RTBENCHv0.4
← Back to scoreboard

Avaturn-1

open · MITavaturn-1-0.1.0last tested 2026-05-01T18:00:00Z

endpoint · open-weights, local inference (avaturn-1-0.1.0)

homepage · https://avaturn.live

All six probes vs. field median

Values normalized to [0,1]. IFC normalized as 1 − |v − 1|.

Raw scores

ProbeValueUnit
GR — Generation Rate0.994ratio
XPD — Cross-Run Pixel Diversity0.184embedding-dist
AVC — Audio→Motion Causality (off-mouth)0.710pearson-r
LWA — Listening-Window Activation3.420ratio
EVD — Emotional-Valence Differentiation4.830fisher-ratio
IFC — Idle-Loop Compression0.940ratio-vs-human

Frame strips

Identical-input replays. Hashes computed by the harness.

support-package-lookup

✓ distinct

grief-disclosure

✓ distinct

Composite history

Disclosure

Open weights (MIT). Inference run locally on the bench A100. Composite leader by construction of architecture, not by tuning.