Arena
Fourteen RL environments: eight public data tasks on Prime Intellect Verifiers and six proprietary spatial tasks from Tera.
An environment is an eval you take the gradient of, so a reward defect is inherited by the model rather than printed beside it. Arena has fourteen environments across two simulator families. Eight public Python data tasks package against Prime Intellect's upstream Verifiers v1 API and use its null harness and subprocess runtime: redaction-pressure, canary-trap, fault-localisation, schema-migration, bot-detection, grand-exchange, grand-exchange-live, and drop-table-inference. Six proprietary TypeScript spatial tasks wrap the fixed-step systems in Tera's browser world: office-nav-v1, office-jobs-v1, studio-ops-v1, drive-101-v1, crow-nav-v1, and california-flight-v1. Lumbridge owns the task, world and reward semantics, not another harness. The six spatial environments expose twenty-five versioned scenarios; CI replays all of them at two seeds. Every reward has counterweights in the same episode. Inaction must earn nothing or less, an oracle or scripted policy must finish positively, safety terminals must be reachable, and snapshots and action traces must replay to the same checksum. The data environments remain Apache-2.0 at git.karti.ai/lumbridge-public/arena; the spatial environments ship through Tera's proprietary distribution. The historical score archive covers seven of the data environments across two model runs; three of those fourteen scores are void and stand as recorded rather than being re-scored, because a number nobody measured is worse than an old one.