Single entry point for every speed claim about msuiter. This document contains no independently retyped numbers: each value is a verbatim, path:line-pinned quotation of a committed bench artifact, extracted by the commands listed at the bottom in the same command window that wrote this file, and sealed by the digests in the final section. To audit any claim, re-run the extraction — do not trust the rendering.
Measured speedups (authoritative)
Source: bench/m1s/result.md — the M1s acceptance-gate bench summary (bench/m1s/result.md:1):
bench/m1s · the M1s acceptance-gate bench summary (the U-M1s-13 companion; the pinned-hardware caliber, not into the CI)
The three measured kernels, quoted exactly from the summary table:
| workload | measured speedup | original target | verdict (vs original) |
|---|---|---|---|
| catalog generation, SBS96 @ 10⁶ mutations | 2.6× |
≥50× | speedup FAIL; <5s PASS |
| KL-NMF, 96×1000, k=10, 100 iterations | 2.3× |
≥100× | FAIL |
| batch NNLS, 101×10³, k=10 | 1.8× |
≥10× | FAIL |
Baselines (from the same cells): SPMG for catalog generation, the C-kernel R package brunet for KL-NMF, and scipy per-sample linear algebra for batch NNLS — see the verbatim rows directly below.
Verbatim table rows (bench/m1s/result.md:11, 13, 14):
| the catalog generation SBS96 (the 10⁶ mutations) | ≥50× SPMG and <5s single-core | 2.6× (ours 3.73s vs SPMG 9.82s); <5s ✓ | **speedup FAIL; <5s PASS** |
| the NMF KL (96×1000, k=10, 100 iter) | ≥100× R NMF | 2.3× (ours 205ms vs brunet 0.47s) | **FAIL** |
| the batch NNLS (101×10³, k=10) | ≥10× the per-sample scipy | 1.8× (ours 6.7ms vs scipy 11.2ms) | **FAIL** |And the summary’s own adjudication of the pre-recalibration targets (bench/m1s/result.md:17):
The ROADMAP M1s acceptance row “catalog generation ≥50× SPMG; NMF ≥100× R NMF” does not hold under this bench —
Framing: the Plan-A recalibration (PI ruling 2026-10-05)
The original ≥50×/≥100×/≥10× rhetoric was written against baselines that were never themselves measured — brunet already ships a C kernel, and SPMG on clean input is vectorised. The PI ruling of 2026-10-05 recalibrated acceptance to the MEASURED C-class baselines, and moved the original rhetoric to a naive-R/Python framing (vignettes/articles/roadmap.qmd:118, verbatim; the blockquote spans vignettes/articles/roadmap.qmd:117-118):
[English twin of the retired ruling original; CJK corpus] phase-2, issue #25, 2026-10-19; the Chinese original lives in git blob sha256 8b905eb8151cde0c75dee4b3d3b61f5efb72cffa0e38729b6119d74615718a91). Acceptance (v0.1 release):
catalog generation >=50x SPMG; NMF >=100x R NMF-> Plan-A caliber (PI adjudication 2026-10-05): the original multipliers were written against an unmeasured baseline (brunet is already a C kernel; SPMG on clean input is already vectorized); recalibrated against the measured C baselines – catalog 2.6x SPMG (3.73s, <5s OK), NMF 2.3x R NMF brunet (C core), NNLS 1.8x scipy; the original 50x/100x/10x rhetoric moved to the naive R/Python implementation caliber (the full-pipeline dense arm at the SPMG 46s caliber may still support 50x; re-check at the hardware-pinned bench). All fixtures green; installable from r-universe. M1s 14/14 closed.
Targets table (ARCHITECTURE cross-reference)
The canonical class targets, verbatim from vignettes/articles/architecture.qmd:316-323:
| Kernel | Baseline | Target |
|---|---|---|
| the catalog generation SBS96 (10⁶ mutations) | SPMG | ≥50× and <5 s single core |
| NMF KL (96×1000, k=10, 100 iter) | R NMF / rust-NMF | ≥100× R NMF; the rust-NMF alignment (tolerance) |
| the batch NNLS (101×10³) | scipy nnls per sample | ≥10× |
| the all-sample fitting | SPA ~0.21 s/sample | aligned or better |
| bootstrap (10³×10³) | sigminer serial | ≥100× |MEASURED: bootstrap vs sigminer – the >=100x target measured FALSE
The targets table’s final row (vignettes/articles/architecture.qmd:323) sets the bootstrap benchmark against sigminer’s serial bootstrapper at >=100x:
| bootstrap (10³×10³) | sigminer serial | ≥100× |Status: MEASURED – the >=100x row is false as a claim and must never be quoted as achieved. The RB-10 authoritative single-shot campaign (bench/results/bootstrap_vs_sigminer_20261013T175859Z/result.md, committed read-only in this PR) adjudicated the row on the pinned campaign host:
| quantity | measured value |
|---|---|
| tier | B |
| ratio (A1 bootstrap / A2 sigminer serial) | 1.536600758177e+01 (15.37x) |
| verdict | MEASURED-false 15.37x |
| secondary A1/A3 (pool-scaling isolation) | 14.2209x |
Evidence chain, all committed inside the outdir: seed 20260928 (G0 master; shape K=1000 S=1000 n_sig=50); the data seal {"sha256": "cc094225d599a6765dbc7958a445609459ded0f32f2481f87c5cb9e8762b3263", ...} with generator gen_campaign.R argv recorded verbatim in seal.json; campaign.rds re-hashes to the sealed digest (byte-idempotent derivation attested: the 2026-10-11 rehearsal outdir carries the identical rds digest from three days earlier under the frozen seed and argv); one probe launch and exactly one full launch in campaign_launches.jsonl (the RB-10 single-shot law – no refiring); co-tenancy attestation in witnesses/host_census.jsonl plus the wrapper’s load1 census (all samples strictly below the 1.5 bar, bar never relaxed); the CI-noise-envelope anchor gate adjudicated the paired anchor witnesses green (median_ok_frac 1.0 over 1250 cells).
What “MEASURED-false” means precisely: the campaign’s oracle compares the timed ratio against the >=100x table claim and returns false – the claim is refuted by measurement at this shape on this pinned host, roughly one order of magnitude short. The row above is kept because ARCHITECTURE is the target registry, not an achievements registry; this document is now the place that records what the target actually measured.
Pinned-host caveat
Every number above is a single-session measurement on one pinned, developer machine whose power state was not controlled — the source attestation (bench/m1s/result.md:3-4):
Date: 2026-09-30 · the local Apple M5 / macOS 26.6.2 / R 4.5.2 / rustc 1.92 / SPMG 1.3.6 (bench/m1s/.venv). The power status not controllable — the numbers are for this pinned-hardware single-session caliber;
The per-part benches repeat the caveat individually – the three reference reports below stay byte-frozen CJK originals (translation only on user demand, charter class “reference”), so each line is a labeled English rendering, not a verbatim quote:
-
bench/m1s/sbs/result.md:18(rendering of the CJK original) > - host: aarch64-apple-darwin20 (Apple Silicon; the power status not controllable – the numbers are for this pinned-hardware caliber) -
bench/m1s/nmf/result.md:18(rendering of the CJK original) > - host: aarch64-apple-darwin20 (Apple Silicon; the power status not controllable – the numbers are for this pinned-hardware caliber) -
bench/m1s/nnls/result.md:12(rendering of the CJK original) > - python 3.13.9 | numpy 2.3.5 | scipy 1.17.0 | darwin (Apple Silicon; the power status not controllable – the numbers are for this pinned-hardware caliber)
Consequences, all by reference: these values are acceptance-gate evidence, not CI-tracked series (the ‘do not enter CI’ stipulation of the title line quoted above); the Plan-A paragraph itself defers the naive-baseline 50× re-check to a pinned-hardware bench (the re-verification at the pinned-hardware bench, quoted above); and from U-M7-04 B1a onward the M7 harness writer stamps host and hw_profile into every timings row so future machine attestation is checkable from the table alone:
$ git log --oneline -1 -- bench/m7/competitors/timings.py
bc3e8ca feat(m7): U-M7-04 M7 closeout — governance truth, hardware-pin machine-checkability, performance anchoring, release material (#14)Re-extraction commands (same window as this file’s creation)
$ sed -n '11p;13p;14p' bench/m1s/result.md
$ grep -n 'Plan-A calibration' vignettes/articles/roadmap.qmd # -> line 107
$ sed -n '316,323p' vignettes/articles/architecture.qmd # targets table incl. bootstrap row 323
$ grep -n 'power status not controllable' bench/m1s/result.md; grep -n 'Apple Silicon' bench/m1s/*/result.md # parent (translated) + the three CJK reference reportsIntegrity seals
SHA-256 of every source quoted above, hashlib/shasum pair-checked in-window; the abc control vector hashed to the same value under hashlib, openssl dgst -sha256 -r, node crypto.createHash and Apple CommonCrypto shasum -a 256:
1886e3937a297a88ebb29fe9f627ce0c12ec94ad127d0bfce22b96dd9f73a7a0 bench/m1s/result.md
b85fd4d2cda15805e3693f6942c31d8bdf9e917d2133153ac5a39c62e9f69320 vignettes/articles/architecture.qmd
51ce5888ab202dfc90de7ed9f853fe98fea0e7d67b78ed24f626186a9fb5500f vignettes/articles/roadmap.qmd
832f0a474752479b7017979d31620f26cd6c2b6aecfecddbce8dc6363cc462e0 bench/m1s/sbs/result.md
9e95633451308d8198b3b57cdb0a4b890da491b8c3922612b125e3d86dd74a7b bench/m1s/nmf/result.md
275ba5541d276cf7ca31bb70a44a6fba8d1657a334b439a9ece37e533389cc91 bench/m1s/nnls/result.md
ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad (control: sha256 of 'abc')