⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.
clm-0094
communitymed ●●○
citable URL: https://halobench.com/records/clm-0094/ — this address never moves; the anchor /records/#clm-0094 keeps resolving
The llama.cpp #25618 F16-V result is a prompt-sensitive exactness result, not a general proof that MTP preserves the target trajectory on Qwen3.8-27B. The independent Strix Halo Vulkan report reproduced the original prose prompt as a byte-exact PASS, but the same target/draft artifacts, commit, f16 K/V cache, server flags, apply-template-to-completion path and greedy reporter sampling produced 0/5 exact-token parity on five other prompts; replaying those five prompts under a separate fixed sampler also produced 0/5 with unchanged first mismatch locations. The bounded conclusion is that one exact MTP trajectory can be real evidence for that trajectory, while varied prompts remain required before claiming baseline/MTP invariance for the runtime.
Source is a GitHub issue comment, not a local HaloBench run. The artifacts and hashes are recorded in con-0004's upstream report rather than as content/runs because no aihydra run record exists for this diagnostic. Do not promote its timing fields as performance metrics: the report itself states the five-prompt timing totals were diagnostic only because token totals differed. Protocol impact: any future MTP-losslessness or exact-invariance dossier needs a varied prompt set and must preserve the failing prompts or raw token arrays; an exact single-prompt control can show a narrow PASS but cannot clear the broader invariance claim.