Home › Evidence › Records › clm-0092

clm-0092

measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0092/ — this address never moves; the anchor /records/#clm-0092 keeps resolving

Ornith-1.0-35B UD-Q4_K_XL has a bounded HO-005 r3 preflight record on the stock 3653e6d6d Vulkan/f16-KV boundary: the Q4 artifact identity was sidecar/local-sha matched, FIT/header/load at c32768 was healthy, the schema-visible guard passed 4/4 (run-0494), and the 5-task tau2 airline smoke passed 5/5 with mean_reward 1.000 (run-0495). The same admitted r3 session also completed one planned fresh-process d262144 throughput cell: 86.1589 t/s prefill (run-0496) and 24.3129 t/s decode (run-0497), with no reviewed DeviceLost/OOM marker. This is only a bounded preflight/smoke plus one model-max throughput cell: it is not a full-26 Q4 capability claim, not a Q4/Q8 equivalence claim, and not an inheritance of Q8_0 capability across a lossy weight-quant change.

verified 2026-08-20 · volatility medium
evidence run-0494 run-0495 run-0496 run-0497

Note — the record's own working

METHOD — HO-005 reviewer-admitted r3 evidence from /home/aihydra/bench-results/ho005-ornith-q4kxl-vulkan-preflight-r3 only. Artifact: /home/aihydra/models/ornith-35b/Ornith-1.0-35B-UD-Q4_K_XL.gguf, 22,324,804,000 bytes, sha256 67081ae4a1a291bd6c72834094ea056332cb3cb5fa15e88536ec7f233a475b71, source unsloth/Ornith-1.0-35B-GGUF per the candidate record. Build/backend: /home/aihydra/src/llama.cpp-vk3653e6d HEAD 3653e6d6d547ec763317d9ecd0ace334a7e21359, build_commit 3653e6d6d, Vulkan0 Radeon 8060S, -ngl 999 -fa 1 -b 2048 -ub 512 --load-mode none, f16 KV, --parallel 1/-np 1; tau2 head 668d3bcd135c02aa3438f987ef45735b7c163ee3. FIT proof: c32768 healthy 2026-08-20T00:17:59Z..00:18:15Z, rc=0, gtt_used_bytes 22444122112. Guard: 2026-08-20T00:18:15Z..00:18:23Z, rc=0, 4/4, native tool call non-empty, isolation skipped only because --parallel 1. Smoke: 2026-08-20T00:18:23Z..00:26:35Z, task ids 0..4, tool_call_messages 23, empty_argument_tool_calls 0, empty_assistant_turns 0, infrastructure_errors 0, rc=0. Non-fatal tau2 log cost-map ERROR lines were present and are not treated as infrastructure errors. Throughput: 2026-08-20T00:26:37Z..00:49:37Z, pp1024/tg256 N=1, parse ok, stddev/sample vectors preserved on run-0496 and run-0497. Failed r1/r2 custody/wrapper attempts are not ingested as metrics. ENERGY STATUS — wall-meter energy is now joined from Home Assistant counter-differences over the preserved run-meta windows: guard eng-0199 (0.1788 Wh), 5-task smoke eng-0200 (16.0646 Wh; 3.2129 Wh/correct for this bounded smoke only), and d262144 throughput eng-0201 (61.9776 Wh gross). These joins do not broaden the claim into full-26 Q4 capability or Q4/Q8 equivalence.

Cited by — computed at build time, never stored

model pages ornith-35b
candidate gate history ornith-35b