Each row points at a file. Where the file does not exist, the row says so.

The blog says every number carries an error bar, or says why it has none, and a provenance. This table is the proof — including the proofs we have not attached yet. The ruler corrects downward and also confirms. Both directions are below.

§1 — the ruler works in both directions

Two real cases. The 25 Aug remeasurement is published on this site. One was corrected downward; the other reproduced. The current value comes from resumo.json.

B1 · corrected downward

2026-08-21 · retracted

10.1×

Cherry-pick of the floor of the band. JSONL absent.

2026-08-25 · current; file attached

9.78×

183.6 → 1,795.1 tok/s. 3 per arm.

B2 · reproduced

2026-08-21 · first reading

29,567tok/s

Four replicas on the same node. JSON absent.

2026-08-25 · current; file attached

29,503tok/s

uncertainty 104.2 (sd across repetitions). n=3 (overlap 118 s). Not every remeasurement knocks the number down.

§2 — the index

Solid ink = read from a file published on this site. Light ink = no archived raw file. Struck through = retracted. Sorting and filtering need no JavaScript: anchors and radio buttons. Without CSS, the first table (by date) stays visible. There is also a TSV.

all 33 with a file 21 no raw file 12 with history 22

measurements by date, newest first
id	what was measured	value	uncertainty	n	date	hardware	raw file	history	replication
`C9b`	GLM-5.3-Flash, HLE text only, strict exact match (K=1, no judge) No judge, and without the numeric tolerance that credited different answers: a floor.	8.75 % 127/1,451 strict: 131 by the file's count, 4 only by the numeric tolerance. No judge: a floor, not comparable to the official score	0.74 (binomial standard error; 2·se 1.48 pp)	1,451 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 9.31 % (131/1,407) retracted The denominator dropped 44 items that had not converged even at 163,840 tokens, and the relative tolerance credited 4 answers that differed from the key.	us: 1 run, file attached. outside: nobody
`C9a`	GLM-5.3-Flash, HLE text only, multiple choice (K=1, no judge) Denominator corrected; with no judge, this is not the official score.	35.98 % 163/453; no judge; answers that did not converge count as wrong	2.25 (binomial standard error; 2·se 4.51 pp)	453 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 36.06 % (163/452) retracted The denominator dropped the multiple-choice item that had not converged even at 163,840 tokens.	us: 1 run, file attached. outside: nobody
`C11`	GLM-5.3-Flash, MMLU-Pro with the official extractor (TIGER-AI-Lab) With the official extractor, the same answers give 82.96 %; the runner's old grader gave 72.59 %.	82.96 % 9,981/12,031 with the official MMLU-Pro extractor; the runner's old grader gave 8,733/12,031 = 72.59 %. 122 unfinished answers count as wrong. Also reading \boxed: 83.48 %. Our own 0-shot prompt, without the official harness format instruction	0.34 (binomial standard error; 2·se 0.69 pp)	12,032 items (K=1); 12,031 in the denominator	2026-09-23	8×H100 80GB, TP8; our runner	mmlu-pro-glm53-20260923	2026-09-21 72.59 % retracted Produced by the runner's old grader: it does not match \boxed{\text{G. …}} and it ignores case ('The answer is exactly…' becomes E). The official extractor on the same answers gives 82.96 %.	us: 1 run, file attached. outside: nobody
`C10b`	GLM-5.3-Flash, DeepSWE 111 tasks, v1lite scaffold variant (K=6) The scaffold did not move the result; the recorded value came from rounded percentages.	64.21 % 81/108 · 68/105 · 71/109 · 66/106 · 62/105 · 62/105; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 62.80 % (2·se 5.41 pp); delta against the baseline −1.11 pp (2·se 5.56) 75.0 · 64.8 · 65.1 · 62.3 · 59.0 · 59.0	5.91 (sd across repetitions; 2·se 4.83 pp)	111 tasks × K=6	2026-09-21	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	2026-09-21 64.18 ± 4.85 retracted A mean of rounded percentages; two repetitions were written down as 58.95 % instead of 59.05 % (62/105).	us: 6 repetitions, file attached. outside: nobody
`C8`	GLM-5.3-Flash, MATH-500 (K=3 seeds), corrected answer normalizer The old number measured the normalizer, not the model. The current one reapplies the corrected normalizer to the same answers.	99.00 % 1,485/1,500 with the new normalizer; the old one gave 76.20 %. 148 distinct pairs reclassified (342 items), read by hand; 15 failures left: 7 truncated, 8 wrong 99.4 · 98.2 · 99.4	0.69 (sd across seeds; 2·se 0.80 pp)	500 × 3 seeds	2026-09-19	8×H100 80GB, TP8; our runner	math500-glm53-20260919	2026-09-19 76.20 % retracted The normalizer marked equivalent answers wrong (x = 4 against 4; \dfrac against \frac).	us: 3 seeds, file attached. outside: nobody
`B8c`	GLM-5.3-Flash, one stream with MTP k=3: nvidia-smi counters (memory busy over time) The same step with MTP keeps memory busier and draws more power: more useful work per step.	26.95 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 99.16 % (fraction of time any kernel is running; this is not per-SM activity) · 308.1 W per GPU 26.8 · 26.8 · 27.2	0.20 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions, file attached. outside: nobody
`B8b`	GLM-5.3-Flash, one stream, no speculation: nvidia-smi counters (memory busy over time) The old label «SM active» was wrong: utilization.gpu is the fraction of time any kernel is running.	21.80 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 98.74 % (fraction of time any kernel is running; this is not per-SM activity) · 256.9 W per GPU 21.9 · 21.8 · 21.8	0.06 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	2026-09-17 «SM active» 98.7 % retracted as a label The column is nvidia-smi utilization.gpu: the fraction of time any kernel is running, not per-SM activity. This node has no per-SM activity counter.	us: 3 repetitions, file attached. outside: nobody
`B8`	GLM-5.3-Flash, one stream: native MTP k=3 against no speculation (17 Sep window) Reproduced from the raw file. The current value comes from the per-request repetitions; tok/s includes TTFT. Synthetic filler prompt at temperature 0: speculative acceptance depends on the text; on real agent tasks MTP gives 1.52× per task (B10).	2.04× 156.05 ± 0.07 → 317.90 ± 2.92 tok/s (±2·se); repetitions between 2.02× and 2.05× 156.1 · 317.9	0.02 (sd across repetitions, propagated; 0.06 and 2.53 tok/s on the arms)	3 per arm	2026-09-17	8×H100 80GB, TP8, native FP8, dedicated vLLM image; 1 request at a time, synthetic 256-token filler prompt, greedy decoding (temperature 0), 512-token output cap, thinking off	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions per arm, file attached. outside: nobody
`C10a`	GLM-5.3-Flash, DeepSWE 111 tasks, our agent (K=3) Reproduced. Our agent, not the vendor's. The first repetition mixes two serving setups: 29 tasks at 20 workers without MTP and a 1,800 s timeout, 82 at 8 workers with MTP k=3 and 3,600 s. 111 of the 113 tasks: the gold patch fails our harness on 2.	65.32 % 67/106 · 70/108 · 72/106; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 63.72 % (2·se 2.65 pp) 63.2 · 64.8 · 67.9	2.40 (sd across repetitions; 2·se 2.77 pp)	111 tasks × K=3	2026-09-09	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	—	us: 3 repetitions, file attached. outside: nobody
`B10`	GLM-5.3-Flash, agentic (8 DeepSWE tasks, 8 workers): per-task duration without MTP divided by MTP k=3 The paired ratio survives; «+40% tasks/h» does not — it measured a single task.	1.52× geometric mean; 95% CI 1.31–1.76 (t on the logs). Resolved 7/8 without MTP and 6/8 with MTP: quality is not comparable at this n 1.9 · 1.6 · 1.9 · 1.4 · 1.1 · 1.4 · 1.8 · 1.4	95% CI 1.31–1.76 (t on the logs)	8 paired tasks, 1 round per arm	2026-09-04	8×H100 80GB, TP8; arms on different days (3 Sep and 4 Sep)	agentico-glm53-mtp-20260903	2026-09-04 13.2 → 18.5 tasks/h (+40%) retracted With 8 tasks on 8 workers the wall clock is the slowest task, and it is the same task on both arms; the pair measures one task.	us: 1 round per arm, file attached. outside: nobody
`B9b`	GLM-5.3-Flash, all-reduce per call in the decode step (cross_device_reduce, TP8 rank 7) Median 5.79 µs; 89 of the 91 calls per step stay near it. The long waits are the first two calls of each step (medians about 1,014 and 242 µs), probably the 8 ranks aligning; measured under the profiler on one rank, and how much of the wait remains without the profiler was not measured.	5.79 µs median per call; mean 23.05 µs, max 5.9 ms. Message of 8 KiB per call (derived from the public config, not measured); the long waits sit in the first two calls of each step	1 trace, no repetition (spread across calls: p90 7.0 µs, mean 23.05 µs)	5,642 calls	2026-08-28	8×H100 80GB, TP8, torch.profiler	perfil-glm53-20260828	—	us: 1 trace, derived summary attached. outside: nobody
`B9`	GLM-5.3-Flash, kernels per decode step (profiler trace, 1 request, TP8 rank 7) The old number does not come from the trace; the current one comes from the summary derived from the trace (the 347 MB trace is not published).	1,694 1,484 on the main stream + 210 on the second, the same on every step; median 2.56 µs per kernel; under the profiler a step takes 7.65 ms and the main stream is busy for 88.7 % of it (ratio of the medians)	1 trace, no repetition (the same count on every step)	62 steps between 63 annotations	2026-08-28	8×H100 80GB, TP8, torch.profiler; decode with no speculation, thinking off	perfil-glm53-20260828	2026-08-29 1,529 kernels per step, 54 steps retracted It was the trace's kernel total (109,211) divided by about 71 steps; the trace has 63 annotated steps.	us: 1 trace, derived summary attached. outside: nobody
`C7`	GLM-5.3-Flash, GPQA Diamond (K=3 seeds) With the corrected grader (only the final line «Answer: X» counts; an answer cut off at the limit is wrong), the same answers give 89.85 %; the record said 90.19 %. The denominator is 197 per seed: one question did not fit the 65,536-token window on any seed (our client estimated its prompt at 1,970 tokens; it is about 2,836).	89.85 % 176/197 · 178/197 · 177/197; only the letter on the final line «Answer: X» counts (the old grader credited 2 answers cut off at the limit and reported 90.19 %); counting as wrong the question that did not fit the context window: 89.39 %. Our runner, T=1.0, thinking on 89.3 · 90.4 · 89.8	0.51 (sd across seeds; 2·se 0.59 pp)	198 × 3 seeds (valid per seed: 197, 197, 197)	2026-08-27	8×H100 80GB, TP8; our runner, 0-shot	gpqa-glm53-20260827-1653	2026-08-27 90.19 % retracted The old grader credited 2 answers cut off at the output ceiling (the same question in seeds 42 and 44, both at 64,629 tokens, the full output budget our runner set for that prompt) that never reached the final line «Answer: X»: one letter came from an «Answer:» written inside the reasoning, the other from a fallback that reads a conclusion in prose. Counting a cut-off answer as wrong, the same answers give 89.85 %.	us: 3 seeds, file attached. outside: nobody
`B4b`	Qwen3.8-Flash-Next, composite conc 32 on one TP8 replica File attached; the current value comes from resumo.json.	2,944.4 tok/s 2,891.4 · 2,967.1 · 2,974.7	46.1 (sd across repetitions)	3	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B4`	Qwen3.8-Flash-Next, 8×H100, single-stream without and with MTP k=3 File attached; the current value comes from resumo.json.	2.61× 136.0 → 354.6 tok/s 136.0 · 354.6	0.06 (sd across repetitions, propagated; 3.0 and 0.6 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B3b`	GLM-5.3-Flash, composite conc 32 on one TP8 replica Composite throughput at 32 streams on one TP8 replica. File attached. The current value comes from resumo.json.	2,661.9 tok/s 2,633.3 · 2,715.7 · 2,636.7	46.6 (sd across repetitions)	3	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions, file attached. outside: nobody
`B3`	GLM-5.3-Flash, 8×H100, single-stream without and with MTP k=5 First reading, file attached. The current value comes from resumo.json.	1.76× 155.2 → 272.6 tok/s 155.2 · 272.6	0.05 (sd across repetitions, propagated; 0.0 and 7.1 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions per arm, file attached. outside: nobody
`B1b`	Speculation alone, paired TP2 The second knob (TP2 instead of TP8) lowers the baseline; speculation alone, same session and same GPUs, is the current factor.	15.5× 112.7 → 1,746.5 tok/s 112.7 · 1,746.5	0.01 (sd across repetitions, propagated; 0.0 and 1.3 tok/s on the arms)	3 per arm	2026-08-26	2×H100 (TP2), shared node; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	partp2-20260826-1950	2026-08-26, paired window 1950 15.5× current; file attached TP2 with no speculation 112.7 tok/s → TP2 + DFlash-64 1,746.5 tok/s; same session, same GPUs, n=3. The 9.78× against TP8 (B1) is the conservative comparison.	us: 3 repetitions per arm, file attached. outside: nobody
`B2`	Aggregate throughput of four replicas on the same node Reproduced. The ruler does not knock a number down on reflex: here the remeasurement confirmed it. File attached.	29,503 tok/s 29,467.0 · 29,421.6 · 29,620.4	104.2 (sd across repetitions)	3 (overlap 118 s)	2026-08-25	8×H100 80GB (four replicas, conc=32 per replica); synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	composto4xc32-20260825-0244	2026-08-21 29,567 tok/s first reading; JSON absent Aggregate of four replicas. Audit B checked the sum and marked simultaneity as not archived. 2026-08-25 29,503 tok/s reproduced; current; file attached conc=32 per replica, n=3, overlap 118 s, aggregate 29,503 ± 104 tok/s. The file has no interference percentage.	us: 3 repetitions, file attached. outside: nobody
`B1`	Single-stream decode speedup The first factor does not survive; the order of magnitude does. The 25 Aug remeasurement is the current value, file attached. Measured on a synthetic filler prompt; speculative acceptance depends on the text, so this is not a real-workload speedup.	9.78× 183.6 → 1,795.1 tok/s 183.6 · 1,795.1	0.01 (sd across repetitions, propagated; 0.1 and 1.4 tok/s on the arms)	3 per arm	2026-08-25	8×H100 80GB (TP8, no speculation; DFlash b64 on TP2). Two knobs. synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap.	single10x-20260825-0035	2026-08-21 10.1× retracted One 90-second reading of TP2 + DFlash-64 (1,782 tok/s, 256-token prompts) divided by 176 tok/s, a TP8 baseline measured with 8,192-token prompts; the comparable 256-token baseline read 187 tok/s, which gives 9.5×. Those window files are not published on this site. 2026-08-25, window 0011 200.7 tok/s (TP8 baseline; n=3, sd 2.0) retracted in favor of 0035 First TP8 baseline reading. File: brutos/single10x-20260825-0011/resumo.json. The scheduled job was not suspended; the DFlash arm was not measured (content gate). Why 200.7 became 183.6 is not shown. 2026-08-25, window 0035 9.78×. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b current; file attached TP8, no speculation, 183.6 tok/s → TP2 + DFlash-64, 1,795.1 tok/s; conc=1, n=3. Two knobs moved. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b.	us: 3 repetitions per arm, file attached. outside: nobody
`C6`	SWE-bench gold-sanity: the gold patch through the same generate→apply→test path 0/25 resolved, 92% apply, n=25. In 10 of the 25 the runner could not collect the test IDs (8 Django, 2 SymPy); 13 never ran a test in the host environment; 2 IDs are not SWE-bench instances. File attached.	0/25 resolved, 92% apply, n=25 before the judge and classifier fix: test_fail 23, infra_error 2; after (swe_gold2.json): test_fail 4, infra_error 21	95% CI 0.0–13.3 % on the resolved rate (from the file; the 2 infra errors are in the denominator)	25	2026-08-04	not stated in a file	swe_gold_sanity_20260804	2026-08-04, before the judge and classifier fix 0/25 resolved, 92% apply; test_fail 23, infra_error 2 current; file attached Gold patch on the same code path. 10 test IDs not collected (8 Django, 2 SymPy), 13 environment failures, 2 IDs that are not SWE-bench instances (the 8% apply gap). File: swe_gold_sanity.json. 2026-08-04, after the judge and classifier fix 0/25 resolved, 92% apply; infra_error 21, test_fail 4 reclassification; file attached Same n=25, same apply 0.92. File: swe_gold2.json.	us: 1 gold-sanity run, file attached. outside: nobody
`C5b`	Terminal-Bench gold46, DeepSeek V4-Flash, after the adapter fix VOID: measured before the gold gate existed; a floor, not a measurement. DeepSeek V4-Flash, text scaffold. Retracted.	no archived raw file	—	44 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	2026-08-11 8/44 = 18.2% retracted Measured before the gold gate existed. A floor, not a measurement. No file published on this site.	us: 0 repetition files — n=44 claimed; JSON absent. outside: nobody
`C5a`	Terminal-Bench 0/20 (first score) VOID: the adapter called a host path inside the container (exit 127)	no archived raw file	—	—	2026-08	not stated in a file (absent)	no archived raw file	original score 0/20 retracted as model capability bash <host path> inside the container → exit 127 → «Verification failed» every time.	us: 0 repetition files — reproducible as an adapter bug in the old code; the 0/20 JSON is absent. outside: nobody
`C4`	Spread of GPQA accuracy at temp 0, same server, fixed seed The phenomenon (temp 0 is not deterministic) survives; a single cause does not	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — K=5, n=40 claimed; JSON absent. outside: nobody
`C3`	Thinking versus chat on GPQA (thinking recipe versus greedy) «zero pp as proof of zero» does not survive; the detection floor is ~5 pp at n=40 K=1	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	original formulation zero pp retracted as proof of a null effect n=40 K=1; the pre-registered kill (+3 pp) is below the floor (~5 pp). It does not reject +8 pp. audit formulation no effect ≥ ~5 pp detected revised reading; JSON absent Same probe. Confound: thinking is temp 1.0 / top_p 0.95, chat is greedy temp 0.	us: 0 repetition files — n=40 K=1; no archived raw file. outside: nobody
`A6`	maj@5 versus K=1 on the same server (GPQA Diamond, temp 0.7) Paired contrast on one server; do not mix in a baseline from another instance	no archived raw file	—	198 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — n=198, K=5 claimed; JSON absent. outside: nobody
`A4`	Training-lab spread (identical runs, only the seed changes) Held until the 8-seed file is archived.	no archived raw file	—	8 (claim; no archived raw file)	2026-07-31	8×H100 (claim conditions; no archived raw file)	no archived raw file	—	us: 0 repetition files — n=8 claimed; the 8-seed file is absent. outside: nobody
`A2`	Isolated effect of the weight-decay ramp (0.01 → 0.10) Null on the sub-Chinchilla horizon; 64% of the composite gain was the baseline learning rate	no archived raw file	—	—	2026-07-31	8×H100, 220M, 1,200 steps (claim conditions)	no archived raw file	original catalog gain attributed to the composite scheduler retracted as the mechanism The off arm used a 500-step warmup (42% of training) and a mean learning rate 2.82× lower. paired decomposition, n=7 WD ramp Δ = +0.0048 (below δ_det = 0.0277) later reading; no archived raw file Null effect of the ramp on this horizon (~0.9 tok/param). It does not generalize to over-training.	us: 0 repetition files — n=7 arms claimed; no archived raw file. outside: nobody
`C1`	Document-quality filter versus random Held: the held-out remeasurement has no archived raw file and is being repeated. The in-sample −64.5% stays retracted.	no archived raw file	—	1 (claim; no archived raw file)	2026-07-30	8×H100, ~300M, 1,500 steps (claim conditions)	no archived raw file	eval on the training set −64.5% PPL retracted The metric was in-sample.	us: 0 repetition files — n=1 seed; no archived raw file. outside: nobody
`A1`	Apparent gain of windowed attention after the causal-mask fix The cautionary tale survives; «all 28 causal configurations» does not	no archived raw file	—	—	2026-07-30	8×H100 (claim conditions)	no archived raw file	before the fix −4.1368 retracted as a real gain The gain was future information. after the fix factorial main effect −0.1523; isolated −0.042 (1 seed); no archived raw file later reading; no archived raw file Same design, physical held-out. The factorial main effect is not the isolated effect. No file published on this site.	us: 0 repetition files — no archived raw file. outside: nobody
`A5`	FP8 eager versus BF16 training throughput (full step) SURVIVES as PyTorch eager without fusion; not as a law of the format	no archived raw file	—	—	2026-07	H100 SXM 80GB (claim conditions)	no archived raw file	bench with w.detach() w8 = 1.51× retracted autograd skipped dW: 6MKN → 4MKN, theoretical ratio 1.50×. The measured 1.51× was the bug. full step without detach 0.89× to 0.94× vs BF16 later reading; no archived raw file Two HBM transposes per linear in the backward of torch._scaled_mm. The sweep JSON is not published on this site.	us: 0 repetition files — 8 configurations from 350M to 3B claimed; no archived raw file. outside: nobody
`B7`	Time to first audio, voice stack ~150 ms does not survive. 531.6 ms sums clocks that do not measure time to first audio end to end.	no archived raw file	—	20 (claim; no archived raw file)	—	not stated in a file (absent)	no archived raw file	later decomposition ~150 ms retracted It is not in the meter, and it is not the same instrument as the voice-window p50. voice window p50 531.6 ms retracted as time to first audio The 531.6 ms sums three clocks that do not measure time to first audio end to end: the STT clock includes the handshake, thinking was left on, and the TTS figure is a control frame. Not an end-to-end first-audio latency. No file published on this site.	us: 0 repetition files — n=20 claimed for the p50; JSON absent. outside: nobody
`B5`	HBM measured on the H100 roofline (phase 0) Plausible against the datasheet; the Phase 0 raw JSON is not published on this site	no archived raw file	—	—	—	H100 SXM (claim conditions)	no archived raw file	same internal record GEMM peak 781 TFLOPS vs 761 TFLOPS two anchors, no JSON The internal record disagrees with itself. Without the phase-0 JSON neither figure is a current value.	us: 0 repetition files — no archived raw file. outside: nobody

measurements by n, largest first; missing n at the end
id	what was measured	value	uncertainty	n	date	hardware	raw file	history	replication
`C11`	GLM-5.3-Flash, MMLU-Pro with the official extractor (TIGER-AI-Lab) With the official extractor, the same answers give 82.96 %; the runner's old grader gave 72.59 %.	82.96 % 9,981/12,031 with the official MMLU-Pro extractor; the runner's old grader gave 8,733/12,031 = 72.59 %. 122 unfinished answers count as wrong. Also reading \boxed: 83.48 %. Our own 0-shot prompt, without the official harness format instruction	0.34 (binomial standard error; 2·se 0.69 pp)	12,032 items (K=1); 12,031 in the denominator	2026-09-23	8×H100 80GB, TP8; our runner	mmlu-pro-glm53-20260923	2026-09-21 72.59 % retracted Produced by the runner's old grader: it does not match \boxed{\text{G. …}} and it ignores case ('The answer is exactly…' becomes E). The official extractor on the same answers gives 82.96 %.	us: 1 run, file attached. outside: nobody
`B9b`	GLM-5.3-Flash, all-reduce per call in the decode step (cross_device_reduce, TP8 rank 7) Median 5.79 µs; 89 of the 91 calls per step stay near it. The long waits are the first two calls of each step (medians about 1,014 and 242 µs), probably the 8 ranks aligning; measured under the profiler on one rank, and how much of the wait remains without the profiler was not measured.	5.79 µs median per call; mean 23.05 µs, max 5.9 ms. Message of 8 KiB per call (derived from the public config, not measured); the long waits sit in the first two calls of each step	1 trace, no repetition (spread across calls: p90 7.0 µs, mean 23.05 µs)	5,642 calls	2026-08-28	8×H100 80GB, TP8, torch.profiler	perfil-glm53-20260828	—	us: 1 trace, derived summary attached. outside: nobody
`C8`	GLM-5.3-Flash, MATH-500 (K=3 seeds), corrected answer normalizer The old number measured the normalizer, not the model. The current one reapplies the corrected normalizer to the same answers.	99.00 % 1,485/1,500 with the new normalizer; the old one gave 76.20 %. 148 distinct pairs reclassified (342 items), read by hand; 15 failures left: 7 truncated, 8 wrong 99.4 · 98.2 · 99.4	0.69 (sd across seeds; 2·se 0.80 pp)	500 × 3 seeds	2026-09-19	8×H100 80GB, TP8; our runner	math500-glm53-20260919	2026-09-19 76.20 % retracted The normalizer marked equivalent answers wrong (x = 4 against 4; \dfrac against \frac).	us: 3 seeds, file attached. outside: nobody
`C9b`	GLM-5.3-Flash, HLE text only, strict exact match (K=1, no judge) No judge, and without the numeric tolerance that credited different answers: a floor.	8.75 % 127/1,451 strict: 131 by the file's count, 4 only by the numeric tolerance. No judge: a floor, not comparable to the official score	0.74 (binomial standard error; 2·se 1.48 pp)	1,451 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 9.31 % (131/1,407) retracted The denominator dropped 44 items that had not converged even at 163,840 tokens, and the relative tolerance credited 4 answers that differed from the key.	us: 1 run, file attached. outside: nobody
`C7`	GLM-5.3-Flash, GPQA Diamond (K=3 seeds) With the corrected grader (only the final line «Answer: X» counts; an answer cut off at the limit is wrong), the same answers give 89.85 %; the record said 90.19 %. The denominator is 197 per seed: one question did not fit the 65,536-token window on any seed (our client estimated its prompt at 1,970 tokens; it is about 2,836).	89.85 % 176/197 · 178/197 · 177/197; only the letter on the final line «Answer: X» counts (the old grader credited 2 answers cut off at the limit and reported 90.19 %); counting as wrong the question that did not fit the context window: 89.39 %. Our runner, T=1.0, thinking on 89.3 · 90.4 · 89.8	0.51 (sd across seeds; 2·se 0.59 pp)	198 × 3 seeds (valid per seed: 197, 197, 197)	2026-08-27	8×H100 80GB, TP8; our runner, 0-shot	gpqa-glm53-20260827-1653	2026-08-27 90.19 % retracted The old grader credited 2 answers cut off at the output ceiling (the same question in seeds 42 and 44, both at 64,629 tokens, the full output budget our runner set for that prompt) that never reached the final line «Answer: X»: one letter came from an «Answer:» written inside the reasoning, the other from a fallback that reads a conclusion in prose. Counting a cut-off answer as wrong, the same answers give 89.85 %.	us: 3 seeds, file attached. outside: nobody
`C9a`	GLM-5.3-Flash, HLE text only, multiple choice (K=1, no judge) Denominator corrected; with no judge, this is not the official score.	35.98 % 163/453; no judge; answers that did not converge count as wrong	2.25 (binomial standard error; 2·se 4.51 pp)	453 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 36.06 % (163/452) retracted The denominator dropped the multiple-choice item that had not converged even at 163,840 tokens.	us: 1 run, file attached. outside: nobody
`A6`	maj@5 versus K=1 on the same server (GPQA Diamond, temp 0.7) Paired contrast on one server; do not mix in a baseline from another instance	no archived raw file	—	198 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — n=198, K=5 claimed; JSON absent. outside: nobody
`B9`	GLM-5.3-Flash, kernels per decode step (profiler trace, 1 request, TP8 rank 7) The old number does not come from the trace; the current one comes from the summary derived from the trace (the 347 MB trace is not published).	1,694 1,484 on the main stream + 210 on the second, the same on every step; median 2.56 µs per kernel; under the profiler a step takes 7.65 ms and the main stream is busy for 88.7 % of it (ratio of the medians)	1 trace, no repetition (the same count on every step)	62 steps between 63 annotations	2026-08-28	8×H100 80GB, TP8, torch.profiler; decode with no speculation, thinking off	perfil-glm53-20260828	2026-08-29 1,529 kernels per step, 54 steps retracted It was the trace's kernel total (109,211) divided by about 71 steps; the trace has 63 annotated steps.	us: 1 trace, derived summary attached. outside: nobody
`C5b`	Terminal-Bench gold46, DeepSeek V4-Flash, after the adapter fix VOID: measured before the gold gate existed; a floor, not a measurement. DeepSeek V4-Flash, text scaffold. Retracted.	no archived raw file	—	44 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	2026-08-11 8/44 = 18.2% retracted Measured before the gold gate existed. A floor, not a measurement. No file published on this site.	us: 0 repetition files — n=44 claimed; JSON absent. outside: nobody
`C3`	Thinking versus chat on GPQA (thinking recipe versus greedy) «zero pp as proof of zero» does not survive; the detection floor is ~5 pp at n=40 K=1	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	original formulation zero pp retracted as proof of a null effect n=40 K=1; the pre-registered kill (+3 pp) is below the floor (~5 pp). It does not reject +8 pp. audit formulation no effect ≥ ~5 pp detected revised reading; JSON absent Same probe. Confound: thinking is temp 1.0 / top_p 0.95, chat is greedy temp 0.	us: 0 repetition files — n=40 K=1; no archived raw file. outside: nobody
`C4`	Spread of GPQA accuracy at temp 0, same server, fixed seed The phenomenon (temp 0 is not deterministic) survives; a single cause does not	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — K=5, n=40 claimed; JSON absent. outside: nobody
`C6`	SWE-bench gold-sanity: the gold patch through the same generate→apply→test path 0/25 resolved, 92% apply, n=25. In 10 of the 25 the runner could not collect the test IDs (8 Django, 2 SymPy); 13 never ran a test in the host environment; 2 IDs are not SWE-bench instances. File attached.	0/25 resolved, 92% apply, n=25 before the judge and classifier fix: test_fail 23, infra_error 2; after (swe_gold2.json): test_fail 4, infra_error 21	95% CI 0.0–13.3 % on the resolved rate (from the file; the 2 infra errors are in the denominator)	25	2026-08-04	not stated in a file	swe_gold_sanity_20260804	2026-08-04, before the judge and classifier fix 0/25 resolved, 92% apply; test_fail 23, infra_error 2 current; file attached Gold patch on the same code path. 10 test IDs not collected (8 Django, 2 SymPy), 13 environment failures, 2 IDs that are not SWE-bench instances (the 8% apply gap). File: swe_gold_sanity.json. 2026-08-04, after the judge and classifier fix 0/25 resolved, 92% apply; infra_error 21, test_fail 4 reclassification; file attached Same n=25, same apply 0.92. File: swe_gold2.json.	us: 1 gold-sanity run, file attached. outside: nobody
`B7`	Time to first audio, voice stack ~150 ms does not survive. 531.6 ms sums clocks that do not measure time to first audio end to end.	no archived raw file	—	20 (claim; no archived raw file)	—	not stated in a file (absent)	no archived raw file	later decomposition ~150 ms retracted It is not in the meter, and it is not the same instrument as the voice-window p50. voice window p50 531.6 ms retracted as time to first audio The 531.6 ms sums three clocks that do not measure time to first audio end to end: the STT clock includes the handshake, thinking was left on, and the TTS figure is a control frame. Not an end-to-end first-audio latency. No file published on this site.	us: 0 repetition files — n=20 claimed for the p50; JSON absent. outside: nobody
`A4`	Training-lab spread (identical runs, only the seed changes) Held until the 8-seed file is archived.	no archived raw file	—	8 (claim; no archived raw file)	2026-07-31	8×H100 (claim conditions; no archived raw file)	no archived raw file	—	us: 0 repetition files — n=8 claimed; the 8-seed file is absent. outside: nobody
`B10`	GLM-5.3-Flash, agentic (8 DeepSWE tasks, 8 workers): per-task duration without MTP divided by MTP k=3 The paired ratio survives; «+40% tasks/h» does not — it measured a single task.	1.52× geometric mean; 95% CI 1.31–1.76 (t on the logs). Resolved 7/8 without MTP and 6/8 with MTP: quality is not comparable at this n 1.9 · 1.6 · 1.9 · 1.4 · 1.1 · 1.4 · 1.8 · 1.4	95% CI 1.31–1.76 (t on the logs)	8 paired tasks, 1 round per arm	2026-09-04	8×H100 80GB, TP8; arms on different days (3 Sep and 4 Sep)	agentico-glm53-mtp-20260903	2026-09-04 13.2 → 18.5 tasks/h (+40%) retracted With 8 tasks on 8 workers the wall clock is the slowest task, and it is the same task on both arms; the pair measures one task.	us: 1 round per arm, file attached. outside: nobody
`C10b`	GLM-5.3-Flash, DeepSWE 111 tasks, v1lite scaffold variant (K=6) The scaffold did not move the result; the recorded value came from rounded percentages.	64.21 % 81/108 · 68/105 · 71/109 · 66/106 · 62/105 · 62/105; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 62.80 % (2·se 5.41 pp); delta against the baseline −1.11 pp (2·se 5.56) 75.0 · 64.8 · 65.1 · 62.3 · 59.0 · 59.0	5.91 (sd across repetitions; 2·se 4.83 pp)	111 tasks × K=6	2026-09-21	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	2026-09-21 64.18 ± 4.85 retracted A mean of rounded percentages; two repetitions were written down as 58.95 % instead of 59.05 % (62/105).	us: 6 repetitions, file attached. outside: nobody
`B1`	Single-stream decode speedup The first factor does not survive; the order of magnitude does. The 25 Aug remeasurement is the current value, file attached. Measured on a synthetic filler prompt; speculative acceptance depends on the text, so this is not a real-workload speedup.	9.78× 183.6 → 1,795.1 tok/s 183.6 · 1,795.1	0.01 (sd across repetitions, propagated; 0.1 and 1.4 tok/s on the arms)	3 per arm	2026-08-25	8×H100 80GB (TP8, no speculation; DFlash b64 on TP2). Two knobs. synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap.	single10x-20260825-0035	2026-08-21 10.1× retracted One 90-second reading of TP2 + DFlash-64 (1,782 tok/s, 256-token prompts) divided by 176 tok/s, a TP8 baseline measured with 8,192-token prompts; the comparable 256-token baseline read 187 tok/s, which gives 9.5×. Those window files are not published on this site. 2026-08-25, window 0011 200.7 tok/s (TP8 baseline; n=3, sd 2.0) retracted in favor of 0035 First TP8 baseline reading. File: brutos/single10x-20260825-0011/resumo.json. The scheduled job was not suspended; the DFlash arm was not measured (content gate). Why 200.7 became 183.6 is not shown. 2026-08-25, window 0035 9.78×. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b current; file attached TP8, no speculation, 183.6 tok/s → TP2 + DFlash-64, 1,795.1 tok/s; conc=1, n=3. Two knobs moved. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b.	us: 3 repetitions per arm, file attached. outside: nobody
`B1b`	Speculation alone, paired TP2 The second knob (TP2 instead of TP8) lowers the baseline; speculation alone, same session and same GPUs, is the current factor.	15.5× 112.7 → 1,746.5 tok/s 112.7 · 1,746.5	0.01 (sd across repetitions, propagated; 0.0 and 1.3 tok/s on the arms)	3 per arm	2026-08-26	2×H100 (TP2), shared node; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	partp2-20260826-1950	2026-08-26, paired window 1950 15.5× current; file attached TP2 with no speculation 112.7 tok/s → TP2 + DFlash-64 1,746.5 tok/s; same session, same GPUs, n=3. The 9.78× against TP8 (B1) is the conservative comparison.	us: 3 repetitions per arm, file attached. outside: nobody
`B2`	Aggregate throughput of four replicas on the same node Reproduced. The ruler does not knock a number down on reflex: here the remeasurement confirmed it. File attached.	29,503 tok/s 29,467.0 · 29,421.6 · 29,620.4	104.2 (sd across repetitions)	3 (overlap 118 s)	2026-08-25	8×H100 80GB (four replicas, conc=32 per replica); synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	composto4xc32-20260825-0244	2026-08-21 29,567 tok/s first reading; JSON absent Aggregate of four replicas. Audit B checked the sum and marked simultaneity as not archived. 2026-08-25 29,503 tok/s reproduced; current; file attached conc=32 per replica, n=3, overlap 118 s, aggregate 29,503 ± 104 tok/s. The file has no interference percentage.	us: 3 repetitions, file attached. outside: nobody
`B3`	GLM-5.3-Flash, 8×H100, single-stream without and with MTP k=5 First reading, file attached. The current value comes from resumo.json.	1.76× 155.2 → 272.6 tok/s 155.2 · 272.6	0.05 (sd across repetitions, propagated; 0.0 and 7.1 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions per arm, file attached. outside: nobody
`B3b`	GLM-5.3-Flash, composite conc 32 on one TP8 replica Composite throughput at 32 streams on one TP8 replica. File attached. The current value comes from resumo.json.	2,661.9 tok/s 2,633.3 · 2,715.7 · 2,636.7	46.6 (sd across repetitions)	3	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions, file attached. outside: nobody
`B4`	Qwen3.8-Flash-Next, 8×H100, single-stream without and with MTP k=3 File attached; the current value comes from resumo.json.	2.61× 136.0 → 354.6 tok/s 136.0 · 354.6	0.06 (sd across repetitions, propagated; 3.0 and 0.6 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B4b`	Qwen3.8-Flash-Next, composite conc 32 on one TP8 replica File attached; the current value comes from resumo.json.	2,944.4 tok/s 2,891.4 · 2,967.1 · 2,974.7	46.1 (sd across repetitions)	3	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B8`	GLM-5.3-Flash, one stream: native MTP k=3 against no speculation (17 Sep window) Reproduced from the raw file. The current value comes from the per-request repetitions; tok/s includes TTFT. Synthetic filler prompt at temperature 0: speculative acceptance depends on the text; on real agent tasks MTP gives 1.52× per task (B10).	2.04× 156.05 ± 0.07 → 317.90 ± 2.92 tok/s (±2·se); repetitions between 2.02× and 2.05× 156.1 · 317.9	0.02 (sd across repetitions, propagated; 0.06 and 2.53 tok/s on the arms)	3 per arm	2026-09-17	8×H100 80GB, TP8, native FP8, dedicated vLLM image; 1 request at a time, synthetic 256-token filler prompt, greedy decoding (temperature 0), 512-token output cap, thinking off	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions per arm, file attached. outside: nobody
`B8b`	GLM-5.3-Flash, one stream, no speculation: nvidia-smi counters (memory busy over time) The old label «SM active» was wrong: utilization.gpu is the fraction of time any kernel is running.	21.80 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 98.74 % (fraction of time any kernel is running; this is not per-SM activity) · 256.9 W per GPU 21.9 · 21.8 · 21.8	0.06 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	2026-09-17 «SM active» 98.7 % retracted as a label The column is nvidia-smi utilization.gpu: the fraction of time any kernel is running, not per-SM activity. This node has no per-SM activity counter.	us: 3 repetitions, file attached. outside: nobody
`B8c`	GLM-5.3-Flash, one stream with MTP k=3: nvidia-smi counters (memory busy over time) The same step with MTP keeps memory busier and draws more power: more useful work per step.	26.95 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 99.16 % (fraction of time any kernel is running; this is not per-SM activity) · 308.1 W per GPU 26.8 · 26.8 · 27.2	0.20 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions, file attached. outside: nobody
`C10a`	GLM-5.3-Flash, DeepSWE 111 tasks, our agent (K=3) Reproduced. Our agent, not the vendor's. The first repetition mixes two serving setups: 29 tasks at 20 workers without MTP and a 1,800 s timeout, 82 at 8 workers with MTP k=3 and 3,600 s. 111 of the 113 tasks: the gold patch fails our harness on 2.	65.32 % 67/106 · 70/108 · 72/106; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 63.72 % (2·se 2.65 pp) 63.2 · 64.8 · 67.9	2.40 (sd across repetitions; 2·se 2.77 pp)	111 tasks × K=3	2026-09-09	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	—	us: 3 repetitions, file attached. outside: nobody
`C1`	Document-quality filter versus random Held: the held-out remeasurement has no archived raw file and is being repeated. The in-sample −64.5% stays retracted.	no archived raw file	—	1 (claim; no archived raw file)	2026-07-30	8×H100, ~300M, 1,500 steps (claim conditions)	no archived raw file	eval on the training set −64.5% PPL retracted The metric was in-sample.	us: 0 repetition files — n=1 seed; no archived raw file. outside: nobody
`A1`	Apparent gain of windowed attention after the causal-mask fix The cautionary tale survives; «all 28 causal configurations» does not	no archived raw file	—	—	2026-07-30	8×H100 (claim conditions)	no archived raw file	before the fix −4.1368 retracted as a real gain The gain was future information. after the fix factorial main effect −0.1523; isolated −0.042 (1 seed); no archived raw file later reading; no archived raw file Same design, physical held-out. The factorial main effect is not the isolated effect. No file published on this site.	us: 0 repetition files — no archived raw file. outside: nobody
`A2`	Isolated effect of the weight-decay ramp (0.01 → 0.10) Null on the sub-Chinchilla horizon; 64% of the composite gain was the baseline learning rate	no archived raw file	—	—	2026-07-31	8×H100, 220M, 1,200 steps (claim conditions)	no archived raw file	original catalog gain attributed to the composite scheduler retracted as the mechanism The off arm used a 500-step warmup (42% of training) and a mean learning rate 2.82× lower. paired decomposition, n=7 WD ramp Δ = +0.0048 (below δ_det = 0.0277) later reading; no archived raw file Null effect of the ramp on this horizon (~0.9 tok/param). It does not generalize to over-training.	us: 0 repetition files — n=7 arms claimed; no archived raw file. outside: nobody
`A5`	FP8 eager versus BF16 training throughput (full step) SURVIVES as PyTorch eager without fusion; not as a law of the format	no archived raw file	—	—	2026-07	H100 SXM 80GB (claim conditions)	no archived raw file	bench with w.detach() w8 = 1.51× retracted autograd skipped dW: 6MKN → 4MKN, theoretical ratio 1.50×. The measured 1.51× was the bug. full step without detach 0.89× to 0.94× vs BF16 later reading; no archived raw file Two HBM transposes per linear in the backward of torch._scaled_mm. The sweep JSON is not published on this site.	us: 0 repetition files — 8 configurations from 350M to 3B claimed; no archived raw file. outside: nobody
`B5`	HBM measured on the H100 roofline (phase 0) Plausible against the datasheet; the Phase 0 raw JSON is not published on this site	no archived raw file	—	—	—	H100 SXM (claim conditions)	no archived raw file	same internal record GEMM peak 781 TFLOPS vs 761 TFLOPS two anchors, no JSON The internal record disagrees with itself. Without the phase-0 JSON neither figure is a current value.	us: 0 repetition files — no archived raw file. outside: nobody
`C5a`	Terminal-Bench 0/20 (first score) VOID: the adapter called a host path inside the container (exit 127)	no archived raw file	—	—	2026-08	not stated in a file (absent)	no archived raw file	original score 0/20 retracted as model capability bash <host path> inside the container → exit 127 → «Verification failed» every time.	us: 0 repetition files — reproducible as an adapter bug in the old code; the 0/20 JSON is absent. outside: nobody

measurements with a file first
id	what was measured	value	uncertainty	n	date	hardware	raw file	history	replication
`B1`	Single-stream decode speedup The first factor does not survive; the order of magnitude does. The 25 Aug remeasurement is the current value, file attached. Measured on a synthetic filler prompt; speculative acceptance depends on the text, so this is not a real-workload speedup.	9.78× 183.6 → 1,795.1 tok/s 183.6 · 1,795.1	0.01 (sd across repetitions, propagated; 0.1 and 1.4 tok/s on the arms)	3 per arm	2026-08-25	8×H100 80GB (TP8, no speculation; DFlash b64 on TP2). Two knobs. synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap.	single10x-20260825-0035	2026-08-21 10.1× retracted One 90-second reading of TP2 + DFlash-64 (1,782 tok/s, 256-token prompts) divided by 176 tok/s, a TP8 baseline measured with 8,192-token prompts; the comparable 256-token baseline read 187 tok/s, which gives 9.5×. Those window files are not published on this site. 2026-08-25, window 0011 200.7 tok/s (TP8 baseline; n=3, sd 2.0) retracted in favor of 0035 First TP8 baseline reading. File: brutos/single10x-20260825-0011/resumo.json. The scheduled job was not suspended; the DFlash arm was not measured (content gate). Why 200.7 became 183.6 is not shown. 2026-08-25, window 0035 9.78×. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b current; file attached TP8, no speculation, 183.6 tok/s → TP2 + DFlash-64, 1,795.1 tok/s; conc=1, n=3. Two knobs moved. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b.	us: 3 repetitions per arm, file attached. outside: nobody
`B10`	GLM-5.3-Flash, agentic (8 DeepSWE tasks, 8 workers): per-task duration without MTP divided by MTP k=3 The paired ratio survives; «+40% tasks/h» does not — it measured a single task.	1.52× geometric mean; 95% CI 1.31–1.76 (t on the logs). Resolved 7/8 without MTP and 6/8 with MTP: quality is not comparable at this n 1.9 · 1.6 · 1.9 · 1.4 · 1.1 · 1.4 · 1.8 · 1.4	95% CI 1.31–1.76 (t on the logs)	8 paired tasks, 1 round per arm	2026-09-04	8×H100 80GB, TP8; arms on different days (3 Sep and 4 Sep)	agentico-glm53-mtp-20260903	2026-09-04 13.2 → 18.5 tasks/h (+40%) retracted With 8 tasks on 8 workers the wall clock is the slowest task, and it is the same task on both arms; the pair measures one task.	us: 1 round per arm, file attached. outside: nobody
`B1b`	Speculation alone, paired TP2 The second knob (TP2 instead of TP8) lowers the baseline; speculation alone, same session and same GPUs, is the current factor.	15.5× 112.7 → 1,746.5 tok/s 112.7 · 1,746.5	0.01 (sd across repetitions, propagated; 0.0 and 1.3 tok/s on the arms)	3 per arm	2026-08-26	2×H100 (TP2), shared node; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	partp2-20260826-1950	2026-08-26, paired window 1950 15.5× current; file attached TP2 with no speculation 112.7 tok/s → TP2 + DFlash-64 1,746.5 tok/s; same session, same GPUs, n=3. The 9.78× against TP8 (B1) is the conservative comparison.	us: 3 repetitions per arm, file attached. outside: nobody
`B2`	Aggregate throughput of four replicas on the same node Reproduced. The ruler does not knock a number down on reflex: here the remeasurement confirmed it. File attached.	29,503 tok/s 29,467.0 · 29,421.6 · 29,620.4	104.2 (sd across repetitions)	3 (overlap 118 s)	2026-08-25	8×H100 80GB (four replicas, conc=32 per replica); synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	composto4xc32-20260825-0244	2026-08-21 29,567 tok/s first reading; JSON absent Aggregate of four replicas. Audit B checked the sum and marked simultaneity as not archived. 2026-08-25 29,503 tok/s reproduced; current; file attached conc=32 per replica, n=3, overlap 118 s, aggregate 29,503 ± 104 tok/s. The file has no interference percentage.	us: 3 repetitions, file attached. outside: nobody
`B3`	GLM-5.3-Flash, 8×H100, single-stream without and with MTP k=5 First reading, file attached. The current value comes from resumo.json.	1.76× 155.2 → 272.6 tok/s 155.2 · 272.6	0.05 (sd across repetitions, propagated; 0.0 and 7.1 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions per arm, file attached. outside: nobody
`B3b`	GLM-5.3-Flash, composite conc 32 on one TP8 replica Composite throughput at 32 streams on one TP8 replica. File attached. The current value comes from resumo.json.	2,661.9 tok/s 2,633.3 · 2,715.7 · 2,636.7	46.6 (sd across repetitions)	3	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions, file attached. outside: nobody
`B4`	Qwen3.8-Flash-Next, 8×H100, single-stream without and with MTP k=3 File attached; the current value comes from resumo.json.	2.61× 136.0 → 354.6 tok/s 136.0 · 354.6	0.06 (sd across repetitions, propagated; 3.0 and 0.6 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B4b`	Qwen3.8-Flash-Next, composite conc 32 on one TP8 replica File attached; the current value comes from resumo.json.	2,944.4 tok/s 2,891.4 · 2,967.1 · 2,974.7	46.1 (sd across repetitions)	3	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B8`	GLM-5.3-Flash, one stream: native MTP k=3 against no speculation (17 Sep window) Reproduced from the raw file. The current value comes from the per-request repetitions; tok/s includes TTFT. Synthetic filler prompt at temperature 0: speculative acceptance depends on the text; on real agent tasks MTP gives 1.52× per task (B10).	2.04× 156.05 ± 0.07 → 317.90 ± 2.92 tok/s (±2·se); repetitions between 2.02× and 2.05× 156.1 · 317.9	0.02 (sd across repetitions, propagated; 0.06 and 2.53 tok/s on the arms)	3 per arm	2026-09-17	8×H100 80GB, TP8, native FP8, dedicated vLLM image; 1 request at a time, synthetic 256-token filler prompt, greedy decoding (temperature 0), 512-token output cap, thinking off	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions per arm, file attached. outside: nobody
`B8b`	GLM-5.3-Flash, one stream, no speculation: nvidia-smi counters (memory busy over time) The old label «SM active» was wrong: utilization.gpu is the fraction of time any kernel is running.	21.80 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 98.74 % (fraction of time any kernel is running; this is not per-SM activity) · 256.9 W per GPU 21.9 · 21.8 · 21.8	0.06 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	2026-09-17 «SM active» 98.7 % retracted as a label The column is nvidia-smi utilization.gpu: the fraction of time any kernel is running, not per-SM activity. This node has no per-SM activity counter.	us: 3 repetitions, file attached. outside: nobody
`B8c`	GLM-5.3-Flash, one stream with MTP k=3: nvidia-smi counters (memory busy over time) The same step with MTP keeps memory busier and draws more power: more useful work per step.	26.95 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 99.16 % (fraction of time any kernel is running; this is not per-SM activity) · 308.1 W per GPU 26.8 · 26.8 · 27.2	0.20 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions, file attached. outside: nobody
`B9`	GLM-5.3-Flash, kernels per decode step (profiler trace, 1 request, TP8 rank 7) The old number does not come from the trace; the current one comes from the summary derived from the trace (the 347 MB trace is not published).	1,694 1,484 on the main stream + 210 on the second, the same on every step; median 2.56 µs per kernel; under the profiler a step takes 7.65 ms and the main stream is busy for 88.7 % of it (ratio of the medians)	1 trace, no repetition (the same count on every step)	62 steps between 63 annotations	2026-08-28	8×H100 80GB, TP8, torch.profiler; decode with no speculation, thinking off	perfil-glm53-20260828	2026-08-29 1,529 kernels per step, 54 steps retracted It was the trace's kernel total (109,211) divided by about 71 steps; the trace has 63 annotated steps.	us: 1 trace, derived summary attached. outside: nobody
`B9b`	GLM-5.3-Flash, all-reduce per call in the decode step (cross_device_reduce, TP8 rank 7) Median 5.79 µs; 89 of the 91 calls per step stay near it. The long waits are the first two calls of each step (medians about 1,014 and 242 µs), probably the 8 ranks aligning; measured under the profiler on one rank, and how much of the wait remains without the profiler was not measured.	5.79 µs median per call; mean 23.05 µs, max 5.9 ms. Message of 8 KiB per call (derived from the public config, not measured); the long waits sit in the first two calls of each step	1 trace, no repetition (spread across calls: p90 7.0 µs, mean 23.05 µs)	5,642 calls	2026-08-28	8×H100 80GB, TP8, torch.profiler	perfil-glm53-20260828	—	us: 1 trace, derived summary attached. outside: nobody
`C10a`	GLM-5.3-Flash, DeepSWE 111 tasks, our agent (K=3) Reproduced. Our agent, not the vendor's. The first repetition mixes two serving setups: 29 tasks at 20 workers without MTP and a 1,800 s timeout, 82 at 8 workers with MTP k=3 and 3,600 s. 111 of the 113 tasks: the gold patch fails our harness on 2.	65.32 % 67/106 · 70/108 · 72/106; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 63.72 % (2·se 2.65 pp) 63.2 · 64.8 · 67.9	2.40 (sd across repetitions; 2·se 2.77 pp)	111 tasks × K=3	2026-09-09	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	—	us: 3 repetitions, file attached. outside: nobody
`C10b`	GLM-5.3-Flash, DeepSWE 111 tasks, v1lite scaffold variant (K=6) The scaffold did not move the result; the recorded value came from rounded percentages.	64.21 % 81/108 · 68/105 · 71/109 · 66/106 · 62/105 · 62/105; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 62.80 % (2·se 5.41 pp); delta against the baseline −1.11 pp (2·se 5.56) 75.0 · 64.8 · 65.1 · 62.3 · 59.0 · 59.0	5.91 (sd across repetitions; 2·se 4.83 pp)	111 tasks × K=6	2026-09-21	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	2026-09-21 64.18 ± 4.85 retracted A mean of rounded percentages; two repetitions were written down as 58.95 % instead of 59.05 % (62/105).	us: 6 repetitions, file attached. outside: nobody
`C11`	GLM-5.3-Flash, MMLU-Pro with the official extractor (TIGER-AI-Lab) With the official extractor, the same answers give 82.96 %; the runner's old grader gave 72.59 %.	82.96 % 9,981/12,031 with the official MMLU-Pro extractor; the runner's old grader gave 8,733/12,031 = 72.59 %. 122 unfinished answers count as wrong. Also reading \boxed: 83.48 %. Our own 0-shot prompt, without the official harness format instruction	0.34 (binomial standard error; 2·se 0.69 pp)	12,032 items (K=1); 12,031 in the denominator	2026-09-23	8×H100 80GB, TP8; our runner	mmlu-pro-glm53-20260923	2026-09-21 72.59 % retracted Produced by the runner's old grader: it does not match \boxed{\text{G. …}} and it ignores case ('The answer is exactly…' becomes E). The official extractor on the same answers gives 82.96 %.	us: 1 run, file attached. outside: nobody
`C6`	SWE-bench gold-sanity: the gold patch through the same generate→apply→test path 0/25 resolved, 92% apply, n=25. In 10 of the 25 the runner could not collect the test IDs (8 Django, 2 SymPy); 13 never ran a test in the host environment; 2 IDs are not SWE-bench instances. File attached.	0/25 resolved, 92% apply, n=25 before the judge and classifier fix: test_fail 23, infra_error 2; after (swe_gold2.json): test_fail 4, infra_error 21	95% CI 0.0–13.3 % on the resolved rate (from the file; the 2 infra errors are in the denominator)	25	2026-08-04	not stated in a file	swe_gold_sanity_20260804	2026-08-04, before the judge and classifier fix 0/25 resolved, 92% apply; test_fail 23, infra_error 2 current; file attached Gold patch on the same code path. 10 test IDs not collected (8 Django, 2 SymPy), 13 environment failures, 2 IDs that are not SWE-bench instances (the 8% apply gap). File: swe_gold_sanity.json. 2026-08-04, after the judge and classifier fix 0/25 resolved, 92% apply; infra_error 21, test_fail 4 reclassification; file attached Same n=25, same apply 0.92. File: swe_gold2.json.	us: 1 gold-sanity run, file attached. outside: nobody
`C7`	GLM-5.3-Flash, GPQA Diamond (K=3 seeds) With the corrected grader (only the final line «Answer: X» counts; an answer cut off at the limit is wrong), the same answers give 89.85 %; the record said 90.19 %. The denominator is 197 per seed: one question did not fit the 65,536-token window on any seed (our client estimated its prompt at 1,970 tokens; it is about 2,836).	89.85 % 176/197 · 178/197 · 177/197; only the letter on the final line «Answer: X» counts (the old grader credited 2 answers cut off at the limit and reported 90.19 %); counting as wrong the question that did not fit the context window: 89.39 %. Our runner, T=1.0, thinking on 89.3 · 90.4 · 89.8	0.51 (sd across seeds; 2·se 0.59 pp)	198 × 3 seeds (valid per seed: 197, 197, 197)	2026-08-27	8×H100 80GB, TP8; our runner, 0-shot	gpqa-glm53-20260827-1653	2026-08-27 90.19 % retracted The old grader credited 2 answers cut off at the output ceiling (the same question in seeds 42 and 44, both at 64,629 tokens, the full output budget our runner set for that prompt) that never reached the final line «Answer: X»: one letter came from an «Answer:» written inside the reasoning, the other from a fallback that reads a conclusion in prose. Counting a cut-off answer as wrong, the same answers give 89.85 %.	us: 3 seeds, file attached. outside: nobody
`C8`	GLM-5.3-Flash, MATH-500 (K=3 seeds), corrected answer normalizer The old number measured the normalizer, not the model. The current one reapplies the corrected normalizer to the same answers.	99.00 % 1,485/1,500 with the new normalizer; the old one gave 76.20 %. 148 distinct pairs reclassified (342 items), read by hand; 15 failures left: 7 truncated, 8 wrong 99.4 · 98.2 · 99.4	0.69 (sd across seeds; 2·se 0.80 pp)	500 × 3 seeds	2026-09-19	8×H100 80GB, TP8; our runner	math500-glm53-20260919	2026-09-19 76.20 % retracted The normalizer marked equivalent answers wrong (x = 4 against 4; \dfrac against \frac).	us: 3 seeds, file attached. outside: nobody
`C9a`	GLM-5.3-Flash, HLE text only, multiple choice (K=1, no judge) Denominator corrected; with no judge, this is not the official score.	35.98 % 163/453; no judge; answers that did not converge count as wrong	2.25 (binomial standard error; 2·se 4.51 pp)	453 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 36.06 % (163/452) retracted The denominator dropped the multiple-choice item that had not converged even at 163,840 tokens.	us: 1 run, file attached. outside: nobody
`C9b`	GLM-5.3-Flash, HLE text only, strict exact match (K=1, no judge) No judge, and without the numeric tolerance that credited different answers: a floor.	8.75 % 127/1,451 strict: 131 by the file's count, 4 only by the numeric tolerance. No judge: a floor, not comparable to the official score	0.74 (binomial standard error; 2·se 1.48 pp)	1,451 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 9.31 % (131/1,407) retracted The denominator dropped 44 items that had not converged even at 163,840 tokens, and the relative tolerance credited 4 answers that differed from the key.	us: 1 run, file attached. outside: nobody
`A1`	Apparent gain of windowed attention after the causal-mask fix The cautionary tale survives; «all 28 causal configurations» does not	no archived raw file	—	—	2026-07-30	8×H100 (claim conditions)	no archived raw file	before the fix −4.1368 retracted as a real gain The gain was future information. after the fix factorial main effect −0.1523; isolated −0.042 (1 seed); no archived raw file later reading; no archived raw file Same design, physical held-out. The factorial main effect is not the isolated effect. No file published on this site.	us: 0 repetition files — no archived raw file. outside: nobody
`A2`	Isolated effect of the weight-decay ramp (0.01 → 0.10) Null on the sub-Chinchilla horizon; 64% of the composite gain was the baseline learning rate	no archived raw file	—	—	2026-07-31	8×H100, 220M, 1,200 steps (claim conditions)	no archived raw file	original catalog gain attributed to the composite scheduler retracted as the mechanism The off arm used a 500-step warmup (42% of training) and a mean learning rate 2.82× lower. paired decomposition, n=7 WD ramp Δ = +0.0048 (below δ_det = 0.0277) later reading; no archived raw file Null effect of the ramp on this horizon (~0.9 tok/param). It does not generalize to over-training.	us: 0 repetition files — n=7 arms claimed; no archived raw file. outside: nobody
`A4`	Training-lab spread (identical runs, only the seed changes) Held until the 8-seed file is archived.	no archived raw file	—	8 (claim; no archived raw file)	2026-07-31	8×H100 (claim conditions; no archived raw file)	no archived raw file	—	us: 0 repetition files — n=8 claimed; the 8-seed file is absent. outside: nobody
`A5`	FP8 eager versus BF16 training throughput (full step) SURVIVES as PyTorch eager without fusion; not as a law of the format	no archived raw file	—	—	2026-07	H100 SXM 80GB (claim conditions)	no archived raw file	bench with w.detach() w8 = 1.51× retracted autograd skipped dW: 6MKN → 4MKN, theoretical ratio 1.50×. The measured 1.51× was the bug. full step without detach 0.89× to 0.94× vs BF16 later reading; no archived raw file Two HBM transposes per linear in the backward of torch._scaled_mm. The sweep JSON is not published on this site.	us: 0 repetition files — 8 configurations from 350M to 3B claimed; no archived raw file. outside: nobody
`A6`	maj@5 versus K=1 on the same server (GPQA Diamond, temp 0.7) Paired contrast on one server; do not mix in a baseline from another instance	no archived raw file	—	198 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — n=198, K=5 claimed; JSON absent. outside: nobody
`B5`	HBM measured on the H100 roofline (phase 0) Plausible against the datasheet; the Phase 0 raw JSON is not published on this site	no archived raw file	—	—	—	H100 SXM (claim conditions)	no archived raw file	same internal record GEMM peak 781 TFLOPS vs 761 TFLOPS two anchors, no JSON The internal record disagrees with itself. Without the phase-0 JSON neither figure is a current value.	us: 0 repetition files — no archived raw file. outside: nobody
`B7`	Time to first audio, voice stack ~150 ms does not survive. 531.6 ms sums clocks that do not measure time to first audio end to end.	no archived raw file	—	20 (claim; no archived raw file)	—	not stated in a file (absent)	no archived raw file	later decomposition ~150 ms retracted It is not in the meter, and it is not the same instrument as the voice-window p50. voice window p50 531.6 ms retracted as time to first audio The 531.6 ms sums three clocks that do not measure time to first audio end to end: the STT clock includes the handshake, thinking was left on, and the TTS figure is a control frame. Not an end-to-end first-audio latency. No file published on this site.	us: 0 repetition files — n=20 claimed for the p50; JSON absent. outside: nobody
`C1`	Document-quality filter versus random Held: the held-out remeasurement has no archived raw file and is being repeated. The in-sample −64.5% stays retracted.	no archived raw file	—	1 (claim; no archived raw file)	2026-07-30	8×H100, ~300M, 1,500 steps (claim conditions)	no archived raw file	eval on the training set −64.5% PPL retracted The metric was in-sample.	us: 0 repetition files — n=1 seed; no archived raw file. outside: nobody
`C3`	Thinking versus chat on GPQA (thinking recipe versus greedy) «zero pp as proof of zero» does not survive; the detection floor is ~5 pp at n=40 K=1	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	original formulation zero pp retracted as proof of a null effect n=40 K=1; the pre-registered kill (+3 pp) is below the floor (~5 pp). It does not reject +8 pp. audit formulation no effect ≥ ~5 pp detected revised reading; JSON absent Same probe. Confound: thinking is temp 1.0 / top_p 0.95, chat is greedy temp 0.	us: 0 repetition files — n=40 K=1; no archived raw file. outside: nobody
`C4`	Spread of GPQA accuracy at temp 0, same server, fixed seed The phenomenon (temp 0 is not deterministic) survives; a single cause does not	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — K=5, n=40 claimed; JSON absent. outside: nobody
`C5a`	Terminal-Bench 0/20 (first score) VOID: the adapter called a host path inside the container (exit 127)	no archived raw file	—	—	2026-08	not stated in a file (absent)	no archived raw file	original score 0/20 retracted as model capability bash <host path> inside the container → exit 127 → «Verification failed» every time.	us: 0 repetition files — reproducible as an adapter bug in the old code; the 0/20 JSON is absent. outside: nobody
`C5b`	Terminal-Bench gold46, DeepSeek V4-Flash, after the adapter fix VOID: measured before the gold gate existed; a floor, not a measurement. DeepSeek V4-Flash, text scaffold. Retracted.	no archived raw file	—	44 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	2026-08-11 8/44 = 18.2% retracted Measured before the gold gate existed. A floor, not a measurement. No file published on this site.	us: 0 repetition files — n=44 claimed; JSON absent. outside: nobody

measurements by id
id	what was measured	value	uncertainty	n	date	hardware	raw file	history	replication
`A1`	Apparent gain of windowed attention after the causal-mask fix The cautionary tale survives; «all 28 causal configurations» does not	no archived raw file	—	—	2026-07-30	8×H100 (claim conditions)	no archived raw file	before the fix −4.1368 retracted as a real gain The gain was future information. after the fix factorial main effect −0.1523; isolated −0.042 (1 seed); no archived raw file later reading; no archived raw file Same design, physical held-out. The factorial main effect is not the isolated effect. No file published on this site.	us: 0 repetition files — no archived raw file. outside: nobody
`A2`	Isolated effect of the weight-decay ramp (0.01 → 0.10) Null on the sub-Chinchilla horizon; 64% of the composite gain was the baseline learning rate	no archived raw file	—	—	2026-07-31	8×H100, 220M, 1,200 steps (claim conditions)	no archived raw file	original catalog gain attributed to the composite scheduler retracted as the mechanism The off arm used a 500-step warmup (42% of training) and a mean learning rate 2.82× lower. paired decomposition, n=7 WD ramp Δ = +0.0048 (below δ_det = 0.0277) later reading; no archived raw file Null effect of the ramp on this horizon (~0.9 tok/param). It does not generalize to over-training.	us: 0 repetition files — n=7 arms claimed; no archived raw file. outside: nobody
`A4`	Training-lab spread (identical runs, only the seed changes) Held until the 8-seed file is archived.	no archived raw file	—	8 (claim; no archived raw file)	2026-07-31	8×H100 (claim conditions; no archived raw file)	no archived raw file	—	us: 0 repetition files — n=8 claimed; the 8-seed file is absent. outside: nobody
`A5`	FP8 eager versus BF16 training throughput (full step) SURVIVES as PyTorch eager without fusion; not as a law of the format	no archived raw file	—	—	2026-07	H100 SXM 80GB (claim conditions)	no archived raw file	bench with w.detach() w8 = 1.51× retracted autograd skipped dW: 6MKN → 4MKN, theoretical ratio 1.50×. The measured 1.51× was the bug. full step without detach 0.89× to 0.94× vs BF16 later reading; no archived raw file Two HBM transposes per linear in the backward of torch._scaled_mm. The sweep JSON is not published on this site.	us: 0 repetition files — 8 configurations from 350M to 3B claimed; no archived raw file. outside: nobody
`A6`	maj@5 versus K=1 on the same server (GPQA Diamond, temp 0.7) Paired contrast on one server; do not mix in a baseline from another instance	no archived raw file	—	198 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — n=198, K=5 claimed; JSON absent. outside: nobody
`B1`	Single-stream decode speedup The first factor does not survive; the order of magnitude does. The 25 Aug remeasurement is the current value, file attached. Measured on a synthetic filler prompt; speculative acceptance depends on the text, so this is not a real-workload speedup.	9.78× 183.6 → 1,795.1 tok/s 183.6 · 1,795.1	0.01 (sd across repetitions, propagated; 0.1 and 1.4 tok/s on the arms)	3 per arm	2026-08-25	8×H100 80GB (TP8, no speculation; DFlash b64 on TP2). Two knobs. synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap.	single10x-20260825-0035	2026-08-21 10.1× retracted One 90-second reading of TP2 + DFlash-64 (1,782 tok/s, 256-token prompts) divided by 176 tok/s, a TP8 baseline measured with 8,192-token prompts; the comparable 256-token baseline read 187 tok/s, which gives 9.5×. Those window files are not published on this site. 2026-08-25, window 0011 200.7 tok/s (TP8 baseline; n=3, sd 2.0) retracted in favor of 0035 First TP8 baseline reading. File: brutos/single10x-20260825-0011/resumo.json. The scheduled job was not suspended; the DFlash arm was not measured (content gate). Why 200.7 became 183.6 is not shown. 2026-08-25, window 0035 9.78×. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b current; file attached TP8, no speculation, 183.6 tok/s → TP2 + DFlash-64, 1,795.1 tok/s; conc=1, n=3. Two knobs moved. TP2 alone = 0.61× of TP8; speculation alone = 15.5×, see B1b.	us: 3 repetitions per arm, file attached. outside: nobody
`B10`	GLM-5.3-Flash, agentic (8 DeepSWE tasks, 8 workers): per-task duration without MTP divided by MTP k=3 The paired ratio survives; «+40% tasks/h» does not — it measured a single task.	1.52× geometric mean; 95% CI 1.31–1.76 (t on the logs). Resolved 7/8 without MTP and 6/8 with MTP: quality is not comparable at this n 1.9 · 1.6 · 1.9 · 1.4 · 1.1 · 1.4 · 1.8 · 1.4	95% CI 1.31–1.76 (t on the logs)	8 paired tasks, 1 round per arm	2026-09-04	8×H100 80GB, TP8; arms on different days (3 Sep and 4 Sep)	agentico-glm53-mtp-20260903	2026-09-04 13.2 → 18.5 tasks/h (+40%) retracted With 8 tasks on 8 workers the wall clock is the slowest task, and it is the same task on both arms; the pair measures one task.	us: 1 round per arm, file attached. outside: nobody
`B1b`	Speculation alone, paired TP2 The second knob (TP2 instead of TP8) lowers the baseline; speculation alone, same session and same GPUs, is the current factor.	15.5× 112.7 → 1,746.5 tok/s 112.7 · 1,746.5	0.01 (sd across repetitions, propagated; 0.0 and 1.3 tok/s on the arms)	3 per arm	2026-08-26	2×H100 (TP2), shared node; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	partp2-20260826-1950	2026-08-26, paired window 1950 15.5× current; file attached TP2 with no speculation 112.7 tok/s → TP2 + DFlash-64 1,746.5 tok/s; same session, same GPUs, n=3. The 9.78× against TP8 (B1) is the conservative comparison.	us: 3 repetitions per arm, file attached. outside: nobody
`B2`	Aggregate throughput of four replicas on the same node Reproduced. The ruler does not knock a number down on reflex: here the remeasurement confirmed it. File attached.	29,503 tok/s 29,467.0 · 29,421.6 · 29,620.4	104.2 (sd across repetitions)	3 (overlap 118 s)	2026-08-25	8×H100 80GB (four replicas, conc=32 per replica); synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	composto4xc32-20260825-0244	2026-08-21 29,567 tok/s first reading; JSON absent Aggregate of four replicas. Audit B checked the sum and marked simultaneity as not archived. 2026-08-25 29,503 tok/s reproduced; current; file attached conc=32 per replica, n=3, overlap 118 s, aggregate 29,503 ± 104 tok/s. The file has no interference percentage.	us: 3 repetitions, file attached. outside: nobody
`B3`	GLM-5.3-Flash, 8×H100, single-stream without and with MTP k=5 First reading, file attached. The current value comes from resumo.json.	1.76× 155.2 → 272.6 tok/s 155.2 · 272.6	0.05 (sd across repetitions, propagated; 0.0 and 7.1 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions per arm, file attached. outside: nobody
`B3b`	GLM-5.3-Flash, composite conc 32 on one TP8 replica Composite throughput at 32 streams on one TP8 replica. File attached. The current value comes from resumo.json.	2,661.9 tok/s 2,633.3 · 2,715.7 · 2,636.7	46.6 (sd across repetitions)	3	2026-08-27	8×H100 TP8, native FP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-glm53-20260827-1218	—	us: 3 repetitions, file attached. outside: nobody
`B4`	Qwen3.8-Flash-Next, 8×H100, single-stream without and with MTP k=3 File attached; the current value comes from resumo.json.	2.61× 136.0 → 354.6 tok/s 136.0 · 354.6	0.06 (sd across repetitions, propagated; 3.0 and 0.6 tok/s on the arms)	3 per arm	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B4b`	Qwen3.8-Flash-Next, composite conc 32 on one TP8 replica File attached; the current value comes from resumo.json.	2,944.4 tok/s 2,891.4 · 2,967.1 · 2,974.7	46.1 (sd across repetitions)	3	2026-08-27	8×H100 TP8, dedicated vLLM image; synthetic filler prompt (a marker word plus one word repeated to 256 tokens), greedy decoding (temperature 0), 512-token output cap	flash-qwen38fn-20260827-1357	—	us: 3 repetitions per arm, file attached. outside: nobody
`B5`	HBM measured on the H100 roofline (phase 0) Plausible against the datasheet; the Phase 0 raw JSON is not published on this site	no archived raw file	—	—	—	H100 SXM (claim conditions)	no archived raw file	same internal record GEMM peak 781 TFLOPS vs 761 TFLOPS two anchors, no JSON The internal record disagrees with itself. Without the phase-0 JSON neither figure is a current value.	us: 0 repetition files — no archived raw file. outside: nobody
`B7`	Time to first audio, voice stack ~150 ms does not survive. 531.6 ms sums clocks that do not measure time to first audio end to end.	no archived raw file	—	20 (claim; no archived raw file)	—	not stated in a file (absent)	no archived raw file	later decomposition ~150 ms retracted It is not in the meter, and it is not the same instrument as the voice-window p50. voice window p50 531.6 ms retracted as time to first audio The 531.6 ms sums three clocks that do not measure time to first audio end to end: the STT clock includes the handshake, thinking was left on, and the TTS figure is a control frame. Not an end-to-end first-audio latency. No file published on this site.	us: 0 repetition files — n=20 claimed for the p50; JSON absent. outside: nobody
`B8`	GLM-5.3-Flash, one stream: native MTP k=3 against no speculation (17 Sep window) Reproduced from the raw file. The current value comes from the per-request repetitions; tok/s includes TTFT. Synthetic filler prompt at temperature 0: speculative acceptance depends on the text; on real agent tasks MTP gives 1.52× per task (B10).	2.04× 156.05 ± 0.07 → 317.90 ± 2.92 tok/s (±2·se); repetitions between 2.02× and 2.05× 156.1 · 317.9	0.02 (sd across repetitions, propagated; 0.06 and 2.53 tok/s on the arms)	3 per arm	2026-09-17	8×H100 80GB, TP8, native FP8, dedicated vLLM image; 1 request at a time, synthetic 256-token filler prompt, greedy decoding (temperature 0), 512-token output cap, thinking off	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions per arm, file attached. outside: nobody
`B8b`	GLM-5.3-Flash, one stream, no speculation: nvidia-smi counters (memory busy over time) The old label «SM active» was wrong: utilization.gpu is the fraction of time any kernel is running.	21.80 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 98.74 % (fraction of time any kernel is running; this is not per-SM activity) · 256.9 W per GPU 21.9 · 21.8 · 21.8	0.06 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	2026-09-17 «SM active» 98.7 % retracted as a label The column is nvidia-smi utilization.gpu: the fraction of time any kernel is running, not per-SM activity. This node has no per-SM activity counter.	us: 3 repetitions, file attached. outside: nobody
`B8c`	GLM-5.3-Flash, one stream with MTP k=3: nvidia-smi counters (memory busy over time) The same step with MTP keeps memory busier and draws more power: more useful work per step.	26.95 % utilization.memory (fraction of time the memory is being read or written). utilization.gpu 99.16 % (fraction of time any kernel is running; this is not per-SM activity) · 308.1 W per GPU 26.8 · 26.8 · 27.2	0.20 (sd across repetitions)	3 repetitions × 8 GPUs at 1 Hz	2026-09-17	8×H100 80GB, TP8; nvidia-smi at 1 Hz on all 8 GPUs	telemetria-glm53-c1-20260917-2110	—	us: 3 repetitions, file attached. outside: nobody
`B9`	GLM-5.3-Flash, kernels per decode step (profiler trace, 1 request, TP8 rank 7) The old number does not come from the trace; the current one comes from the summary derived from the trace (the 347 MB trace is not published).	1,694 1,484 on the main stream + 210 on the second, the same on every step; median 2.56 µs per kernel; under the profiler a step takes 7.65 ms and the main stream is busy for 88.7 % of it (ratio of the medians)	1 trace, no repetition (the same count on every step)	62 steps between 63 annotations	2026-08-28	8×H100 80GB, TP8, torch.profiler; decode with no speculation, thinking off	perfil-glm53-20260828	2026-08-29 1,529 kernels per step, 54 steps retracted It was the trace's kernel total (109,211) divided by about 71 steps; the trace has 63 annotated steps.	us: 1 trace, derived summary attached. outside: nobody
`B9b`	GLM-5.3-Flash, all-reduce per call in the decode step (cross_device_reduce, TP8 rank 7) Median 5.79 µs; 89 of the 91 calls per step stay near it. The long waits are the first two calls of each step (medians about 1,014 and 242 µs), probably the 8 ranks aligning; measured under the profiler on one rank, and how much of the wait remains without the profiler was not measured.	5.79 µs median per call; mean 23.05 µs, max 5.9 ms. Message of 8 KiB per call (derived from the public config, not measured); the long waits sit in the first two calls of each step	1 trace, no repetition (spread across calls: p90 7.0 µs, mean 23.05 µs)	5,642 calls	2026-08-28	8×H100 80GB, TP8, torch.profiler	perfil-glm53-20260828	—	us: 1 trace, derived summary attached. outside: nobody
`C1`	Document-quality filter versus random Held: the held-out remeasurement has no archived raw file and is being repeated. The in-sample −64.5% stays retracted.	no archived raw file	—	1 (claim; no archived raw file)	2026-07-30	8×H100, ~300M, 1,500 steps (claim conditions)	no archived raw file	eval on the training set −64.5% PPL retracted The metric was in-sample.	us: 0 repetition files — n=1 seed; no archived raw file. outside: nobody
`C10a`	GLM-5.3-Flash, DeepSWE 111 tasks, our agent (K=3) Reproduced. Our agent, not the vendor's. The first repetition mixes two serving setups: 29 tasks at 20 workers without MTP and a 1,800 s timeout, 82 at 8 workers with MTP k=3 and 3,600 s. 111 of the 113 tasks: the gold patch fails our harness on 2.	65.32 % 67/106 · 70/108 · 72/106; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 63.72 % (2·se 2.65 pp) 63.2 · 64.8 · 67.9	2.40 (sd across repetitions; 2·se 2.77 pp)	111 tasks × K=3	2026-09-09	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	—	us: 3 repetitions, file attached. outside: nobody
`C10b`	GLM-5.3-Flash, DeepSWE 111 tasks, v1lite scaffold variant (K=6) The scaffold did not move the result; the recorded value came from rounded percentages.	64.21 % 81/108 · 68/105 · 71/109 · 66/106 · 62/105 · 62/105; our agent; independent repetitions at T=1.0; counting as failures the tasks whose agent history did not fit the 262,144-token context: 62.80 % (2·se 5.41 pp); delta against the baseline −1.11 pp (2·se 5.56) 75.0 · 64.8 · 65.1 · 62.3 · 59.0 · 59.0	5.91 (sd across repetitions; 2·se 4.83 pp)	111 tasks × K=6	2026-09-21	8×H100 80GB, TP8; our agent and runner	deepswe-glm53-20260908	2026-09-21 64.18 ± 4.85 retracted A mean of rounded percentages; two repetitions were written down as 58.95 % instead of 59.05 % (62/105).	us: 6 repetitions, file attached. outside: nobody
`C11`	GLM-5.3-Flash, MMLU-Pro with the official extractor (TIGER-AI-Lab) With the official extractor, the same answers give 82.96 %; the runner's old grader gave 72.59 %.	82.96 % 9,981/12,031 with the official MMLU-Pro extractor; the runner's old grader gave 8,733/12,031 = 72.59 %. 122 unfinished answers count as wrong. Also reading \boxed: 83.48 %. Our own 0-shot prompt, without the official harness format instruction	0.34 (binomial standard error; 2·se 0.69 pp)	12,032 items (K=1); 12,031 in the denominator	2026-09-23	8×H100 80GB, TP8; our runner	mmlu-pro-glm53-20260923	2026-09-21 72.59 % retracted Produced by the runner's old grader: it does not match \boxed{\text{G. …}} and it ignores case ('The answer is exactly…' becomes E). The official extractor on the same answers gives 82.96 %.	us: 1 run, file attached. outside: nobody
`C3`	Thinking versus chat on GPQA (thinking recipe versus greedy) «zero pp as proof of zero» does not survive; the detection floor is ~5 pp at n=40 K=1	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	original formulation zero pp retracted as proof of a null effect n=40 K=1; the pre-registered kill (+3 pp) is below the floor (~5 pp). It does not reject +8 pp. audit formulation no effect ≥ ~5 pp detected revised reading; JSON absent Same probe. Confound: thinking is temp 1.0 / top_p 0.95, chat is greedy temp 0.	us: 0 repetition files — n=40 K=1; no archived raw file. outside: nobody
`C4`	Spread of GPQA accuracy at temp 0, same server, fixed seed The phenomenon (temp 0 is not deterministic) survives; a single cause does not	no archived raw file	—	40 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	—	us: 0 repetition files — K=5, n=40 claimed; JSON absent. outside: nobody
`C5a`	Terminal-Bench 0/20 (first score) VOID: the adapter called a host path inside the container (exit 127)	no archived raw file	—	—	2026-08	not stated in a file (absent)	no archived raw file	original score 0/20 retracted as model capability bash <host path> inside the container → exit 127 → «Verification failed» every time.	us: 0 repetition files — reproducible as an adapter bug in the old code; the 0/20 JSON is absent. outside: nobody
`C5b`	Terminal-Bench gold46, DeepSeek V4-Flash, after the adapter fix VOID: measured before the gold gate existed; a floor, not a measurement. DeepSeek V4-Flash, text scaffold. Retracted.	no archived raw file	—	44 (claim; no archived raw file)	2026-08	not stated in a file (absent)	no archived raw file	2026-08-11 8/44 = 18.2% retracted Measured before the gold gate existed. A floor, not a measurement. No file published on this site.	us: 0 repetition files — n=44 claimed; JSON absent. outside: nobody
`C6`	SWE-bench gold-sanity: the gold patch through the same generate→apply→test path 0/25 resolved, 92% apply, n=25. In 10 of the 25 the runner could not collect the test IDs (8 Django, 2 SymPy); 13 never ran a test in the host environment; 2 IDs are not SWE-bench instances. File attached.	0/25 resolved, 92% apply, n=25 before the judge and classifier fix: test_fail 23, infra_error 2; after (swe_gold2.json): test_fail 4, infra_error 21	95% CI 0.0–13.3 % on the resolved rate (from the file; the 2 infra errors are in the denominator)	25	2026-08-04	not stated in a file	swe_gold_sanity_20260804	2026-08-04, before the judge and classifier fix 0/25 resolved, 92% apply; test_fail 23, infra_error 2 current; file attached Gold patch on the same code path. 10 test IDs not collected (8 Django, 2 SymPy), 13 environment failures, 2 IDs that are not SWE-bench instances (the 8% apply gap). File: swe_gold_sanity.json. 2026-08-04, after the judge and classifier fix 0/25 resolved, 92% apply; infra_error 21, test_fail 4 reclassification; file attached Same n=25, same apply 0.92. File: swe_gold2.json.	us: 1 gold-sanity run, file attached. outside: nobody
`C7`	GLM-5.3-Flash, GPQA Diamond (K=3 seeds) With the corrected grader (only the final line «Answer: X» counts; an answer cut off at the limit is wrong), the same answers give 89.85 %; the record said 90.19 %. The denominator is 197 per seed: one question did not fit the 65,536-token window on any seed (our client estimated its prompt at 1,970 tokens; it is about 2,836).	89.85 % 176/197 · 178/197 · 177/197; only the letter on the final line «Answer: X» counts (the old grader credited 2 answers cut off at the limit and reported 90.19 %); counting as wrong the question that did not fit the context window: 89.39 %. Our runner, T=1.0, thinking on 89.3 · 90.4 · 89.8	0.51 (sd across seeds; 2·se 0.59 pp)	198 × 3 seeds (valid per seed: 197, 197, 197)	2026-08-27	8×H100 80GB, TP8; our runner, 0-shot	gpqa-glm53-20260827-1653	2026-08-27 90.19 % retracted The old grader credited 2 answers cut off at the output ceiling (the same question in seeds 42 and 44, both at 64,629 tokens, the full output budget our runner set for that prompt) that never reached the final line «Answer: X»: one letter came from an «Answer:» written inside the reasoning, the other from a fallback that reads a conclusion in prose. Counting a cut-off answer as wrong, the same answers give 89.85 %.	us: 3 seeds, file attached. outside: nobody
`C8`	GLM-5.3-Flash, MATH-500 (K=3 seeds), corrected answer normalizer The old number measured the normalizer, not the model. The current one reapplies the corrected normalizer to the same answers.	99.00 % 1,485/1,500 with the new normalizer; the old one gave 76.20 %. 148 distinct pairs reclassified (342 items), read by hand; 15 failures left: 7 truncated, 8 wrong 99.4 · 98.2 · 99.4	0.69 (sd across seeds; 2·se 0.80 pp)	500 × 3 seeds	2026-09-19	8×H100 80GB, TP8; our runner	math500-glm53-20260919	2026-09-19 76.20 % retracted The normalizer marked equivalent answers wrong (x = 4 against 4; \dfrac against \frac).	us: 3 seeds, file attached. outside: nobody
`C9a`	GLM-5.3-Flash, HLE text only, multiple choice (K=1, no judge) Denominator corrected; with no judge, this is not the official score.	35.98 % 163/453; no judge; answers that did not converge count as wrong	2.25 (binomial standard error; 2·se 4.51 pp)	453 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 36.06 % (163/452) retracted The denominator dropped the multiple-choice item that had not converged even at 163,840 tokens.	us: 1 run, file attached. outside: nobody
`C9b`	GLM-5.3-Flash, HLE text only, strict exact match (K=1, no judge) No judge, and without the numeric tolerance that credited different answers: a floor.	8.75 % 127/1,451 strict: 131 by the file's count, 4 only by the numeric tolerance. No judge: a floor, not comparable to the official score	0.74 (binomial standard error; 2·se 1.48 pp)	1,451 (K=1)	2026-09-23	8×H100 80GB, TP8; our runner	hle-glm53-20260923	2026-09-23 9.31 % (131/1,407) retracted The denominator dropped 44 items that had not converged even at 163,840 tokens, and the relative tolerance credited 4 answers that differed from the key.	us: 1 run, file attached. outside: nobody

§ — raw files published on this site

Each directory is the package the page read. The link goes to the package README, with a sha256 per file.

§3 — replication

Each row says how many times we repeated it and whether anyone outside did. In this index, anyone outside = nobody. That is written down. It is not an omission.

§4 — what stays out

A number the audit knocked down, even with a caveat.
A number typed by hand as a current value.
A production identifier, a hostname, a disk path, and whatever the audit knocked down.

Verdict sources: audits A, B and C (2026-08-24), and the 2026-09-25 revision. The page reads the file or writes the absence sentence. History entries of rows without a file are copied from internal records and cannot be checked on this site.

Changelog — 2026-09-25, rev 2

The page is in English. Rows say published on this site.
The uncertainty column names the kind in each cell: standard deviation across seeds, standard deviation across repetitions, or binomial standard error.
C5b, Terminal-Bench gold46 on DeepSeek V4-Flash, is VOID and retracted: measured before the gold gate existed; a floor, not a measurement.
B4 and B4b point at the package published on this site, dated 2026-08-27. The current value comes from its resumo.json.
A1 shows the isolated effect beside the factorial main effect. No archived raw file.
B7 notes that 531.6 ms sums clocks that do not measure time to first audio end to end.