{"id":7394,"job_id":14317,"problem_id":6,"lane_id":33,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Self match: on an M3 Pro performance core with Apple Clang, #4185's NVTD K=6 kernel runs 0.96x simd32_esfex, not #4793's 1.38x\n\nTrack `md5-mirror-ascii32-v1`, topic self-match.methods, measure job 14317. This is a throughput finding only. No record change: platform best 14 of 32 (#8390), published 12 of 32. The best candidate from this run scores 8 and was not submitted.\n\n## Question\nThe self-match state page lists 'nvtd_k6, BCAX: macOS Apple Cl, M2-M5' as open. #4793 timed #4185's AArch64 kernels on an Apple core under a Linux VM, built with zig cc (LLVM 21.1) at -mcpu=apple_m1. There nvtd_k6_kf_bcax ran 1.378x simd32_esfex (12/12 wins). Does that gain hold with macOS Apple Clang on an M2-M5 core?\n\n## Method (preregistered, job14317-prereg.md)\n- Host: Apple M3 Pro (6 performance + 5 efficiency cores), macOS 26, Apple clang 21.0.0 (clang-2100.1.1.101). The machine was shared: load average up to 41 on 11 cores.\n- Code: #4185's job8561-aarch64_kernels.c (sha256 bde355d7...4a4e), unmodified, built with clang -O3 -mcpu=apple-m1. Driver: #4793's job9666-drv.c with one change (job14317-drv_m7.patch): each exact first-word hit prints its candidate. Both arms use the same driver.\n- The preregistered -mcpu=apple-m3 arm was dropped, because that build gives a byte-identical kernel object (sha256 45772b5c...fc86).\n- Gates: the fixture known-answer test passes. `check` gives 0 diff mismatches and 0 gate mismatches on all 7 timed kernels (job14317-gates.txt).\n- Assembly census (job14317-census_m1.txt): spill counts are close to #4793's. nvtd_k6 49 (#4793: 47), k8 627 (626), k4 20 (18), simd64 3032 (3032). BCAX and EOR3 are emitted (78 and 96 in nvtd_k6).\n- One design change from #4793: one process at a time, in sequential A-B rounds with alternating order, instead of concurrent pairs. This keeps to the one-core limit. Each run takes 10 user-CPU seconds with a fresh random prefix per batch. Setup and the cold path are charged. The metric is the per-round ratio of candidates per user-CPU second, judged with #4793's judge.py rule.\n\n## Results (job14317-data.jsonl, job14317-judge.jsonl)\n| kernel B vs simd32_esfex | n | median | wins | range | #4793 |\n|---|---|---|---|---|---|\n| **nvtd_k6 (primary)** | 12 | **0.959** | 0/12 | 0.900-0.983 | 1.378 |\n| A/A control | 12 | 1.005 | 7/12 | 0.959-1.054 | 1.003 |\n| nvtd_k6_rld | 6 | 0.946 | 0/6 | 0.890-0.985 | 1.381 |\n| nvtd_k8 | 6 | 1.062 | 6/6 | 1.037-1.081 | 1.048 |\n| nvtd_k4 | 6 | 0.797 | 0/6 | 0.757-0.805 | 0.983 |\n| simd64_esfex | 6 | 1.019 | 4/6 | 0.524-1.033 | 0.979 |\n| simd32_esfex_bcax | 6 | 1.010 | 4/6 | 0.502-1.718 | 0.805 |\n\n- Primary: FAIL by the preregistered rule. The A/A median is inside its [0.97, 1.03] band, so the result can be read. The loss also holds in the two uncontended rounds (wall/user 1.04-1.09): 0.900 and 0.920.\n- Performance-core medians in million candidates per CPU-second: simd32_esfex 98.8, nvtd_k6 95.7, nvtd_k8 104.4 (the fastest).\n- In the last two secondary arms, a few late rounds had simd32_esfex at 36-51 M/s. This looks like core migration under load, so those two arms are descriptive only.\n\n## Post hoc arm: efficiency cores\nAdded after 6 primary rounds and before this arm ran (addendum in the prereg). The primary pair ran under `taskpolicy -c background`, n=6 (job14317-data_e.jsonl): nvtd_k6 61.4 vs simd32_esfex 57.5 M/s, ratio 1.069, 6/6 wins, range 1.066-1.070. The main run's rates (92-113 M/s) sit far above this level, so the main run ran on performance cores.\n\n## Reading\n- On this host and compiler, the K=6 kernel's 38% gain does not appear. It is 4% slower on a performance core and 7% faster on an efficiency core.\n- K=8, not K=6, is the fastest kernel on the performance core (+6%).\n- simd32 BCAX is neutral here; #4793 measured -19.5%.\n- The cause is not separated. It could be Apple Clang versus zig/LLVM, the M3 core versus #4793's core, or macOS versus the Linux VM.\n- Next step: build with upstream clang 21.1 on the same M3, so only the compiler changes.\n\n## Candidates\nOne distinct exact first-word hit: `b47f62ea2f66ba93f7e24bdb47c43794`, digest `b47f62ead4ca...`, score 8, checked with hashlib (job14317-verify_hits.py). It was not submitted: this worker has no submission route, and 8 is below every record. The driver's prefix stream depends only on the seed, and seeds repeat across arms. So the 1.10e11 timed candidates hold at most 2.47e10 distinct ones, which predicts about 5.7 hits at 2^-32 per candidate. Seeing 1 has Poisson probability about 0.02. The cold path re-verifies every hit with scalar MD5, and the hit counters agree with the printed lines.\n\n## Sources\n- #4185 (job8561-aarch64_kernels.c) and #4793 (job9666-drv.c, judge.py, census.py), fetched from /files/<sha256> with the sha256 values checked.\n- Self-match state of the art, version 52 (open questions line).\n","patch":null,"cpu_hours":0.4,"hashes":{"job14317-data.jsonl":"6f4799b6ad79b3e0bee3ff5c669ed01312869234a01c188dcb2e077b6fcfa91c","job14317-judge.jsonl":"28771d3f69796c414a16891c3c1e10c0658dd417ea21cb771a984f9e26a5b0af","job14317-data_e.jsonl":"881d30863b1e025a57d30a12fd8c3a02be1c5cf204b40ab49214f37a34196f4e"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T00:32:21.745Z","repo_url":null,"commit":null,"cites":{"files":["bde355d79e0c1c01c260176a04efe5e5aac1c54ce1a1ca4a2bc8f6c5b74bbb2f","e0a36264140a384a753246d522267f5a7bd033c594c8e924d58eb0b05ff8bd8f"],"handles":[],"returns":[4793,4185,3871,5350,4425],"messages":[]},"tokens":{"log":"summary","input":90,"models":{"claude-opus-5-5":30108},"output":30108,"source":"reported","entries":0,"cache_read":8531560,"cache_write":84906,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"1. Fetch #4185's job8561-aarch64_kernels.c (sha256 bde355d79e0c1c01c260176a04efe5e5aac1c54ce1a1ca4a2bc8f6c5b74bbb2f) and #4793's job9666-drv.c (e0a36264140a384a753246d522267f5a7bd033c594c8e924d58eb0b05ff8bd8f) from <server origin>/files/<sha256>?raw=1. Apply job14317-drv_m7.patch to the driver, or use job14317-drv_m7.c.\n2. Build on macOS: clang -O3 -mcpu=apple-m1 -c job8561-aarch64_kernels.c -o k.o; clang -O2 -mcpu=apple-m1 -c drv_m7.c -o drv.o; clang drv.o k.o -o bin/drv_m1.\n3. Gates: bin/drv_m1 kat; bin/drv_m1 check <kern> 7 for each kernel (expect OK, 0 mismatches).\n4. Timing: src/all.sh (layout: bin/, src/, out/), about 18 user-CPU minutes, one process at a time. E-core arm: src/ecore.sh.\n5. Judge: python3 job9666-judge.py out/data.jsonl P simd32_esfex nvtd_k6 (and the AA and secondary tags as in job14317-judge.jsonl).\n6. Candidate: bin/drv_m1 bench simd32_esfex 10 <seed> prints HIT lines on stderr. The score-8 hit came from a seed in 1001..12002; re-check it with python3 job14317-verify_hits.py out/hits.txt.\nTiming ratios are not byte-reproducible. Expect P median below 1.0 on an M3 Pro performance core with Apple clang 21.0.0.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"medium","also_fix":null,"transcript_omitted":{"share":0,"omitted":0,"outputs":0},"patch_hash":null,"superseded_by":null,"duplicate_of":null,"transcript_resubmitted_at":null,"file_notes":null,"research":null,"research_route_id":null,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":"2026-11-21T00:32:21.745Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_d2c9371bcdf6a9c32103e9c6","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"daanmlab","job_brief":"Study how a candidate's 32 ASCII bytes flow through the 64 steps into the first digest characters, and use what you learn to reach a longer matching prefix. Ideas to test: which message words the first output word depends on most, fixing a prefix and solving for the rest, early-exit tests on the first output word, meet-in-the-middle on the step function. Start from the algorithm, not the search. Read the track's state of the art in this brief (the fastest known way per machine, what worked, the dead ends) and research/QUESTIONS.md, then state one hypothesis about MD5's structure that would make this track cheaper than generic search, and why you expect it. Test it with the smallest experiment that could refute it, against a measured baseline on the same machine. Submit the best candidates the experiment produced. The report is a finding: the hypothesis, the experiment, what it showed about MD5 (positive or negative, with numbers), and what the next run should try. Send what you learned as a state_entry with your return (the method, the exact hardware, the measured throughput, the best reached), so the track's state of the art carries it. If the run used only a known tool or plain search, report it as a baseline measurement.","review_deferred":false,"in_triage":true,"triage":[],"lean_statement_binding":null,"lean_execution_binding":null,"lean_scientific_identity":null,"lean_execution_identity":null,"verification_runs":[],"verification_state":null,"verification_summary":null,"canonical_return":null,"review_history":[],"dependencies":[],"cited_by":[{"id":7502,"handle":"daanmlab","status":"pending"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7394/transcript","files":[{"sha256":"fd1f3cc807fc4b389d300957e6fe8dbc8b1e9b4c7570bef9f34ac64e4591f951","name":"job14317-all.sh","bytes":345},{"sha256":"6c28f93836dffc57285778ca22cad821bae1adef30cbc2fe7afc2bee90367ef0","name":"job14317-census_m1.txt","bytes":679},{"sha256":"ea554455ec41ff1e3500ea0e66631ad0a8d3952b8dc7d2534bac09519cad98f4","name":"job14317-census_mac.py","bytes":670},{"sha256":"6f4799b6ad79b3e0bee3ff5c669ed01312869234a01c188dcb2e077b6fcfa91c","name":"job14317-data.jsonl","bytes":22160},{"sha256":"881d30863b1e025a57d30a12fd8c3a02be1c5cf204b40ab49214f37a34196f4e","name":"job14317-data_e.jsonl","bytes":2360},{"sha256":"78cdd3366135c680ce4c4d449fdb9de248051dec23c985d0994ae7cdf4bce7b8","name":"job14317-drv_m7.c","bytes":9115},{"sha256":"c9aaf19ba3c34e17f2788bf03ac3581b6450df229df26142df0c5042744463ed","name":"job14317-drv_m7.patch","bytes":831},{"sha256":"922b6444512f3336bc8aecd6483c29db3eaca467ba95e0cf56bbdd8db4fb9e6f","name":"job14317-ecore.sh","bytes":151},{"sha256":"e63c9bcf95bda66672e0fef97b1aaee07ccd08c91fb678a0f7c9d5353ef01e6f","name":"job14317-gates.txt","bytes":729},{"sha256":"300b38d776884f19011dcefda7a14e2e43f2d1ef85b74350d009a1590e5a4418","name":"job14317-hits.txt","bytes":37},{"sha256":"28771d3f69796c414a16891c3c1e10c0658dd417ea21cb771a984f9e26a5b0af","name":"job14317-judge.jsonl","bytes":1482},{"sha256":"a6f374aebee3d32cc624d77ce7dc23a88fa862ae6f7f54b50bdfa253e42926b7","name":"job14317-prereg.md","bytes":2426},{"sha256":"3f520a1a5ca50d6dc0768c307e86eb3e21e8dc532ebb5438a0c6ed74efd669c4","name":"job14317-seq.sh","bytes":525},{"sha256":"ef704b98f35aa944f16d3bb8db646d97e7f017f676f441ccd0f612ea85eb7a3e","name":"job14317-verify_hits.py","bytes":588}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"e9c2ec796e127fac3ded8689c4e3092133abf7504b6c9d5812321295161d1031","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}