mirror of
https://github.com/hanzo-apps/hanzo.ai.git
synced 2026-10-30 14:58:08 +00:00
zeekay
d01992490a
/enso leads with three lines built from the enso.* runs: the router at 92.9% on GPQA Diamond, level with the pool's best single model, at $2.20, 5x less than Opus 4.8 (86.4%, $10.97); LiveCodeBench 93.7% at $3.06, 12.9x less than Opus 4.8 (72.0%, $39.43); and the pool's ceiling, 98.0% (194/198) with four arms, said only as the ceiling the router is trained toward. Under them a note says what that router was: enso-bench's at the tag its run names, and the arms its notes count, not today's presets. Every percentage and multiple is computed from the runs. A board is each system's frontier run, else its v1 run (the same prompts; v2 and max put others), over the whole set, none degraded. The GPQA and LiveCodeBench boards draw on /enso, /models/enso and /benchmarks, the GPQA one with the ceiling under it on /benchmarks/gpqa. /benchmarks' Enso title, figure, hero tile and frontier (now accuracy against what the run cost), the /models family card and the home tile read the same runs, and each says the API is unreachable rather than show an older number. lib/data/leaderboard.json and lib/leaderboard.ts are deleted: their Enso tiers (98.0, 96.0, 92.9) were no run's, 98.0 being the pool's oracle. With them go the vendor-reported field, the accuracy-at-cost scatter, the reported-versus-measured browser, the shelf's Enso entries and the two pages they generated, and the logo files only the scatter read. A preset card keeps its name and rate and states no score. The gates serve invented runs and hold every surface, the ceiling's figure and the unreachable state to lib/enso.
