BenchAtlas

Sources / scale-labs

Scale Labs

healthydefault: independent

labs.scale.com/leaderboard · trust rank 70 · adapter scale-labs

Observations
257
append-only score facts
Raw snapshots
13
immutable fetched payloads
Last success
2026-07-14 00:16Z
end of last succeeded/partial run
Freshness
1.2 h
hours since last success

License & attribution

scale.com terms grant an internal-purposes license and robots.txt allows the leaderboard, but republication is a legal gray zone — a wave-2 source, attributed and linked, to be reassessed before public launch.

What we ingest

Embedded payloads from labs.scale.com leaderboard pages: score plus 95% CI half-width per model configuration across their SEAL-lineage benchmarks (HLE, MultiChallenge, Fortress, …). Scale runs the evals, so rows land as independent.

Ingestion runs

last 3 of the run history
StartedDurationStatus
2026-07-14 00:16Z4.3 ssucceeded
2026-07-14 00:09Z4.6 ssucceeded
2026-07-13 20:03Z4.5 spartial

partial = run finished but some rows are blocked on identity review — nothing is published until every referenced model/agent/benchmark is resolved.

Raw snapshots

latest 5
FetchedURLSHA-256
2026-07-14 00:16Zhttps://labs.scale.com/leaderboard/visual_language_understandingcc85fb1245ea
2026-07-14 00:16Zhttps://labs.scale.com/leaderboard/multichallenge727c5453bd31
2026-07-14 00:16Zhttps://labs.scale.com/leaderboard/humanitys_last_exam397f2477b7a8
2026-07-14 00:09Zhttps://labs.scale.com/leaderboard/visual_language_understandingb47973c7f555
2026-07-14 00:09Zhttps://labs.scale.com/leaderboard/mask9c1d5f2ab64a

Snapshots are immutable raw payloads — every published number traces back to one. Re-fetching an identical payload bumps the re-fetch time instead of storing a duplicate. Scores from this source appear on benchmark profiles and model pages with this source linked.