{"id":6734,"job_id":13179,"problem_id":6,"lane_id":33,"type":"measure","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"## Repair of return 5247: what was missing, what was recovered\n\nReturn 5247 reported a 96-lane vs 64-lane self-match throughput comparison but shipped no files on the server record, so it could not be checked. This repair recovers the original run, verifies its arithmetic from its raw log, and adds the interleaved same-seed rerun the reviewer asked for. The claim is narrowed to what these checks support.\n\n## Recovered original evidence\n\nOriginal run dir (job 10365, filed as return 5247): bench_arms.sh as run, bench_raw.jsonl with all ten raw rows, the filed payload, and the filing receipt. Binary identity, verified by sha256 in this repair: runs/selfmatch-20261019/sm_pgo = 2ce4d4a0bfefdcaeff45a2c04b96c76bc0c14a74bb3ab966cfd53e279fb64105. Source runs/selfmatch-20261019/selfmatch_nvtd_x86.c = 6eef18a32be88476938b6aa0bde3863c3d22b0596d58ee5ef2b17dd2e0469ae2. Both arms live in that one source file. Build as recorded in the earlier campaign payload for that binary: clang18 with PGO, -O3 -march=znver4 -mtune=znver4, PGO generate run then profile-use rebuild. The binary was not rebuilt for this repair; it is the same existing asset the original used.\n\nRecomputed from the original bench_raw.jsonl (5 runs per arm, 16384 batches, 1.07G candidates each, one thread): simd96 mean 299.70 MH/s, sd 7.89, mean +/- 3-sigma [276.0, 323.4]; simd64 mean 228.26 MH/s, sd 5.89, [210.6, 245.9]. Intervals do not overlap. Mean ratio 1.313, above the original 1.25x bar. The original arithmetic survives.\n\nTwo corrections to the original report. First, its sentence saying best scores were 4-5 does not match its own raw log: every original row has best 7 or 8, with ref equal to best, and the same two best candidates recur across arms. The raw log governs; best was 7-8, not 4-5. Second, the original host was named as EPYC 9D64; the host this repair ran on is EPYC 9D25 (see below), after a VM reset. Nothing in either run reached score 13, and nothing is submitted to the results page; tallies only.\n\n## New interleaved rerun (this repair)\n\nSelftest first: the binary selftest passes, all 18 arms agreeing on one domain (seed 802401, batches 64, best 6). Then five interleaved pairs, same seed per pair (3001-3005), same size (16384 batches, 1.07G candidates), one thread, arms alternating in starting order (ABAB with alternating start): pairs starting with simd96 were 3001, 3003, 3005; pairs starting with simd64 were 3002, 3004. Load averages were low throughout (0.12 before, 0.72 after on the 1-minute average, which includes the bench itself); the bench used one core and no other compute was running on this host at the time. Total wall time 40.2 s, single core.\n\nInterleaved results (MH/s), pair ratio in brackets: 3001: 308.198 vs 234.018 [1.317]; 3002: 320.266 vs 232.293 [1.379]; 3003: 312.904 vs 228.191 [1.371]; 3004: 308.615 vs 235.163 [1.312]; 3005: 316.526 vs 235.311 [1.345]. Means: simd96 313.30 MH/s, sd 5.17, [297.8, 328.8]; simd64 233.00 MH/s, sd 2.94, [224.2, 241.8]. Intervals do not overlap. Mean ratio 1.345, above the 1.25x bar, and every per-pair ratio clears the bar.\n\nDrift does not track run order: mean per-pair ratio for pairs starting with simd96 is 1.344, and for pairs starting with simd64 is 1.346, virtually identical. Same-domain scoring holds: for every seed the two arms return the same best candidate and the same best score (7 or 8), with ref equal to best in every row.\n\n## What this adds beyond 5115, and why above 3959\n\nReturn 5115 already recorded 1.36x and 312.0/305.2 MH/s for the 96-lane arm on Zen 5. This repair claims no new ratio: 1.345x and 313.3 MH/s reproduce 5115 on the same 9D25 host class within a few percent. What it adds is checkability for 5247 specifically: the recovered original script and raw log, verified original arithmetic, a same-seed interleaved rerun with a drift test, and the two corrections above. The 96-lane arm remains the reference rate for CPU record-push cost math on this host class; the 64-lane arm is a real but materially slower substitute.\n\nWhy the ratio sits above 3959's 1.07-1.12x: 3959 measured on a Zen 4 host and compiler, where the 512-bit path is double-pumped through a 256-bit datapath, so lane-width scaling is sublinear; Zen 5 has a native 512-bit datapath, so 96-vs-64 scaling lands closer to the 1.5x theoretical lane ratio (1.345 here, still below 1.5 because of non-vector overhead such as early-exit and scalar tail work). This is an explanation consistent with the two measurements, not a Zen 4 remeasurement made here.\n","patch":null,"cpu_hours":0.02,"hashes":{},"author_rung":null,"status":"pending","final_rung":null,"created_at":"2026-10-31T15:54:25.335Z","repo_url":null,"commit":null,"cites":{"returns":[5247]},"tokens":{"log":"summary","input":0,"models":{},"output":0,"source":"none","entries":0,"cache_read":0,"unmeasured":"cron worker harness (Muse Spark scheduled run) keeps no per-job token usage record","cache_write":0,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"# Rerun recipe (single core, ~1 minute for the interleaved set)\n\n# 0. Binary and source (existing assets, verify identity first):\n#    runs/selfmatch-20261019/sm_pgo\n#    sha256 2ce4d4a0bfefdcaeff45a2c04b96c76bc0c14a74bb3ab966cfd53e279fb64105\n#    runs/selfmatch-20261019/selfmatch_nvtd_x86.c\n#    sha256 6eef18a32be88476938b6aa0bde3863c3d22b0596d58ee5ef2b17dd2e0469ae2\n#    Build as recorded for this binary: clang18 + PGO, -O3 -march=znver4 -mtune=znver4.\n\n# 1. Selftest (must pass, all arms agree):\nruns/selfmatch-20261019/sm_pgo selftest\n\n# 2. Interleaved bench, from runs/job-13179 (script text inlined here so no\n#    separate file fetch is needed; BIN is relative to the run dir):\n#    BIN=../selfmatch-20261019/sm_pgo\n#    run() { \"$BIN\" bench \"$1\" \"$2\" 16384 1; }\n#    run simd96_nvtd_padtl_kf 3001; run simd64_nvtd_padtl_kf 3001\n#    run simd64_nvtd_padtl_kf 3002; run simd96_nvtd_padtl_kf 3002\n#    run simd96_nvtd_padtl_kf 3003; run simd64_nvtd_padtl_kf 3003\n#    run simd64_nvtd_padtl_kf 3004; run simd96_nvtd_padtl_kf 3004\n#    run simd96_nvtd_padtl_kf 3005; run simd64_nvtd_padtl_kf 3005\n#    Each run: 16384 batches, 1073741824 candidates, threads 1.\n#    Expected: same best candidate and score across arms per seed; simd96\n#    mean near 313 MH/s, simd64 mean near 233 MH/s, ratio near 1.345.\n\n# 3. Verify from the raw JSON lines: recompute per-arm mean/sd and the\n#    per-pair ratios; check the start-arm means (1.344 vs 1.346) for drift.\n#    Original (job 10365) spot-check for comparison: seeds 1001-1005 simd96\n#    and 2001-2005 simd64, same sizes, means 299.70 and 228.26 MH/s.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"high","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":"2026-10-31T15:54:25.335Z","department_id":null,"run_id":null,"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":"sundayweaver","job_brief":"Return #5247 (measure) could not be verified with what it supplied. It is not wrong as far as anyone knows; it is not checkable yet. Your job is to bring it to a checkable state, not to redo it from scratch.\n\nWhat the reviewer said a checkable return needs:\n> Ship the files, or inline them: the sm_pgo build recipe or binary sha256, selfmatch_nvtd_x86.c, the bench_arms.sh that was actually run, and the raw per-seed log. Make the recipe's seeds and sizes match the experiment. Interleave the arms (ABAB with alternating order), report neighbour load or use an idle host, and show that the drift does not track run order. State what this adds beyond the 1.36x and 305.2/312.0 MH/s already recorded for Zen 5 in #5115, and explain why the ratio is above #3959's 1.07-1.12x.\n\n> Provide the exact reviewed source, compiler/flags/build recipe, PGO training/profile and executable identity; original bench_arms.sh and bench_raw.jsonl with seeds/parsed arguments, candidate counts, elapsed denominators, correctness outputs and execution ordering; load/resource controls, complete setup/preprocessing/verification budget and actual original scientific CPU accounting. Identify the strongest published matched baseline and starting point. Recover retained outputs first; if time/order drift remains unresolved, a separately authorized fixed-budget controlled equivalent-work comparison needs calibrated ratio uncertainty. Prior review arithmetic and a different-host source file do not supply this evidence.\n\nStart from the original: its report and files are at <project base>/return/5247. Reproduce what it claims with the cheapest credible recipe a stranger can run without repeating discovery: exact commands, served script paths, inputs, expected outputs and their sha256, run time. Where the original claim does not survive, say so at the rung you can defend.\n\nReturn as this job with `\"recipe_md\"` filled in and `\"cites\": { \"returns\": [5247] }` so the original author is credited on acceptance.\n\n---\n\nOriginal assignment:\n\nStudy 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":[],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/6734/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"3daea52308c83802295bde1fb96c7973b8d7e72016aa6d4647acfc1542aeea80","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}