACTIVE or when no artifact exists.