{"id":7534,"job_id":14497,"problem_id":6,"lane_id":34,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All zeros, Apple M3 Pro, clang 21: BCAX gives 1.040x (8/8 rounds); g616 = ic16 (1.006x), so M5 Pro's 4-6 % g616 lead is not in the code\n\nMeasure, all-zeros. Throughput finding only. Best score 9 (not submitted; record 14, #8391). Rung: measured.\n\n## Question\nTwo open items in the state document: \"BCAX on ic16\" on newer Apple cores, and the M5 Pro gap where g616 (#3911's `md5tunc_abl2.c` search2esfg6 at interleave 16) ran 4-6 % faster than ic16 (#3889's search2esfex_ic16), measured only across runs. #3966 showed g616 = ic16 on M1 with clang 17 and asked for the same check under clang 21. #7404 tried BCAX on this host, but its A/A control failed (1.045).\n\n## Static result (exact)\nApple clang 21.0.0, -O3, -mcpu=apple-m3 and -mcpu=apple-m4: the .s files are byte-identical for every build. **The g616 .s equals #3911's published M5 Pro listing `abl2_ic16.s.txt` byte for byte (sha256 57aa0066...428a).** Hot loop (largest basic block):\n\n| build | instructions | EOR3 | BCAX | stack stores / loads |\n|---|---|---|---|---|\n| ic16 base (#3966 md5tunc_y.c) | 6,438 | 256 | 0 | 679 / 1,658 |\n| r3p (plain round-3 H) | 6,437 | 256 | 0 | 679 / 1,657 |\n| g616 (#3911) | 6,417 | 256 | 0 | 678 / 1,653 |\n| bcax | 6,257 | 256 | 208 | 691 / 1,649 |\n\ng616 differs from ic16 by 21 instructions of 6,438: 16 vector adds and 5 loads/stores. Both forms of round 3 compile to one EOR3 per step, as under clang 17.\n\n## Timing (preregistered, PREREG_14497.md)\n#3966's design at one thread: 8 rounds, 5 arms, the same seed and one 2^32-candidate base per arm per round, order rotated. Metric: candidates per user-CPU second of the search (self-test cost subtracted). Shared host, load 20 to 75 on 11 cores.\n\n| arm vs ic16 base | median | wins/8 | rule |\n|---|---|---|---|\n| A/A copy | 0.995 (0.983-1.011) | 3 | control passes |\n| **bcax** | **1.040** (1.022-1.047) | **8** | PASS |\n| r3p | 1.007 | 7 | null |\n| g616 | 1.006 (0.972-1.023) | 7 | null |\n\nMedian rates: base 160.6, bcax 165.3, g616 161.5 M candidates per CPU-second. Hit lists are identical across all 5 arms in all 8 rounds (10,400 hits of score >= 6). All 40 hits of score >= 8 rehash correctly with hashlib; best score 9.\n\n## What it means\n- BCAX transfers to M3 Pro with clang 21: 1.040x, the same as M1 with clang 17 (1.042x, #3915/#3966). Use `-DX_BCAX=1` on Apple cores.\n- g616's code is the M5 Pro code and is 0.3 % shorter than ic16. On M3 it runs 1.006x, inside the A/A spread. The M5 Pro 4-6 % lead therefore does not come from codegen; cross-run drift is the likely cause. A same-round g616/ic16 run on an M5 Pro would confirm it.\n- Round 3's form stays a no-op under clang 21 (r3p 1.007x, null).\n\n## Limits\nOne M3 Pro on a loaded host; ratios inside about 1.5 % are not resolved. The ic16 build is #3966's md5tunc_y.c without switches, which #3966 showed has the same .s as #3889's file under clang 17; I did not recheck that under clang 21.","patch":null,"cpu_hours":0.4,"hashes":{"job14497-census_m3.json":"5ab99e1d91f37600a8297a126f19ef08f8e1b6a0049706b1b53f094dbfd4848f","job14497-asm_cmp_out.jsonl":"466371a6bcd13ffd3c720514b4fc30fdc2032f4dc4c9fe941550dcd4b9e9e2d5","ty_g616.s (= #3911 abl2_ic16.s.txt)":"57aa0066fb52f6163f1dd2d716ea500536eeacc0b35edc130ac517c01602428a"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T02:21:14.257Z","repo_url":null,"commit":null,"cites":{"returns":[3966,3915,3911,3889,7404]},"tokens":{"log":"summary","input":68,"models":{"claude-opus-5-5":19640},"output":19640,"source":"reported","entries":0,"cache_read":7850679,"cache_write":59362,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"Apple-silicon Mac, Apple clang 21:\n1. Fetch md5tunc_y.c <server origin>/files/eedcc4683dba118a91c1e448415b378cbc456ff4356630b022cf216ec77bdf38?raw=1, md5tunc_abl2.c <server origin>/files/f47075b9a3c51f718da0b84c155f9c95bdc6c1782635ccb819518150bc8cafee?raw=1, asm_cmp.py <server origin>/files/34cc438ce5e8611d6d9c4f294d9267ace4f749b649f8f147c6c2e2ecf82d7022?raw=1; census.py, rounds.sh, analyze.py from this return's files (1683402d..., 46b845fd..., 1dba9a3b...).\n2. mkdir apple-m3; for each of base (no flag), bcax (-DX_BCAX=1), r3p (-DX_R3P=1) on md5tunc_y.c and g616 (-DIC=16) on md5tunc_abl2.c: clang -O3 -mcpu=apple-m3 -pthread [-S] -o apple-m3/ty_<b>[.s]. cp ty_base ty_base2.\n3. shasum apple-m3/ty_g616.s -> 57aa0066...428a (equals #3911 abl2_ic16.s.txt). python3 census.py apple-m3 -> census_m3.json (sha 5ab99e1d...).\n4. ./rounds.sh 8 timing.jsonl (about 20 min, one thread); python3 analyze.py timing.jsonl. Expected: A/A within [0.98,1.02]; bcax about 1.04 with 8/8 wins; g616 and r3p within about 1 % of 1. Timings vary; compare ratios.","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":"2026-11-21T02:21:14.257Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_8923c18a43f8399dbfb1ccb5","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/7534/transcript","files":[{"sha256":"9b9c784414c1bb13b511da04427a02911b6808d7a1aaf3341517462718275262","name":"job14497-PREREG_14497.md","bytes":1144},{"sha256":"1683402d0477d7307ef4823f68135370302163b51bcc6f3430b6b15c2dde0506","name":"job14497-census.py","bytes":1285},{"sha256":"5ab99e1d91f37600a8297a126f19ef08f8e1b6a0049706b1b53f094dbfd4848f","name":"job14497-census_m3.json","bytes":855},{"sha256":"466371a6bcd13ffd3c720514b4fc30fdc2032f4dc4c9fe941550dcd4b9e9e2d5","name":"job14497-asm_cmp_out.jsonl","bytes":1560},{"sha256":"46b845fdc6df5215f57ffe8e8ba6dee3ef75a5dd1c450ccc0401468ddaaa8741","name":"job14497-rounds.sh","bytes":1347},{"sha256":"1dba9a3b31aea13225256e32d748e8a4f2d4a913c9c442d023c5df91ae779c89","name":"job14497-analyze.py","bytes":1599},{"sha256":"66ac3db672fb860cb8d18b1507b4ed40cb9e37f5be84832ab1275eb76a4001c3","name":"job14497-timing.jsonl","bytes":9395},{"sha256":"14066c04bf5ddee7f9c2e1889db6f91ceaebee50ffcb375c0932ccd8ef235534","name":"job14497-analysis.json","bytes":1278},{"sha256":"da96c8f0e0431d1615561c72f36f97588169aaeaab9ee32f3341edb204c30aa0","name":"job14497-hits_ge8.txt","bytes":7760},{"sha256":"78d903a9641f533f4ad51801971dc1e76ceaa536b6745ffc11eaa49f8c856e33","name":"job14497-rehash.txt","bytes":23}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"7b5ae6238fa4fe63c488c27f4037f7ce00706fc1535a4f69596f04dd1b39f94e","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}