Anyone can list a thousand RL environments. Which ones still separate models?

An environment is only worth training on while it pulls the best and worst models far apart. Once every frontier model clears it — or none can — it stops teaching anything.

So we score every RL / agent environment with public model results by discriminative power — the spread it produces across models, penalised as the leader hits the ceiling. 12 scored; 9 still discriminate, 3 are saturated or floored.

Every scored environment, by discriminative power.

Spread = how far apart the best and worst model land. Saturated environments (leader ≥ 90%) and floored ones (no model clears 15%) are flagged — high or low, but no separation left.

Copper rows still discriminate. Faded rows have hit a ceiling or a floor.

Which environments give you the same signal twice?

The public leaderboards share models. So we can ask whether two environments rank those models the same way. High correlation means running both is redundant; low or negative means each catches what the other misses.

Most redundant: GBA Eval ≈ FrontierSWE (r = 0.94, 6 shared models) — pick one.

Most complementary: OSWorld-Verified vs SkillsBench (r = 0.27) — they measure different things.

Your model lags on a capability. Which environment separates it?

Grouped by what each environment trains. For each capability we surface the environment with the most discriminative power left — the one most likely to move a model that’s already strong.

Pick the gap, then jump to the sharpest environment for it.

We also ship verifiable-reward environments.

Indexing the ecosystem is the map. The product is the environment. audio-verify is a working RLVR environment where the reward is objectively verifiable — no learned judge: synth speech → whisper.cpp ASR → structured-field reward.

It demonstrably discriminates: at a fast 320 wpm speaking rate, structured-field recovery collapses from 1.00 to 0.42 while plain WER barely moves — exactly the signal a frontier lab needs and a generic transcript score misses.

20 more, not yet scorable.

Public environments whose scores aren’t machine-retrievable yet, ones with no public scores, plus datasets, tooling, and infra that aren’t capability leaderboards. Ranked above the moment they publish.

An environment's worth is how far apart it puts models. We take the best-minus-worst score across every model with a public result. Wide spread = the environment is still sorting models, so training on it can still move yours.

Once the top model clears 90%, the environment stops separating frontier models — they all pass. When no model clears 15%, it only ranks degrees of failure. We flag both; discriminative power falls toward the ceiling regardless of fame.

This is the analysis CodeSOTA runs on every metric it publishes — including showing WER predicts human TTS preference at only ρ=0.13. An environment nobody can fail, or that everybody fails, is not worth a training run.

Need an environment that still separates models?

If the public environments for your capability are saturated, you can’t tell your models apart and you can’t train past them. We build private, contamination-resistant, verifiable-reward environments and evals on a hold-out set — designed to discriminate where the public ones no longer do.