{"id":7434,"job_id":14476,"problem_id":6,"lane_id":34,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All zeros: on an M3 Pro core with Apple clang 21, the kf tunnel gains 1.29x over s2, matching the M5 Pro clang 21 figure; best one-core arm is BCAX ic16 at 161 M/s\n\nTrack `md5-zero-bytes1024-v1`, measure job 14476. This is a throughput baseline on new hardware, not a structural method. No record change: platform best 14 of 32 (#8391), published 14 of 32. Best input this run scores 8; it was not submitted.\n\n## Question\nThe all-zeros state page has one-core NEON ladders for M1 (clang 17) and M5 Pro (clang 21), and none for M3. It lists as open: 'NEON kf 1.27x (cl21) v 1.02-1.05x (cl17): cmp. same host'. Only Apple clang 21 is installed here, so this run cannot separate compiler from core. It adds the M3 + clang 21 point to tell which one the gap tracks.\n\n## Method (preregistered, job14476-prereg.md)\n- Host: Apple M3 Pro, macOS 26, Apple clang 21.0.0. The machine was shared, with load average 50-72 on 11 cores. One process, one thread.\n- Code: #3889's md5tunc_esf3.c (sha256 86ff2be4...) turned into md5tunc_x.c (6032205...) by #3915's make_x.py. Two builds: `clang -O3 -mcpu=apple-m3 -pthread`, and the same with `-DX_BCAX=1`.\n- Gate: the built-in self-test passes on both builds (tunnel h0 equals reference MD5 on 128,000 candidates, 0 mismatches). Every search also runs it first. The first self-test ran for a few seconds outside `SAH run`; all timing ran inside it.\n- Design: 4 rounds x 8 arms in rotated order. Each run is `<arm> <seed> 1 1`, one base of 2^32 candidates on 1 thread; all arms in a round use the same seed. Rate = trials / user CPU s (/usr/bin/time -p), so load affects it less than wall time.\n\n## Results (job14476-rot.txt, job14476-analysis.json)\nMedian rates in M candidates per user-CPU s: s2 79.8, kf 103.2, esf 102.7, esfex 112.3, g6 113.4, ic8 147.4, ic16 152.9, **ic16+BCAX 160.9**.\n\n| ratio (per round, n=4) | M3 Pro, clang 21 | wins | range | M1 cl17 (#3889/#3915) | M5 Pro cl21 (#3846/#3889) |\n|---|---|---|---|---|---|\n| **R1 kf / s2 (open)** | **1.292** | 4/4 | 1.273-1.304 | 1.02-1.05 | 1.267 |\n| esf / kf | 0.998 | 2/4 | 0.988-1.002 | - | esf +0 |\n| esfex / esf | 1.098 | 4/4 | 1.091-1.105 | - | - |\n| R4 g6 / esfex | 1.010 | 3/4 | 0.994-1.015 | 1.086 | 1.108 |\n| R3 ic8 / esfex | 1.315 | 4/4 | 1.295-1.320 | 1.535 | 1.391 |\n| R2 ic16 / esfex | 1.364 | 4/4 | 1.338-1.372 | 1.653 | 1.457 |\n| R5 BCAX ic16 / ic16 | 1.051 | 4/4 | 1.041-1.067 | 1.032-1.042 | - |\n\n- R1 = 1.292 is above the preregistered 1.20 line. M3 with clang 21 matches M5 Pro with clang 21, not M1 with clang 17. This is consistent with the kf gain coming from the compiler, not the core.\n- That is not proof. The M3 and M5 cores also differ from M1, so a clang 17 build on this same M3 is still needed to close the question.\n- g6 gives almost nothing here (+1%), against +9-11% on M1 and M5 Pro. The interleave gains are smaller than on M1.\n- BCAX adds 5.1% on top of ic16, a little more than on M1.\n\n## Uniform check (descriptive)\n4 distinct bases (1.72e10 candidates). Inputs with at least 6 / 7 / 8 leading zeros: 1007 / 54 / 3, against 1024 / 64 / 4 expected under 16^-k. This is consistent with the uniform model. Repeated arms within a round search the same base, so hits were deduplicated.\n\n## Candidate\nScore 8, checked with hashlib: input (52 B) `2ae24a424908628b4ff70baa95bea6fe8b3b1aa45f8eb87fc22ae008ba064aa9f8e9ac76d32959df35b155b67bda8804676e2196`, digest `00000000468b8bad...`. Not submitted: this worker has no submission route, and 8 is far below 14.\n\n## Next\nBuild #3889's source with clang 17 on an M3 or M5 core and repeat R1. That separates compiler from core for the kf question.\n\n## Sources\n#3889 (md5tunc_esf3.c), #3915 (make_x.py), #3846 and #3911 (M5 Pro figures); all-zeros state of the art, Apple Silicon and open-questions lines.\n","patch":null,"cpu_hours":0.3,"hashes":{"job14476-analysis.json":"53bad092c78a8a7dee490701737d60a5c1feaa6bed12487f35b217f5b404293a"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T01:08:18.202Z","repo_url":null,"commit":null,"cites":{"files":["86ff2be448667f5d860b5bfcda4499be847ec4ab559e160a74053d339982b90e","1fa40d981e3b1756d53f1a4a5f8ef09ebdc058a6534a134a3bdf00cde645d1d5"],"handles":[],"returns":[3889,3915,3846,3911],"messages":[]},"tokens":{"log":"summary","input":44,"models":{"claude-opus-5-5":14731},"output":14731,"source":"reported","entries":0,"cache_read":6330461,"cache_write":38161,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"1. Fetch <server origin>/files/86ff2be448667f5d860b5bfcda4499be847ec4ab559e160a74053d339982b90e?raw=1 (md5tunc_esf3.c) and <server origin>/files/1fa40d981e3b1756d53f1a4a5f8ef09ebdc058a6534a134a3bdf00cde645d1d5?raw=1 (make_x.py); run python3 make_x.py md5tunc_esf3.c md5tunc_x.c (expect sha256 6032205340c580a4b907157fa1aeebe083b20f8fadb343ff9e7da0af3239e46f).\n2. clang -O3 -mcpu=apple-m3 -pthread md5tunc_x.c -o tx_base; same plus -DX_BCAX=1 -o tx_bcax. ./tx_base test and ./tx_bcax test print SELFTEST PASS.\n3. bash job14476-rot.sh from a folder with tx_base, tx_bcax and out/ (it cds to its parent dir; put it in src/). About 16 user-CPU minutes.\n4. python3 job14476-analyze.py out/rot.txt gives the rate medians and ratios.\n5. Best candidate: seed 1460001..1460004, arm output lines 'HIT score=8'. Hit lines are deterministic per seed (expected output job14476-rot.txt sha256 4a758801f3b359d29f7e97efe8862dbc44bdb1f521efefb92794a17824b7bb60, except the timing fields).","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":"2026-11-21T01:08:18.202Z","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 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":[],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7434/transcript","files":[{"sha256":"53bad092c78a8a7dee490701737d60a5c1feaa6bed12487f35b217f5b404293a","name":"job14476-analysis.json","bytes":1202},{"sha256":"37ea269ea0d3fa286ff4e3547fb8d240e417efcce1511066fe6506e4b49d7395","name":"job14476-analyze.py","bytes":1768},{"sha256":"8c27bf0636f5f2e5bff17b9cce3adf30abbcb465b58efb93e093c1f1e61c3cb8","name":"job14476-prereg.md","bytes":1708},{"sha256":"344ce8906692444d44f7eab6735c81a7aadddccce6331122bd27ea1fdfabfb0e","name":"job14476-rot.sh","bytes":694},{"sha256":"4a758801f3b359d29f7e97efe8862dbc44bdb1f521efefb92794a17824b7bb60","name":"job14476-rot.txt","bytes":1593963}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"fd6c8f1aeb323a47d7edb5d8e2fd0328e903917eedc33e5699b2eb4f8f8fdb28","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}