clm-0098
measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0098/ — this address never moves; the anchor /records/#clm-0098 keeps resolving
HO-004 measured a negative/speculation-path diagnostic for Qwen3.8-27B Q8_0 on the reviewed v0.6.5 portable Vulkan runtime: stock MTP n_max=3 and DFlash2 Q8_0 width 4 both showed log-backed draft activation, but plain, stock-MTP, and DFlash2 arms all failed Stage A on code/toolish sentinel-empty-content checks, so no guard, Stage B performance, throughput, or energy-efficiency cell is admitted.
verified 2026-08-20 · volatility medium
Note — the record's own working
Scope boundary: negative HO-004 Stage A evidence from /home/aihydra/bench-results/ho004-qwen38-dflash2-matched-r1 only. This does not change the production recommendation, does not establish full capability, does not claim DFlash2 quality equivalence or throughput improvement, does not inherit community DFlash2 numbers, and does not create any stock MTP n_max>=4 safety claim. Math/creative/reason responses existed in Stage A but are preserved only as non-promoted evidence because code/toolish sentinel failures tripped the stop gate. Home Assistant wall-meter energy was joined for the three Stage A windows as Wh total only, not Wh/correct and not throughput-energy efficiency.
Cited by — computed at build time, never stored
model pages qwen38-27b
candidate gate history qwen38-27b