Home › Evidence › Records › clm-0095

clm-0095

measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0095/ — this address never moves; the anchor /records/#clm-0095 keeps resolving

HO-005 measured a reviewer-admitted Ornith-1.0-35B UD-Q4_K_XL production-depth performance frontier for three exact guarded fingerprints: Vulkan f16/f16 KV, ROCm f16/f16 KV, and Vulkan q8_0/q8_0 KV. The admitted llama-bench cells are schema-visible as run-0503 through run-0522, with guards run-0500 through run-0502, all preserving the exact artifact/runtime/KV/batch/load-mode boundaries and the reviewer caveats.

Note — the record's own working

Scope boundary: this is guarded performance and energy-ingestion bookkeeping for the exact HO-005 r2 matrix at /home/aihydra/bench-results/ho005-ornith-q4q8-tuning-matrix-r2. It does not claim Q4/Q8 capability equivalence, Q5_K_M behavior, q4_0 KV behavior, unmeasured lossy-KV quality, or a broad production recommendation. Q5_K_M and q4_0/q4_0 KV remain explicitly excluded from this matrix. The r1 failed attempt and r2 intermediate rc=1 wrapper/supervision event are preserved as runner-mechanics deviations, not admitted metrics. Energy remains structured unjoined on these records because this hbrunner profile lacked Home Assistant history credentials at ingestion time; every run carries the exact window(s) needed for later wall-meter Wh-total joins.

Cited by — computed at build time, never stored

candidate gate history ornith-35b