[88plug]

H7x (H7 run to its full protocol on the Xeon 6741P, pre-registration #196, amendment V1 #235) ran once on crescent on 2026-11-01 and reads REFUTED.

What ran

• Deploy: ~/k3dgpu-xhost/k3d-gpu-h7x moved from 986887b (the PR ledger(H7x): pre-register H7 to its full protocol on the Xeon 6741P #196 branch) to main at 9978945 (0.1.85). ~/k3dgpu-xhost/eat-src moved from 0d3438d to EAT release-1.1.0 (0c726f57), as V1's Reproduce says; its chart passed h7x.sh's sha256 check.
• Tools: k3d v5.9.0, kubectl v1.37.2 (cross-host amendment K1; bench/xhost/tools.sh had already run at 13:29Z), helm v4.1.1, node image cryptoandcoffee/k3d-gpu:0.1.61 (sha256:8cd264aa...).
• Render pre-check (H7X_RENDER_ONLY=1, no cluster): against smoke 6, only the vLLM image and the pod's fsGroup: 1001 changed, the limit is the fit's 22Gi, as V1 describes. OVMS keeps the pre-registered digest; main's agent manifest now also names its 2026.4.1-gpu tag.
• Setup smoke 7 (V1 item 5), 14:31-14:36Z: B Ready in 115 s, C in 126 s, both on CPUs 0-7,48-55, no request sent.
• The hold, 14:37:24-15:22:13Z, under bench/xhost/hold.sh on the shared bench lock. It started at once (load1 3.16, no aiperf process, disk 32%) and took 2689 s of its 10800 s cap. It was the only H7x hold: one hold of a cap of one.

Result

All 11 blocks (A B B A A B B A A B, then C) are valid. No pod restarted. All 198 scored requests were ok with their full output.

	A: OVMS 2026.4.1, fp16 IR	B: EAT 1.1.0 stock vLLM CPU v0.28.0, bf16	C: B without --enforce-eager (1 block)
decode tok/s, median	12.964	10.607	11.990
prefill TTFT on ~1870 tokens, s	3.005	2.078	1.816
conc4 aggregate tok/s	48.47	39.16	46.23
decode incremental J/token	3.61	4.17	3.83

• R = 1.2222 (gap 0.222), and R_bw = 1.2222.
• The five block-pair ratios are 1.2225, 1.2139, 1.2222, 1.2480 and 1.2188. Every one is above 1 with a gap of at least 0.15.
• bench/apb/h7x.py reads REFUTED. Re-scoring the copied records here gives byte-identical output.
• Meaning: at matched 16-bit weights, OVMS still decodes 22% faster than the vLLM CPU engine EAT ships now, so OVMS's decode lead is not only weight precision. The prediction was R 1.10 (0.85-1.40). The measured R is inside that range but past the 15% line.
• Reported, not gating:
  • C/B is 1.130, inside the predicted 1.0-1.3.
  • vLLM/OVMS prefill TTFT is 0.69, just below the predicted 0.7-4.

Changes

• EXPERIMENTS.md:
  • The H7x row now has its verdict and summary.
  • A falsification-log line is added.
  • A new section, "H7x run (Xeon 6741P, 2026-11-01)", has the Reproduce header, the per-block table, the decision-rule reading and the disclosures.
  • No prediction, rule or amendment text is changed.
• docs/intel/technical-report.md: REFUTED goes to 19 and PENDING to 11 (rebased onto H60's ledger(H60): SUPPORTED on the Xeon 6741P; Flash-Next reads 0.82 of the Gemma control's DRAM share #285). H7x is added to the REFUTED table, to the list of rows that changed no default, and to the list of Xeon-host rows.
• Records:
  • bench/xhost/results/2026-11-01/H7x/hold/: every file of the run, without the kubeconfig. .gitignore drops *.log, so hold.log, chain.log and the run directory's up.log and energy.log are force-added; each is byte-identical to crescent's copy.
  • bench/xhost/results/2026-11-01/H7x/render-precheck/
  • bench/xhost/results/2026-11-01/H7x/campaign.log (force-added the same way)
  • bench/xhost/results/2026-10-21/H7x-smokes/smoke7/, next to smokes 1-6 as V1 asks, with its hold.log, chain.log and up.log force-added the same way.

Disclosures

• The run kept no pod listing. Smokes 2-5 showed that the node image's baked intel-gpu-plugin ran on the older launcher, and this run cannot say whether it ran this time. No arm requests a GPU. This run's launcher printed "No accelerators requested".
• vLLM's OpenMP thread binding was not recorded, because engine logs are kept only for an engine that fails to come up.
• The host's serving stack (serving_cpu, the sum over every listed container) used 2.10-2.38 CPUs throughout (median 2.24), of which the two SGLang serving pairs 1.86-2.12 (median 1.99). The idle W of every block lies within 0.6 W (131.1-131.7 W).

Checks

These ran locally after the rebase, and claims_check, claims_check_test, ledger_drift_check, ledger_drift_check_test and pe_premise_check ran again with the same results after the log commit (d4ba6b6):

• tests/claims_check.sh: clean, 27 files.
• tests/claims_check_test.sh: 83 passed.
• python3 tests/ledger_drift_check.py: 42 files agree with 82 rows.
• tests/ledger_drift_check_test.py: OK, 34 tests.
• tests/pe_premise_check.sh: ok.
• bench/apb/test_h7x.py, test_stage.py and test_apb.py: OK.

Crescent afterwards

• No k3d cluster and no k3d network or volume is left. The bench lock and slot locks are free, and no process of this run remains.
• Both serving pairs answered through dynamo-frontend :8100 at 15:23Z. Each returned "OK" while its own sglang:num_requests_total rose: :30000 went from 52 to 53, and :30001 from 46 to 47.