Evidence level:
lab_repeated. Every number is re-derived from raw per-request records by the query named in the claim registry. Three repeated runs per cell; one variable changed at a time against the same pinned baseline (Qwen3.6-35B Q4_K_M, llama.cpp Vulkan, one GTR9 Pro unit).
Why this study
A verdict for one configuration is not a verdict for its neighbours. This study swaps exactly one identity component at a time — the model family, the quantization, the compute backend — and measures what actually transfers.
Swap 1: a second model family reproduces the failure mode
Our earlier finding was that a reasoning model under a tight completion budget returns HTTP 200 with an empty answer. On gemma-4-26B-A4B in its default mode at a 1024-token budget, 8.3% of requestsrepeated of everyday requests came back empty for the same reason — the reasoning pass consumed the budget. Milder than the Qwen thinking cell, but present: the failure mode is not unique to one model family — both families measured here show it. Whether it holds across reasoning models generally is exactly what this corpus is built to keep testing. The answer trails the stream here too: the first streamed token arrives at 282msrepeated while the first answer token arrives at 10768msrepeated.
Elsewhere gemma is a capable tenant of the same box: needle retrieval at 95.8% of requestsrepeated across the 2k–32k ladder, strict-JSON task success at 93.8% of requestsrepeated (its failures were unparseable RU generation outputs, not wrong values), but on the unanswerable controls it produced an explicit admission of absence in 75.0% of requestsrepeated of requests. The failed quarter did not fabricate codes — those requests returned empty answers, the reasoning pass having consumed the whole token budget. Qwen3.6 answered every control with an explicit decline.
Swap 2: Q8_0 bought nothing these workloads can see
The 35 GB Q8_0 artifact against the 20 GB Q4_K_M, reasoning off: strict-JSON task success stayed at 100.0% of requestsrepeated — identical to Q4_K_M — while decode slowed from 15.9ms/tokenrepeated to 18.7ms/tokenrepeated per token, memory-bandwidth-bound as expected on unified LPDDR5X. On these workloads the larger quant is a strictly worse deal. That is a statement about these gates and corpora — a quality difference may well exist where correctness is harder than our gates measure.
Swap 3: ROCm runs on gfx1151 — but slower
The ROCm build completed every request (100.0% of requestsrepeated on strict JSON, first answer token at 215msrepeated), so the backend is usable. But decode ran at 18.7ms/tokenrepeated per token against Vulkan’s 15.9ms/tokenrepeated — the Vulkan backend is the faster choice on this box today.
The cautionary tale: our own pre-flight smoke test, which read the server-reported timings of a single tiny completion, suggested ROCm was substantially faster. The full client-side cells reversed that verdict. A micro-sample of server telemetry and a user’s stopwatch are different instruments; this lab publishes the stopwatch.
What this does not establish
Disabling gemma’s reasoning was not tested (the switch is template-specific). Both quants sit at the ceiling of our correctness gates — harder tasks could separate them. The ROCm image is a floating upstream tag, pinned here by its digest. Swaps 1–3 ran on one unit; only the baseline configuration has unit-to-unit replication. None of this compares Strix Halo to any other device.
Evidence
Every run shipped the standard bundle (manifest, per-request records including failures, gate outcomes, runtime log, checksums); the numbers above are re-derived from those records in CI. Harness and corpora: agmind-bench.