Performance anchoring: frozen floors, witnesses and reproduction
Source:vignettes/performance-anchoring.Rmd
performance-anchoring.Rmdmsuiter’s benchmark layer is governed, not narrated. Floor thresholds
are frozen constants re-derived from rep-100 witness bytes (never from
short-probe magnitudes), every calibration run mints a witnessed
registry under bench/results/, and the two sealed
head-to-head campaigns are quoted verbatim below with their witness
paths. Nothing on this page retypes a number by hand: each figure is
routed from committed bytes at build time, and the extraction chunks
fail closed if a committed artefact ever goes missing or empty.
The frozen floors and their witnessing registry
The floor table’s epoch hashes the canonical JSON of every
floor-affecting constant (FLOOR_TABLE_FIELDS), so any
re-calibration flips it by construction; the calibration registry pins
the witness digests the floors were sealed from.
if (is.na(PKG)) {
cat("rendered outside the package root: the governance artefacts are not reachable here\n")
} else {
cal_f <- file.path(PKG, "bench", "results",
"m7_floor_calibration_p7cap_legacy_20261027-083955", "CALIBRATION.json")
gov_f <- file.path(PKG, "bench", "m7", "tools", "matrix_governance.py")
anc_f <- file.path(PKG, "bench", "results",
"m7_floor_calibration_p7cap_legacy_20261027-083955", "anchors_fragment.json")
if (!all(file.exists(c(cal_f, gov_f, anc_f)))) {
cat("governance artefacts not on disk (check tarball): read them in the repository under bench/results/m7_floor_calibration_p7cap_legacy_20261027-083955/ and bench/m7/tools/matrix_governance.py\n")
} else {
cat("---- ", basename(cal_f), " ----\n", sep = "")
cal <- readLines(cal_f, warn = FALSE)
.say(.must(.grep_fixed(cal, '"schema"'), "calibration schema"))
.say(.must(.grep_fixed(cal, '"probes"'), "calibration probe count"))
.say(.must(.grep_fixed(cal, '"freshness"'), "freshness law"))
.say(.must(.grep_fixed(cal, '"gate"'), "gate law"))
.say(.must(.grep_fixed(cal, '"reps"'), "reps law"))
dig <- .must(.grep_fixed(cal, '.json": "'), "witness digests")
cat("pinned witness digests:", length(dig), "entries; first three:\n")
.say(dig[1:3])
cat("\n---- ", basename(anc_f), " (head) ----\n", sep = "")
.say(.must(head(readLines(anc_f, warn = FALSE), 14L), "anchors head"))
cat("\n---- ", basename(gov_f), " (floor law bytes) ----\n", sep = "")
gov <- readLines(gov_f, warn = FALSE)
.say(.must(.grep_fixed(gov, "FLOOR_TABLE_FIELDS = ("), "floor table fields"))
.say(.must(.grep_fixed(gov, '"_SWEEP_FIT", "_SWEEP_RAW"'),
"floor table fields tail"))
.say(.must(.grep_fixed(gov, "FLOOR_RATIO = 3.0"), "floor ratio constant"))
.say(.must(.grep_fixed(gov, "def floor_table_epoch()"), "epoch derivation"))
.say(.must(.grep_fixed(gov, "RE_ADJUDICATION_MECHANISM = ("), "re-adjudication law"))
hit <- which(vapply(gov, function(l) grepl("('sigminer', 'clock', 1000):", l,
fixed = TRUE), logical(1)))
if (!length(hit)) stop("fail-closed: the rep-100 floor anchor cell scan yielded no rows\n", domain = NA)
.say(gov[seq(hit[1], hit[1] + 3L)])
nw <- .must(.grep_fixed(gov, "'witness': ("), "floor witness pointers")
cat("witnessed floor-anchor cells:", length(nw), "\n")
}
}
#> ---- CALIBRATION.json ----
#> "schema": "m7-floor-calibration-1",
#> "probes": 93,
#> "freshness": "fresh per-leg cachedir per probe; a 'cached' replay is DEFECT, never a floor",
#> "gate": "--floor-gate off: a probe measures, it never self-adjudicates",
#> "reps": "42 grid witnesses at REP-100 + 18 main-arm probe-ladder witnesses at REP-{1,5,10} (the committed 60-class shape)"
#> pinned witness digests: 60 entries; first three:
#> "sigminer__clock__N1000__rep100.json": "2280ea84afdfd11b7956d6e01bf9b20a03ef2c83e96cd4bbf3357ef7ce1f3741",
#> "sigminer__comp_clock_only__N10000__rep100.json": "c1e177949ba95e2775887755df35f1f8e2cf0cad228c7e1fad1c6cf4e51d0e23",
#> "sigminer__comp_clock_only__N1000__rep100.json": "31f1218456a457ddea1afa50ee8f181d79683fad6a88edddfdffba71790332a4",
#>
#> ---- anchors_fragment.json (head) ----
#> {
#> "sigminer|clock|N1000": {
#> "cpu_seconds": 1786.2061079999999,
#> "reps": 100,
#> "wall_seconds": 1787.16,
#> "witness": [
#> "m7_floor_calibration_p7cap_legacy_20261027-083955",
#> "witnesses",
#> "sigminer__clock__N1000__rep100.json"
#> ]
#> },
#> "sigminer|comp_clock_only|N100": {
#> "cpu_seconds": 1265.724519,
#> "reps": 100,
#>
#> ---- matrix_governance.py (floor law bytes) ----
#> FLOOR_TABLE_FIELDS = ("_FLOOR_ANCHORS", "FLOOR_RATIO",
#> "_SWEEP_FIT", "_SWEEP_RAW")
#> FLOOR_RATIO = 3.0
#> def floor_table_epoch() -> str:
#> RE_ADJUDICATION_MECHANISM = (
#> ('sigminer', 'clock', 1000): {
#> 'witness': ('m7_floor_calibration_p7cap_legacy_20261027-083955', 'witnesses', 'sigminer__clock__N1000__rep100.json'),
#> 'reps': 100, 'cpu_seconds': 1786.2061079999999, 'wall_seconds': 1787.16,
#> },
#> witnessed floor-anchor cells: 189Re-extract any of the above yourself — the commands print the same bytes:
The P10 bridge-face claim, quoted from its witness
The batch-bridge result (the only speedup statement on this page) is quoted from the bytes of its landed commit, never retyped: the frozen-geometry A0 before/after pair, bit-identical output, witnessed in the commit body.
if (is.na(PKG)) {
cat("rendered outside the package root: no repository to read the commit from\n")
} else {
outf <- tempfile(fileext = ".txt")
st <- tryCatch(system2("git",
c("-C", PKG, "show", "-s", "--format=%B", "2296f49"),
stdout = outf, stderr = FALSE, wait = TRUE),
error = function(e) NA_integer_)
on.exit(unlink(outf), add = TRUE)
ok <- !is.na(st) && identical(as.integer(st), 0L)
if (!ok || !file.exists(outf) || file.size(outf) <= 0) {
cat("git not available in this context: read the bytes with `git show -s --format=%B 2296f49` in a checkout\n")
} else {
b <- readLines(outf, warn = FALSE)
l1 <- .must(.grep_fixed(b, "per-run before"), "A0 before figure")
l2 <- .must(.grep_fixed(b, "landing ratio"), "landing ratio figure")
.say(unique(c(l1, l2)))
cat("(witness: commit 2296f49 body; laws in tests/testthat/test-kernel-batch-bridge.R)\n")
}
}
#> git not available in this context: read the bytes with `git show -s --format=%B 2296f49` in a checkoutThe sealed head-to-head adjudications
Peer-comparison numbers exist here only as sealed, adjudicated
campaign verdicts with their witness paths — never as headlines. The
bootstrap campaign’s stored verdict is reproduced verbatim from its
committed result.md:
if (is.na(PKG)) {
cat("rendered outside the package root: the campaign witness is not reachable here\n")
} else {
res_f <- file.path(PKG, "bench", "results",
"bootstrap_vs_sigminer_20261013T175859Z", "result.md")
if (!file.exists(res_f)) {
cat("campaign witness not on disk (check tarball): bench/results/bootstrap_vs_sigminer_20261013T175859Z/result.md\n")
} else {
cat("---- ", basename(res_f), " (verbatim) ----\n", sep = "")
.say(.must(readLines(res_f, warn = FALSE), "sealed campaign verdict"))
}
}
#> ---- result.md (verbatim) ----
#> # bootstrap-vs-sigminer campaign — result
#>
#> - tier: B
#> - ratio(A1/A2) = 1.536600758177e+01
#> - verdict: MEASURED-false 15.37x
#>
#> ## Disclosure
#> - seed: 20260928 (G0 master; the arms carry it into their own RNG regimes; RNG-stream equality is NOT claimed — data bytes carry identity)
#> - shape: K=1000 S=1000 n_sig=50
#> - heavy arms: single authoritative timed shot each (RB-10); repeatability
#> evidence lives in the probe witnesses (warmup + 3 timed, median)
#> - msuiter catalog construction happens outside the timed regions
#> (harness glue); both arms are timed on their documented entry point
#> - host: ShixiangWang-WorkSpacedeMacBook-Pro.local R: R version 4.5.2 (2025-10-31) hw_profile: host-R campaign | Apple Silicon | 10 cores | hw.ncpu=10 hw.memsize=25769803776 | R 4.5.2
#> - data seal: {"sha256": "cc094225d599a6765dbc7958a445609459ded0f32f2481f87c5cb9e8762b3263", "campaign_rds": "campaign.rds","host": "ShixiangWang-WorkSpacedeMacBook-Pro.local", "hw_profile": "host-R campaign | Apple Silicon | 10 cores | hw.ncpu=10 hw.memsize=25769803776 | R 4.5.2", "generator": "gen_campaign.R","argv": "-- --out bench/results/bootstrap_vs_sigminer_20261013T175859Z --k 1000 --s 1000 --sig 50 --seed 20260928"}
#> - co-tenancy census: witnesses/host_census.jsonl (verbatim docker ps)
#> - secondary A1/A3 (pool-scaling isolation): 14.2209The adjudication voice — why a MEASURED-false verdict
stays published rather than being re-measured into a nicer shape — is
the governance record in vignettes/articles/performance.qmd;
the co-tenancy attestation census for this campaign rides at
bench/results/bootstrap_vs_sigminer_20261013T175859Z/witnesses/host_census.jsonl.
Reproduction
The frozen matrix and its floor gate are reproducible from the
committed drivers; the probe legs that mint floor witnesses run under
the freshness law (fresh per-leg cachedir,
--floor-gate off, full witness provenance):
python3 bench/m7/run_bench.py --help # the matrix driver
python3 bench/m7/tests/run_all.py # the governance suite
Rscript -e 'devtools::load_all(); ms_sitrep()' # the installation sitrepThe one-command reproduction face is
inst/repro/reproduce.sh, banked by the paper lane and
attested in vignettes/articles/reproduction-proof.qmd. Its
default --verify face dry-verifies every sealed number
against committed bytes without containers, without network and without
long campaigns. This page LINKS that face, it never forks it: exactly
one reproduction script exists in the repository, no second script
anywhere re-implements the --verify face, and the law
tests/testthat/test-reproduce-link.R keeps that census
true.
Every performance claim msuiter makes in prose (README, manual pages, this vignette) must name its witness path; a number without a witness path is not a claim, and this page holds none.