clm-0127
The KingJones Qwen3.8-Flash-Next full-STRIX ROCmFP4 series is admitted only as BOUNDED/PARTIAL on AI Hydra (cfg-0180/cfg-0181). Exact identity is artifact c6770d7442a06bf1d78edf28cec83e1ec93afdd34664c23ff898807b6b9349fa, runtime 36e9acd40e10a87cd3c3ef8ec734668757dc8520, Ubuntu 26.04, gfx1151 and HIP 7.1.52801-9999. The corrected one-slot served cohort used c262144, f16 K/V, b2048/ub512, thinking off, 512 context checkpoints and a 2048-token checkpoint cadence under the disclosed runtime-default-mmap exception. Standard Tau2 Airline smoke was 5/5. The full seed-42 N=1 arm attempted all 26 tasks: 24 were evaluated, 19 passed and five scored zero; tasks 23 and 24 were infrastructure errors with no score. Therefore 19/24 is the evaluated result and 19/26 is all-attempt pass coverage, not a zero-imputed suite score. The 24 evaluated trajectories carried 245 tool calls, zero empty-argument calls, zero empty assistant turns without a tool, and two tool outputs marked error inside otherwise scored-pass trajectories. The 8,192-token ceiling was inherited from clm-0119's different five-task Qwen3.8-27B IU4/Kairic Edge matrix, where 6k and 8k both scored 4/5 and 8k was selected as headroom; it was not tail-tested for this full-26 STRIX arm. An original-log audit found exactly two final calls at that ceiling, tasks 23 and 24. A later two-task 16k recovery is held because its retained transport was not lossless, so no mixed-budget aggregate amends this strict 8k result. A separate post-hoc wall-meter join (eng-0276) recovers 419.1478 Wh total and 20.5395 active Wh per 19 correct answers from retained exact UTC edges. It is not campaign-time metering compliance: the dated 10.1 W idle baseline and whole-box attribution remain explicit. Corrected cache_prompt:false prompt processing is admitted at exact server prompt_n 8192, 32768, 65536, 131072 and 260096 (run-0640..run-0644). Generation rates are admitted only for nonzero-output 32768, 65536 and 131072 rows. The zero-byte 8192/260096 synthetic predicted rates are excluded. Full-window d262144 returned HTTP 400 without timing or prompt_n (run-0645). Controlled 32k and 131k prefix-reuse traces are separate records (run-0646/run-0647), not ladder repetitions. Representative whole-file mincore, GTT, MemAvailable, page-cache and process receipts remain bounded observations on the cited runs. Per-PLE mincore is incomplete: 30 of 40 before/after records failed after five of six layout tensors, so no PLE fault-tracking, complete PLE working-set or resident-page claim is made. Raw mmap engine rows remain visible (run-0648..run-0652), but d0/d32768 prompt-processing CVs of 3.731412%/3.131244% exceed the protocol trigger; no unrestricted aggregate floor, best-config or cross-model claim is admitted. Thermal observations reached 98 C but never the confirmed >=100 C stop. The exact owner released cleanly with no lease or model process. Publisher comparison is attribution-bounded by con-0017: that immutable page used ROCm 7.2.4, different prompt depths/context allocation and checkpoint evidence, and STRIX_LEAN for its c262 rows. Similar 32k-131k magnitudes are a caveat, not equivalence or causal attribution to a runtime version.