[88plug]

H54q is INCONCLUSIVE, the verdict its pre-registration rated most likely (0.65).

On crescent's Xeon 6741P, over GSM8K items 0-199 with Gemma 4 26B-A4B:

• The QAT-matched OpenVINO IR (OVQ: INT4_SYM, group 32, built from Google's QAT bf16 weights) stayed inside the 5-point non-inferiority margin of the bf16 QAT reference (R).
• OpenVINO's published base-model IR (OVB) did not score below OVQ. The second half of the claim therefore does not hold, and no REFUTED bound is crossed.

Result (h54.py score-q; paired bootstrap over items, 10,000 resamples, seed 1)

Arm	Correct of 200
R: vLLM CPU v0.32.0, QAT bf16	194 (0.970)
OVQ: OVMS 2026.4.1	191 (0.955)
OVB: OVMS 2026.4.1	192 (0.960)
Lg: llama.cpp b11749, QAT Q4_0 GGUF	193 (0.965)
Vi: vLLM v0.32.0, Intel AutoRound	not deliverable

• OVQ - R: -0.015, 95% CI -0.035 to 0.000. The lower bound clears -0.05, so the SUPPORTED half holds.
• OVB - OVQ: +0.005, CI -0.015 to +0.025. It is neither below 0 nor above 0, so the SUPPORTED half fails and the REFUTED bound is not met.
• Not gated:
  • Lg - R: -0.005 (CI -0.025 to +0.015).
  • OVB - R: -0.010 (CI -0.035 to +0.015).
• Predictions:
  • R in [0.85, 0.96]: missed. R read 0.970, above the range.
  • OVQ - R, Lg - R and OVB - R: each inside its range.
  • "OVB < OVQ resolves": did not resolve, as the 0.8 side predicted.
• Every build scored 95.5-97%. Only 5 items differ between OVB and OVQ, too few to separate them, so the falsification log closes this item set for these builds.

What ran

Amendment Q2, the runner. Q1 required a runner amendment before the first hold. Q2 was pushed at 19:34:33Z (11e076d), before any smoke or hold. As with H50's E3 and H51's S1, it was not merged when the holds ran; the run section discloses this.

• arm.sh gained an item mode. Each arm answers its GSM8K items 16 in flight, with H54a's chat and grader unchanged. Every finished request is written as it lands, and no request starts after an item stop 90 s before the hold's deadline.
• An item's first error re-asks it once; a second error counts as wrong.
• A server that stops answering fails its arm, and that run of errors is discarded.
• OVQ, OVB and Lg apply the exact manifests H54a's scored runs applied. R and Vi derive from H54a's recorded B1g render. R's manifest is byte-identical to the Vq manifest H54a's second smoke booted.
• h54.py score-q assembles items 0-199: H54a's records for items 0-49 of OVQ, OVB and Lg, and this run's records for the rest.

Before the holds, at 85e7726 (the runner plus qcheck.sh):

• Every weight file's sha256 matched the pre-registration's list, read under the bench lock.
• Both serving pairs answered through Dynamo :8100, each with its request counter rising.
• A discarded wiring smoke ran: R, OVQ and Lg were ready and answered two non-GSM8K questions. No GSM8K item was asked.

The holds (qcampaign.sh), 3 of the 3 Q1 allows:

Hold	Time (UTC)	What it did
1	19:46-20:20	R answered items 0-199 in 542 s. OVQ answered items 50-199 in 1,253 s.
2	20:22-20:59	OVB answered items 50-199 in 1,098 s. Lg answered items 50-199 in 672 s.
3	20:59-21:15	Vi only, not deliverable. vLLM v0.32.0 stops on the AutoRound checkpoint with "has no parameter 'w2_qweight'", as in H54a's smoke.

• No item errored or was cut, and no server stopped answering.
• aiperf was idle in all 521 sampler windows.

Crescent after the run

• No k3d cluster is left, no lock is held, and no process of this run is running.
• Both serving pairs answered a chat through the Dynamo :8100 front door, and each pair's num_requests_total rose:
  • 19:37:01Z, before: :30000 157 to 158, :30001 52 to 53.
  • 21:15:57Z, after: :30000 158 to 159, :30001 54 to 55.
• H54q used no GPU and stopped no container.

Changes

• bench/h54/:
  • h54.py: the q-* commands and score-q.
  • arm.sh: the item mode.
  • hold.sh: row q and the smoke row smokeq.
  • reproduce.sh: the H54q manifests' sha256.
  • qcampaign.sh and qcheck.sh, both new.
  • q-smoke.jsonl: the smoke's two non-GSM8K questions.
  • test_h54.py: H54q tests, already in CI.
• bench/h54/results/2026-10-29/H54q/: every record from the three holds, the smoke and the checks.
• EXPERIMENTS.md:
  • amendment Q2;
  • the H54q row's verdict and result;
  • a falsification-log entry;
  • the "H54q run" section, with its Reproduce header.
• docs/intel/technical-report.md: the counts now read 29 INCONCLUSIVE and 18 PENDING (after H52b, merged from main), and there is an H54q INCONCLUSIVE row.
• docs/intel.md: the H54q row is updated.
• README.md: the QAT-conversion paragraph now says H54q's run did not promote the conversion. The optimizer still promotes it only on a SUPPORTED H54q.

Checks run locally

• bench/h54/test_h54.py: OK, 54 tests. Repeated runs were stable.
• bench/h54/diag/test_b1g_diag.py: OK.
• bench/h54gpu/test_h54gpu.py: OK.
• bench/eatflow/crescent/test_holds.py: OK.
• bench/eatflow/crescent/test_crescent.py: OK.
• tests/bench_test.sh: 71 passed, 0 failed.
• tests/claims_check.sh: 26 files clean. tests/claims_check_test.sh: 81 passed.
• python3 tests/ledger_drift_check.py: 41 files agree with 81 ledger rows. tests/ledger_drift_check_test.py: OK.
• tests/pe_premise_check.sh: ok. tests/pe_premise_check_test.sh: 16 passed.
• tests/model_cards_license_test.py: OK.
• tests/ci_wiring_test.sh: 69 passed.
• tests/docs_test.sh with mkdocs 1.6.1: 63 passed.
• console/ledger-rules.json regenerates byte-identical.
• shellcheck v0.11.0 at warning level is clean on bench/h54/*.sh.
• The scorer, run on the committed records, reproduces crescent's score-q.json byte for byte.