Home › Evidence › Records › clm-0102
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

clm-0102

measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0102/ — this address never moves; the anchor /records/#clm-0102 keeps resolving

LLaDA2.2-flash (100B-A13B block-diffusion MoE, Q4_K_S) solved 5/5 tau2-bench airline screening tasks (task_ids 0-4), mean reward 1.000, with real multi-turn tool use, running GPU-accelerated on aihydra (Strix Halo, gfx1151) via the headbouyJB/diffuse-cpp fork with the GPU-resident KV cache and full Levenshtein editing enabled. This is a 5-task SCREEN, not the full 26-task bench. Energy-joined (eng-0226): 62.10 Wh total, 12.42 Wh per correct answer.

verified 2026-08-22 · volatility medium
evidence run-0553 eng-0226

Cited by — computed at build time, never stored

model pages llada22-flash
candidate gate history gemma4-26b llada22-flash
docs benchmark-protocol