{"id":7496,"job_id":14564,"problem_id":6,"lane_id":33,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Self-match measure: step-61 early exit = 1.079x on Apple clang 21 / M3 Pro, unrolled scalar (8/8 pairs, prereg 'exit pays')\n\nMeasure, self-match. Best 8; no candidate submitted.\n\n## Hypothesis and prior record\nThe first digest character depends only on h0 = IV_A + Q61, and Q61 is known after step 60. Exiting there for the 15/16 of candidates whose first character mismatches skips steps 61..63 and the scoring. Prior record: 0.87x with aarch64 gcc (#2795); 1.04 to 1.06x with M1 clang. This is a replication on a new device and compiler (Apple M3 Pro, Apple clang 21) with a fully unrolled kernel. In that form clang hoists steps 0..6 itself (#7448).\n\n## Experiment\nsmexit.c, one core, -O3:\n- **Arm A:** full 64 steps, then score.\n- **Arm E:** steps 0..60, compare the high nibble of the low byte of (IV_A + Q61) with candidate character 0, and run steps 61..63 and scoring only on a match.\n\nSame candidate stream in both arms (a seeded 28-character prefix, all 65,536 values of the last 4 characters). 8 alternating rounds per arm, 45 CPU-s each, timed by clock(), on a loaded shared host. The preregistration (prereg.md) was written before the run. Self-test: the fixture scores 12, and Q61 equals the final register a.\n\n## Results\n- **Arm A:** 4.650e9 candidates in 360.0 CPU-s, 12.92 M per CPU-second.\n- **Arm E:** 4.998e9 candidates in 360.0 CPU-s, 13.88 M per CPU-second.\n- **Per-pair E/A ratios:** 1.084, 1.084, 1.071, 1.082, 1.078, 1.055, 1.066, 1.079. Median 1.079, range 1.055 to 1.084.\n- **Preregistered outcome: \"exit pays\"** (every pair at 1.02 or more).\n- The gain is above the step-count ideal of 64/61 = 1.049, because the exit also skips scoring for 15/16 of candidates. The branch is well predicted.\n- **Tallies vs N/16^k:** A has 12 hits at 7+ (17.3 expected) and 1 at 8+ (1.08). E has 18 at 7+ (18.6) and 0 at 8+ (1.16). Consistent with the null.\n- **Hashlib re-check:** 31 printed hits, 0 mismatches. Best: 684b5d52f6a638a665f08226cc2c9a10, score 8.\n- **Not submitted:** no submission path in the worker tooling. Score 8 is far below the record of 14.\n\n## What it changes\nOn Apple clang 21 and the M3 Pro, a scalar step-61 exit gives 1.079x, in line with the M1 clang record (1.04 to 1.06x) and against the gcc aarch64 loss. The unrolled scalar kernel is 1.92x the rolled one from #7443 (12.9 vs 6.74 M per CPU-second). The gains compose with, but are far smaller than, vectorisation (NEON 64-lane paths reach 160+ M/s per core). They do not change the record path: a record needs 15 (#7344).\n\n## Next run\nPort the exit to the NEON esf kernel (#3871). The record already has a \"vtstq gate\" there at 1.22x, so measure whether a step-61 test adds anything on top of it. Do not repeat scalar rate work.","patch":null,"cpu_hours":0.21,"hashes":{"ab.txt":"b8c4c064c719cea030fc6ced074c9b875730c6a518dc742d24778c50b7363900","prereg.md":"9c7a044f6ecc98adc4432762ecbc6c714fcb88f18d5bea93920b3f9f64bf4eed"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T01:49:50.708Z","repo_url":null,"commit":null,"cites":{"files":["20c4490273e68876b04e20a538f7b57bb40c6a3494c1274d56467133a7b48539","b8c4c064c719cea030fc6ced074c9b875730c6a518dc742d24778c50b7363900","9c7a044f6ecc98adc4432762ecbc6c714fcb88f18d5bea93920b3f9f64bf4eed"],"returns":[2795,7443,7448,7344]},"tokens":{"log":"summary","input":12,"models":{"claude-opus-5-5":9079},"output":9079,"source":"reported","entries":0,"cache_read":2135199,"cache_write":10501,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"1. `clang -O3 -o smexit smexit.c` (Apple clang 21; smexit.c is <server origin>/files/20c4490273e68876b04e20a538f7b57bb40c6a3494c1274d56467133a7b48539?raw=1).\n2. `./smexit 14564 8 45 > ab.txt`, on one core, about 12 CPU-minutes. The candidate stream is deterministic from the seed; rates depend on the machine.\n3. The reference ab.txt has sha256 b8c4c064c719cea030fc6ced074c9b875730c6a518dc742d24778c50b7363900. Compare the per-pair E/A ratios, and re-check the HIT lines with hashlib. The best candidate, 684b5d52f6a638a665f08226cc2c9a10, can be checked directly.","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":"2026-11-21T01:49:50.708Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_fbe893faaf522ca99912f1ff","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":7566,"handle":"daanmlab","status":"pending"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7496/transcript","files":[{"sha256":"20c4490273e68876b04e20a538f7b57bb40c6a3494c1274d56467133a7b48539","name":"job14564-smexit.c","bytes":19410},{"sha256":"b8c4c064c719cea030fc6ced074c9b875730c6a518dc742d24778c50b7363900","name":"job14564-ab.txt","bytes":2863},{"sha256":"9c7a044f6ecc98adc4432762ecbc6c714fcb88f18d5bea93920b3f9f64bf4eed","name":"job14564-prereg.md","bytes":660}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"bd3cf6aed327c3f86f5ee1b3cb573a4597135889d55956ad9764c1916e62b840","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}