{"id":7566,"job_id":14608,"problem_id":6,"lane_id":34,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All-zeros measure: the step-61 exit gives no gain with cheap scoring (E/A 0.997, 8/8 pairs within ±2%); #7496's self-match gain came from skipped scoring\n\nMeasure. Best 8; not submitted.\n\n## Hypothesis\nReturn #7496 measured a 1.079x step-61 exit gain on the self-match kernel (Apple M3 Pro, Apple clang 21, unrolled scalar). That is above the step-count ideal of 64/61 = 1.049, which suggests most of the gain came from skipping the scoring for 15/16 of candidates, not from the 3 skipped steps. Test: run the same kernel with all-zeros scoring, a cheap leading-zero nibble count instead of a character compare. If skipping 3 steps is what pays, the ratio should stay near 1.05.\n\n## Experiment\nzexit.c is #7496's smexit.c with two changes: the score counts leading zero hex characters, and the exit test is \"first nibble of h0 = IV_A + Q61 is 0\". Everything else is identical: candidate stream, unroll and A/B design. 8 alternating rounds per arm, 45 CPU-s each, on one core of a loaded host. The preregistration (prereg.md) was written before the run. Self-test: the all-zeros fixture scores 13. The compiler is the Command Line Tools clang 21.0.0 (build 2100.1.1.101, the same build as #7496), because the Xcode clang driver began refusing to run (an unaccepted Xcode licence).\n\n## Results\n- **Arm A (full):** 5.093e9 candidates in 360.0 CPU-s, 14.15 M per CPU-second.\n- **Arm E (exit):** 5.066e9 candidates in 360.0 CPU-s, 14.07 M per CPU-second.\n- **Per-pair E/A ratios:** 1.003, 1.004, 0.995, 1.002, 0.991, 0.984, 0.982, 0.998. Median 0.997, range 0.982 to 1.004.\n- **Preregistered outcome: \"no gain\"** (every pair within 0.98 to 1.02).\n- **Tallies vs N/16^k:** A has 16 hits at 7+ (19.0 expected) and 0 at 8+ (1.19). E has 19 at 7+ (18.9) and 1 at 8+ (1.18). Consistent with the null.\n- **Hashlib re-check:** 37 hits, 0 mismatches. Best: the 32 ASCII bytes 63aeea722e8503e60c0ea0fdfdcd678e, score 8. Not submitted: no submission path in the worker tooling, and far below the record of 14.\n\n## What it shows\nOn this core and compiler, skipping steps 61..63 for 15/16 of candidates saves nothing measurable. The branch and the loss of overlap cost about what the 3 steps would. The self-match gain of #7496 comes from avoiding the per-candidate character-compare scoring. So the exit pays only where scoring is expensive.\n\nThe all-zeros playbook's \"exit code\" trick (#2821/#3474: a hoisted nibble mask on (IV_A + a)) is the right form: make scoring cheap, and do not branch early. This agrees with the 1.04 to 1.06x M1 clang record only for kernels with costly scoring.\n\n## Next run\nNone on scalar CPU. Vectorised kernels already score with masks. The record needs 15 (return #7346), which is GPU work.","patch":null,"cpu_hours":0.21,"hashes":{"ab.txt":"940d858740cf1bb258973a764e870394bdc5b500c0da6339e52964a85ef57e35","prereg.md":"597db4e1a32888024d84fe8d94a2aaa3ff5e6ffe7cbbb1fbdb8d78c573b2cf90"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-21T02:35:00.335Z","repo_url":null,"commit":null,"cites":{"files":["a423d7731da95739e1a11149837486a69db72d84c386abe47c9a36adf8c478e6","940d858740cf1bb258973a764e870394bdc5b500c0da6339e52964a85ef57e35","597db4e1a32888024d84fe8d94a2aaa3ff5e6ffe7cbbb1fbdb8d78c573b2cf90"],"returns":[7496,7346,2795,2821,3474]},"tokens":{"log":"summary","input":18,"models":{"claude-opus-5-5":8664},"output":8664,"source":"reported","entries":0,"cache_read":3339198,"cache_write":10602,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"1. `clang -O3 -o zexit zexit.c` (Apple clang 21.0.0, build 2100.1.1.101; zexit.c is <server origin>/files/a423d7731da95739e1a11149837486a69db72d84c386abe47c9a36adf8c478e6?raw=1).\n2. `./zexit 14608 8 45 > ab.txt`, on one core, about 12 CPU-minutes.\n3. The reference ab.txt has sha256 940d858740cf1bb258973a764e870394bdc5b500c0da6339e52964a85ef57e35. Compare the per-pair E/A ratios, and re-check the HITs with hashlib (md5 of the 32 ASCII bytes, leading zero hex characters).","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":"2026-11-21T02:35:00.335Z","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 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/7566/transcript","files":[{"sha256":"a423d7731da95739e1a11149837486a69db72d84c386abe47c9a36adf8c478e6","name":"job14608-zexit.c","bytes":19411},{"sha256":"940d858740cf1bb258973a764e870394bdc5b500c0da6339e52964a85ef57e35","name":"job14608-ab.txt","bytes":3199},{"sha256":"597db4e1a32888024d84fe8d94a2aaa3ff5e6ffe7cbbb1fbdb8d78c573b2cf90","name":"job14608-prereg.md","bytes":649}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"3ecdb3e1dd32a172694c4859f9d1424a5ddd370c71abf043155de5ff8c1ee1fb","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}