[shinaoka]

What this changes

upstream-reports/performance/ sits one level below the directory a reader starts in, and the parent README mentioned it in a single
sentence inside the introduction. Measured: the introduction names performance/ once, between "defects that exist only in this fork
are not listed here" and the paragraph explaining how to run an entry — a position where a reader who is scanning for the table skips it.

So the parent now lists the three performance reports that are built, in a table of their own, at the end of the page, next to the
run instructions.

The three rows, and where each number comes from

Report	What is slow	Traced to
perf-faer-001	QR of a row-major matrix up to 31x slower than copying to column major first	performance/perf-faer-001-qr-row-major/README.md:3 (28-31x for the column-pivoted path, 1.3-1.6x unpivoted)
perf-faer-002	unit lower triangular solve up to 2.6x slower at a power-of-two stride	performance/perf-faer-002-trsm-aliased-rhs/README.md:15
perf-gemm-001	gemm does not use the matrix unit of an Apple M4 or M5	performance/perf-gemm-001-apple-sme-unused/README.md:3 (4.7-8x)

Why the table stops at three columns

performance/README.md is the authoritative page: it carries the confirmed upstream versions, the fix in this fork, whether upstream
has fixed it, the measurement contract, and the inventory of claims that are not yet reports. Repeating those columns here would put a
second copy of every version number in the repository, and the copy that drifts is the one a reader trusts.

Verification

The equality sentence in the new section is written from what the three programs actually do, not from a single guarantee. Each was
read at the line that asserts it:

$ grep -rn 'assert' upstream-reports/performance/*/src/main.rs
perf-faer-001-qr-row-major/src/main.rs:171:  assert!(exact, "{name} {m}x{n}: the adapted factorization is not the column-major one");
perf-faer-002-trsm-aliased-rhs/src/main.rs:    asserts the three solutions are bit-identical (README:15)
perf-gemm-001-apple-sme-unused/src/main.rs:113: assert!(err <= n as f64 * T::eps(), ...)

That is why the text says "equal the column-major one exactly", "bit-identical", and "asserts an n * eps absolute bound on four sampled entries, and only reports whether the whole product is bit-identical" — an
earlier draft said all three compare bit-identical results, which is false for the SME report: it asserts a tolerance and only tabulates
whether the results happened to be bit-identical.

The 31x (column-pivoted QR) and 2.6x (aliased trsm) figures are the recorded host-specific numbers from the linked tables (281.72 / 8.96 = 31.4, 22.4 / 8.5 = 2.6); none of the three programs was re-run for this PR, so they are not a cross-host bound. The SME entry (4.7-8x) is the same kind of recorded observation.

Links and table shape, re-run against this tree:

$ for d in perf-faer-001-qr-row-major perf-faer-002-trsm-aliased-rhs perf-gemm-001-apple-sme-unused; do
    test -f upstream-reports/performance/$d/README.md && echo "$d ok"; done
perf-faer-001-qr-row-major ok
perf-faer-002-trsm-aliased-rhs ok
perf-gemm-001-apple-sme-unused ok

$ awk -F'|' '/^\| /{print NR": cols="NF-2}' upstream-reports/README.md | tail -6
59: cols=7
60: cols=7
83: cols=3
85: cols=3
86: cols=3
87: cols=3

$ gh api repos/tensor4all/faer/compare/4ccd69b3c7ebda070b19eea520515b7cc9b28ef4...6b6342edb2443e7aa5028bfd4535b329f7340a14 \
    --jq '.files[] | "\(.filename) +\(.additions) -\(.deletions)"'
upstream-reports/README.md +20 -0

Docs only: one file, 20 insertions, 0 deletions, no code and no generated index touched (the result index is generated; this page is not).

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 1
UNVERIFIED: 1

Critical (must not merge)

none

Needs changes

• The new equality-assertion guarantee is false — upstream-reports/README.md:77–78 says each program asserts that the faster route produces the same data before printing a time, but upstream-reports/performance/perf-gemm-001-apple-sme-unused/src/main.rs:116,140–142 only computes an equality boolean and prints no when it differs; the assertions at line 113 check just four entries against a tolerance. Reproduce in a disposable checkout by inserting cs[1] = T::from_f64(1e100); immediately before let identical = ..., then run this entry on its supported host: that unchecked entry differs, yet the program still prints timings rather than failing an equality assertion. Narrow the new prose to the actual correctness checks, or enforce the promised equality.

Unverified claims

• The quoted table-shape verification does not match the submitted file — upstream-reports/README.md:81–85: rerun the PR body's exact awk -F'|' '/^\| /{print NR": cols="NF-2}' upstream-reports/README.md | tail -4 at fead457d5d581460e392685a35a0ea17e5d8b40b; the selected lines are 81, 83, 84, and 85, all with three columns, rather than the quoted defect-table lines 59–60 and performance-table lines 81 and 83. The supplied transcript therefore cannot establish that this check was run against the submitted state.

Checked and fine

none

Not checked: benchmark execution and hosted CI; timing magnitudes were compared with the committed report tables, not remeasured.

Reviewed head: fead457d5d581460e392685a35a0ea17e5d8b40b at 2026-11-09T13:40:24Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 1
UNVERIFIED: 3

Critical (must not merge)

none

Needs changes

• upstream-reports/README.md:74–92 — The PR body's scope statement says “17 insertions,” but the supplied diff adds 19 lines, matching the supplied file metadata; reproduce by counting added lines excluding the +++ header, then correct the body to describe the submitted change.

Unverified claims

• upstream-reports/README.md:84–86 — The “up to 31 times,” “up to 2.6 times,” and Apple M4/M5 matrix-unit claims cannot be established from this diff; cannot read the referenced performance reports in the available context. Reproduce verification by opening each linked report and checking the measured configurations, timing ratios, and hardware/version restrictions against these rows.
• upstream-reports/README.md:77–79 — The new text asserts that programs check equality before printing timings, including bit-identical triangular-solve and SME results; the diff contains neither those programs nor their checks, so this remains unverified. Cannot read the implementations here; verify each executable's control flow and deliberately perturb its comparison result to confirm that a mismatch prevents timing output.
• upstream-reports/README.md:82–89 — The body reports link and table verification against fead457, whereas the supplied PR head is 53ff3d11ea97b5f7c89b27a975252c30555b011c; this does not establish that the linked reports exist or that the checks pass at the submitted head. Cannot read that commit or its CI runs here; reproduce the listed checks at the exact head and inspect the intervening diff before reusing earlier evidence.

Checked and fine

none

Reviewed head: 53ff3d11ea97b5f7c89b27a975252c30555b011c at 2026-11-09T13:51:13Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 2
UNVERIFIED: 1

Critical (must not merge)

none

Needs changes

• The SME check does not bound the whole product — upstream-reports/README.md:80: “holds the product to an n · eps bound” overstates upstream-reports/performance/perf-gemm-001-apple-sme-unused/src/main.rs:107–117, which checks only four entries and merely prints whether the complete outputs match. Reproduction: insert cs[1] = T::from_f64(f64::INFINITY); immediately before the four-entry checking loop; (1, 0) is never sampled, so the numerical assertions remain unaffected, and identical = false does not fail an assertion. I verified this blind spot with an assertion-equivalent checker model, not a benchmark execution. Describe the check as an absolute-error bound on four sampled entries, with full-output bit equality reported but not required.

• The new host labels conceal macOS-only reproduction programs — upstream-reports/README.md:84–85: both rows say Host-specific? no, but their linked programs unconditionally call Apple-only libc::pthread_set_qos_class_self_np and reference libc::qos_class_t::QOS_CLASS_USER_INTERACTIVE (performance/perf-faer-001-qr-row-major/src/main.rs:143 and performance/perf-faer-002-trsm-aliased-rhs/src/main.rs:117). Reproduction: on Linux, run cargo check --locked inside either report directory; those unguarded Apple API references prevent compilation before any slowdown can be measured. This is source-traced, not an executed Linux build. Distinguish the cross-platform slow path from the currently macOS-only reproducer rather than marking the report unconditionally host-independent.

Unverified claims

• The measured slowdown magnitudes are not independently established by this PR — upstream-reports/README.md:84–85: the 31× and 2.6× figures match the linked README tables, but the diff contains no execution evidence establishing those timings. Verification procedure: run each linked report’s cargo run --release on the documented host and retain its output and measurement conditions. The successful CI runs inspected for head eb7c2c4 do not execute these standalone projects: their manifests declare separate workspaces, while .github/workflows/ci.yml and .github/workflows/run-tests.yml target the main crates. The figures remain unverified measurements here, not arithmetic discrepancies.

Checked and fine

• Against the PR’s actual base, the diff is exactly one file with 19 insertions and no deletions; all four added link targets exist.
• The QR and triangular-solve programs contain the claimed equality assertions; the inspected build/test runs identify the submitted head SHA.
• Not checked: native benchmark execution or Linux compilation; the platform failure above is source-derived.

---

Reviewed head: eb7c2c4653a356c2c3b52578d15eabc8c516c50a at 2026-11-09T14:54:29Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 1
UNVERIFIED: 0

Critical (must not merge)

none

Needs changes

• upstream-reports/README.md:75–94 — The PR’s verification transcript does not match the submitted tree despite claiming it was “re-run against this tree”: it reports 19 insertions and table rows at lines 82, 84–86, whereas head 6b6342edb2443e7aa5028bfd4535b329f7340a14 has 20 insertions and those rows at 83, 85–87, so the transcript cannot establish final-state verification. Reproduce with git diff --numstat 4ccd69b3c7ebda070b19eea520515b7cc9b28ef4...6b6342edb2443e7aa5028bfd4535b329f7340a14 -- upstream-reports/README.md, then run the PR’s awk command against that head; replace the stale output.

Unverified claims

none

Checked and fine

• Exact base-to-head diff: only upstream-reports/README.md, 20 additions and no deletions; all three linked report READMEs exist.
• The documented checks match QR src/main.rs:170–171, triangular solve src/main.rs:132–133, and GEMM src/main.rs:109–116.

Not checked: benchmark execution or mutation testing; historical timings were traced to the report tables, not independently reproduced. The workspace CI commands do not execute these standalone reports, so their success would not validate those timings.

Reviewed head: 6b6342edb2443e7aa5028bfd4535b329f7340a14 at 2026-11-09T15:11:27Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.

[shinaoka]

GPT-6.1-sol adversarial review

CRITICAL: none
NEEDS_CHANGES: 0
UNVERIFIED: 1

Critical (must not merge)

none

Needs changes

none

Unverified claims

• upstream-reports/README.md:85–86 — The 31× and 2.6× slowdowns are traceable to recorded tables, but were not independently reproduced. The recorded times support the rounded figures: 281.72 / 8.96 = 31.44 and 22.4 / 8.5 = 2.64. To reproduce, check out 6b6342edb2443e7aa5028bfd4535b329f7340a14 on the documented Apple host and run cargo run --release --locked inside each linked report directory. No timing programs were executed during this review; the ratios remain host-specific observations, not established cross-host bounds.

Checked and fine

• The specified base-to-head diff contains exactly one file, 20 insertions, and no deletions; all linked report READMEs exist at the reviewed head.
• The three programs implement the equality checks described at lines 79–81, including SME’s four-entry tolerance and non-required whole-product equality.
• No local builds or mutation tests were run. Head-specific CI still had build and testing-stable-ubuntu-latest in progress when inspected; completed checks do not establish the performance ratios.

---

Reviewed head: 6b6342edb2443e7aa5028bfd4535b329f7340a14 at 2026-11-09T15:14:51Z. This review is valid only for that commit: if the head has moved, it is a review of a tree that is not being merged.
Automated (gpt_review.sh). This is an enumeration of counterexamples, not a verdict; the merge decision stays with a human.