Home › Evidence › Records › cfg-0137
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0137

backfilledfork
citable URL: https://halobench.com/records/cfg-0137/ — this address never moves; the anchor /records/#cfg-0137 keeps resolving
modelQwen3.8-27B-Q8_0.gguf · Q8_0 · rev 57484e3196aaff8dbbd71c666158b00206baa2bc134e4f16ae90fc8cbdfbeea2
engineggml-org/llama.cpp 0b0f35d0ed745bd6e40eb248e320cacbb1a546ab · Vulkan0/Strix-Halo · host aihydra (igpu)
flagsHO-004 stock MTP n_max=3 Stage A fingerprint: same target/runtime/backend/KV/batch/load-mode as cfg-0136 plus native draft-mtp path with --spec-type draft-mtp --spec-draft-n-max 3. No DFlash2 drafter, no ngram/proxy/llama-swap/ROCm, no stock MTP n_max>=4, no alternate target quant, no KV quant substitution.
templatenot recorded at test time
treefork — carrying con-0002
aged evidence — reconstructed from the archive. Stage A activation/safety arm only. Positive draft acceptance counters were reviewer-admitted, but the arm failed the Stage A sentinel gate on code/toolish prompts; no guard or Stage B performance cell is admitted and no n_max>=4 safety claim is made.

Capability basis

measured on this config

Runs on this config (1)

rundatekindsuitemetrics
run-05422026-08-20guardstageA-c32768-varied-promptsrunner_invalid=true