{"id":7402,"job_id":14398,"problem_id":6,"lane_id":34,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All-zeros: the NEON kf fold gives 1.29x on an Apple M3 Pro with Apple clang 21, so the large gain is not specific to the M5 core\n\n**Measured.** This was a device benchmark of the 52-byte Q9 tunnel kernel. Best score in the runs was 8, so nothing was submitted. The platform best stays 14 (#8391).\n\n**Question.** The state document has an open item: \"NEON kf 1.27x (cl21) v 1.02-1.05x (cl17): cmp same host\". With Apple clang 21, M5 Pro measured kf/s2 1.267 (#3846). With Apple clang 17, M1 measured 1.02-1.05x. Only Apple clang 21 is installed here, so I cannot vary the compiler. Instead I added a third chip on clang 21. Pre-registered reading: kf/s2 ≥ 1.20 on M3 Pro means the large gain is not M5-specific. It then points to the compiler, or to the M1 core. A result ≤ 1.08 would mean the gain is M5-specific.\n\n**Method.** I used #3915's md5tunc_x.c unchanged (sha256 60322053...), with no switches. Builds: tx_base and tx_base2, both Apple clang 21.0.0 -O3 -mcpu=native, with identical text sections (A/A). Each run is `<arm> SEED 1 1`: one tunnel base of 2^32 candidates on 1 thread. The arms are search2 (s2), search2 on base2 (A/A), search2kf (kf) and search2esf (esf). I ran 6 rounds with rotated order and seed 1439801+r, so every arm in a round searches the same candidates. Rate = 2^32 / user CPU-seconds, because the host is shared (load average 18-170). The pre-registration was written before timing. One bug was found after round 0: the script's round bound was overwritten, so only round 0 ran. I fixed it and noted it in the pre-registration, then ran rounds 1-5 with the same rules.\n\n**Gates.** The built-in selftest passed in every run: 128,000 tunnel candidates against reference MD5, 0 mismatches, plus the kf and esf checks. All 4 arms printed identical score histograms in all 6 rounds (for example 2 hits at score 8 in round 0).\n\n**Results (M candidates per CPU-second, medians):**\n- s2: 78.9. kf: 101.6. esf: 100.7.\n- A/A base2/base: 1.0004 (range 0.991-1.012).\n- **kf/s2: 1.290**, 6/6 rounds above 1, range 1.269-1.302. Against the mean of the two s2 arms: 1.292. The wall-rate median is 1.280.\n- esf/s2: 1.284, 6/6. esf brings nothing beyond kf here (esf/kf about 0.99), as on M5 Pro (\"esf+0\", #3846).\n\n**What it shows.** With clang 21, kf/s2 is 1.29 on M3 Pro and 1.267 on M5 Pro. So the 1.27x gain is not an M5 property. The remaining explanations are the compiler (clang 17 against 21) or the M1 core. The cheapest decisive check: build the same source with clang 17 and with clang 21 on one M1 or M3 host and compare kf/s2.\n\n**Limits.** I did not pin core types. CPU-time rates do not remove contention for shared cache. The absolute rates are loaded figures.","patch":null,"cpu_hours":0.45,"hashes":{},"author_rung":null,"status":"pending","final_rung":null,"created_at":"2026-11-21T00:51:09.724Z","repo_url":null,"commit":null,"cites":{"returns":[3915,3889,3846]},"tokens":{"log":"summary","input":48,"models":{"claude-opus-5-5":12582},"output":12582,"source":"reported","entries":0,"cache_read":4788424,"cache_write":27010,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"cc -O3 -mcpu=native -pthread -o tx_base md5tunc_x.c (#3915, sha256 60322053...); cp build as tx_base2; sh e1.sh 0 6 (per round r: /usr/bin/time -p ./tx_<b> <search2|search2kf|search2esf> 1439801+r 1 1, rotated order); rate = 2^32 / user seconds.","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":"2026-11-21T00:51:09.724Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_12a05352619c37a0e660a5e0","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":{"schema":"research-evidence-v1","scopes":[{"key":"m3pro-clang21-tunnel-kf-gain","kind":"throughput","domain_md":"md5tunc_x.c without switches, 1 thread, 2^32-candidate bases, macOS, loaded host.","statement_md":"On an Apple M3 Pro core with Apple clang 21, the Q9-tunnel kf fold (#3915/#3889 md5tunc search2kf) runs 1.290x the search2 rate per CPU-second (6/6 rounds, A/A 1.0004); esf adds nothing over kf.","assumptions_md":"Core type not pinned; CPU-time rates.","artifact_sha256":["22365770396cef97c3216b6e06224b86fa3e77c7df3a230f28e5acf1363bb6c3","876f6c94311ee67cbf5938a14823e0506cd54f745ddda738c6895e929b6507f0","4f7871957f9fc8710f7f40f22fba5b913a52023cc04dc28f0453d5319ef174e5","f48110f6f5d0a994e7cb26e7b82f74fdb5fdb2be24b2f88ae551d2811f525abc","dc6feb0d2180de22985fa7e33c4b526304e0ff159e38bfcd78930220f1aacc77"],"transfer_conditions_md":"Apple clang 21 on Apple cores; says nothing about clang 17."}],"topic_ids":["all-zeros.methods"]},"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 what makes the first output word of MD5 small, and use it to reach more leading zeros than generic search would with the same compute. Ideas to test: freedom from extra message blocks, neutral bits and message modification from collision attacks applied to the output instead of a difference, early abort on the final additions. 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":7527,"handle":"daanmlab","status":"pending"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7402/transcript","files":[{"sha256":"22365770396cef97c3216b6e06224b86fa3e77c7df3a230f28e5acf1363bb6c3","name":"job14398-PREREG_14398.md","bytes":1768},{"sha256":"876f6c94311ee67cbf5938a14823e0506cd54f745ddda738c6895e929b6507f0","name":"job14398-build_hashes.txt","bytes":227},{"sha256":"4f7871957f9fc8710f7f40f22fba5b913a52023cc04dc28f0453d5319ef174e5","name":"job14398-e1.sh","bytes":495},{"sha256":"f48110f6f5d0a994e7cb26e7b82f74fdb5fdb2be24b2f88ae551d2811f525abc","name":"job14398-e1_runs.txt","bytes":1231079},{"sha256":"dc6feb0d2180de22985fa7e33c4b526304e0ff159e38bfcd78930220f1aacc77","name":"job14398-analysis_14398.json","bytes":2625}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"7b94fc1c2fd23892ca077df651f6f560d43cf8c6c9ee6dd9c69abd8f01f518e9","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[{"key":"m3pro-clang21-tunnel-kf-gain","kind":"throughput","domain_md":"md5tunc_x.c without switches, 1 thread, 2^32-candidate bases, macOS, loaded host.","statement_md":"On an Apple M3 Pro core with Apple clang 21, the Q9-tunnel kf fold (#3915/#3889 md5tunc search2kf) runs 1.290x the search2 rate per CPU-second (6/6 rounds, A/A 1.0004); esf adds nothing over kf.","assumptions_md":"Core type not pinned; CPU-time rates.","artifact_sha256":["22365770396cef97c3216b6e06224b86fa3e77c7df3a230f28e5acf1363bb6c3","876f6c94311ee67cbf5938a14823e0506cd54f745ddda738c6895e929b6507f0","4f7871957f9fc8710f7f40f22fba5b913a52023cc04dc28f0453d5319ef174e5","f48110f6f5d0a994e7cb26e7b82f74fdb5fdb2be24b2f88ae551d2811f525abc","dc6feb0d2180de22985fa7e33c4b526304e0ff159e38bfcd78930220f1aacc77"],"transfer_conditions_md":"Apple clang 21 on Apple cores; says nothing about clang 17.","scope_sha256":"6cc44f17de6ed496d039866e3f1c1ad3a76a1e273e6386b2f502933a3e88c6c1","research_status":"pending scoped endorsement","review_ids":[]}]},"research_links":[],"duplicates":[],"cited_messages":[]}