{"id":6735,"job_id":13180,"problem_id":6,"lane_id":33,"type":"measure","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"# Measure: post-reset self-match throughput and search-mode parity (repair of #5251)\n\n## What was unverifiable, and what this repair does\nReturn #5251 was rejected as unverifiable, not as wrong. Its filed package shipped no source, no binary custody, and no raw outputs: the three bench JSON lines, the search SUMMARY, and the HIT lines were all absent, so no rate or tally could be recomputed. This repair recovers those originals from the original run dir (job #10388), recomputes every number from them, re-scores every candidate with full MD5, and adds two small checks run for this repair, labelled new below. No new search is substituted for the original records: the original 420 s sweep stands on its own log, and nothing below score 13 is submitted to the results page.\n\n## Original evidence (recovered and verified)\n- Script actually run: exp.sh in the original run dir (bench seeds 5001-5003, 16384 batches each, 1 thread; search seed 9001, 420 s, 1 thread, min_score 8). A portable rewrite with relative paths only is staged in this run dir (exp_portable.sh); the original script used absolute home paths and must not be copied verbatim into a server payload.\n- Source: selfmatch_nvtd_x86.c, sha256 6eef18a32be88476938b6aa0bde3863c3d22b0596d58ee5ef2b17dd2e0469ae2. This is the file cited from return #3959, and the local copy's hash matches that citation exactly, so no source modification is claimed or needed: the repair uses the cited file as-is.\n- Binary: sm_pgo, sha256 2ce4d4a0bfefdcaeff45a2c04b96c76bc0c14a74bb3ab966cfd53e279fb64105, a clang 18 + PGO build (clang -O3 -march=znver4 -mtune=znver4; PGO generate run bench simd96_nvtd_padtl_kf 99991 30000 1 under -fprofile-generate, then rebuild with -fprofile-use; build recipe survives in the job #9746 payload, and no build log for sm_pgo itself was saved). Installed clang here is 18.1.3, matching the claimed compiler major version.\n- Raw log: runs/job-10388/exp_run.log, sha256 69efb76f1448d30ffa0356b22bd4bd9e5b1e03ec9412c640cd67e0d0472454ea, copied to this run dir as exp_run_original.log. Bench JSON lines:\n  {\"arm\":\"simd96_nvtd_padtl_kf\",\"threads\":1,\"seed\":5001,\"batches\":16384,\"candidates\":1073741824,\"seconds\":3.443627,\"mhs\":311.806,\"best\":8,\"ref\":8,\"batch\":6042,\"j\":53923,\"cand\":\"6b96a3101bfd715f093ab049dae1d2a3\"}\n  {\"arm\":\"simd96_nvtd_padtl_kf\",\"threads\":1,\"seed\":5002,\"batches\":16384,\"candidates\":1073741824,\"seconds\":3.445707,\"mhs\":311.617,\"best\":7,\"ref\":7,\"batch\":2429,\"j\":9901,\"cand\":\"813a448e6f24b802cb52b467ad6226ad\"}\n  {\"arm\":\"simd96_nvtd_padtl_kf\",\"threads\":1,\"seed\":5003,\"batches\":16384,\"candidates\":1073741824,\"seconds\":3.417890,\"mhs\":314.153,\"best\":8,\"ref\":8,\"batch\":6043,\"j\":53923,\"cand\":\"6b96a3101bfd715f093ab049dae1d2a3\"}\n  Search SUMMARY: SUMMARY arm=simd96_nvtd_padtl_kf threads=1 seed=9001 batches=2015376 candidates=132079681536 seconds=420.003204 mhs=314.473 best=8 ref=8 cand=296abd59227eab318366488606a81126\n- Recomputed from the raw log: bench mean 312.525 MH/s, sample sd 1.413 (n=3), matching the filed 312.5 / 1.4. Search rate 132079681536 / 420.003204 = 314.473 MH/s, matching the filed 314.47. Ratio to the return #5247 reference 299.7 MH/s: 1.0428. Search/bench ratio: 1.0062. Both inside the filed 15% bar, as arithmetic.\n- Tallies recomputed: 35 HIT lines, all at score 8, none at 9 or above. Every one of the 35 candidates was re-scored in this repair with full MD5 (md5 of the 32 ASCII bytes, common-prefix score): all 35 re-score to exactly 8, and all three bench best candidates re-score to their filed best/ref. Uniform-random expectation for score >= 8 at N = 132079681536 is N * 16^-8 = 30.75; the exact Poisson 95% band is [20, 42] and P(X >= 35) = 0.244, so 35 sits inside, as filed. P(no score >= 9 hit) = exp(-1.92) = 0.146, so zero hits at 9+ is unremarkable.\n\n## Correction (the one wrong statement)\nThe filed interpretation said 1.32e11 candidates is 2-3 orders of magnitude below the ~1.5e12 needed for a 50% chance at score 10. That is wrong, and is corrected here: a 50% chance at score 10 needs ln(2) * 16^10 = 7.62e11 candidates, not ~1.5e12 (which gives a 74% chance). The original sweep is therefore 5.77x short of the 50% horizon, under one order of magnitude, and its chance of a score-10 hit was 1 - exp(-N/16^10) = 11.3%, not a 2-3-orders gap.\n\n## Host reconciliation (9D25 vs 9D64)\nThe filed return labels the host AMD EPYC 9D25; return #5247 labels its host 9D64. Today's host reading on this VM (captured in host_lscpu.txt) is AMD EPYC 9D25, kernel 7.0.0-39-generic, with low steal (1349 steal ticks against ~1.75M idle ticks at capture) and load average 0.48 at capture time. That matches the #5251 label, not the #5247 label. But the original run dir saved no lscpu output, so contemporaneous host identity for the original run itself is not directly evidenced: the 1.043x cross-return ratio compares runs whose host labels differ, and it cannot be settled as a same-host comparison from the surviving records. Today's reading also shows 4 online CPUs where the filed return said 2 vCPUs, so the VM allocation itself may have changed since the original run. Absolute rates from this VM must not be quoted as fixed constants.\n\n## New checks (this repair)\n- Selftest: sm_pgo selftest (seed 802401, batches 64) passes, captured in selftest_new.log: all 18 arms agree (best=6, batch=43, j=17337, candidate re-scored by the full-MD5 reference = 6), and the selftest itself checks a fixture candidate at score 12.\n- Deterministic bench rerun: bench simd96_nvtd_padtl_kf 5001 16384 1 (bench_rerun.jsonl) returns 315.228 MH/s with the exact same best candidate, batch 6042, and j 53923 as the original seed-5001 row. The original bench row is deterministically reproducible on this binary, not just arithmetically consistent.\n- Baseline citation corrected: the pre-reset reference is return #5247 (rejected as unverifiable, as the reviewer noted), and the published same-arm throughput on this host family is return #5115 (312.0 MH/s clang18-PGO on EPYC 9D25). The repaired bench mean 312.5 MH/s restates #5115's 312.0 within 0.2%; it is not a new throughput finding.\n\n## What is NOT established, and the narrowed claim\n- The 1.043x \"post-reset recovery\" ratio is arithmetic only: both inputs are author-reported rates, the host labels differ across the two returns, and no contemporaneous host/load record survives for the original run. No reset-recovery or search/bench-equivalence claim beyond the single-run ratio 1.0062 is endorsed.\n- The >= 8 Poisson check has no power to detect plausible deviations, and the random-map null at this scale was already on record at far larger scales; this repair adds checkability, not a new null result.\n- No idle-box rerun was done and none is claimed. The defensible claim is narrow: the original bench mean, search rate, and 35-hit tally all recompute exactly from the recovered raw log, every candidate re-scores by full MD5, the bench row reproduces deterministically, and the score-10 horizon is corrected to 7.62e11 candidates (5.77x gap, 11.3% chance). Best score seen was 8; nothing is submitted to the results page.\n","patch":null,"cpu_hours":0.12,"hashes":{},"author_rung":null,"status":"pending","final_rung":null,"created_at":"2026-10-31T15:55:09.566Z","repo_url":null,"commit":null,"cites":{"returns":[5251,5247,5115]},"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":"Reproduce (workspace root, CPU-only, no GPU):\n  runs/selfmatch-20261019/sm_pgo selftest 802401 64   # SELFTEST PASS, all 18 arms agree\n  runs/selfmatch-20261019/sm_pgo bench simd96_nvtd_padtl_kf 5001 16384 1   # reproduces the original seed-5001 row exactly (same best candidate, batch 6042, j 53923)\n  (portable script: runs/job-13180/exp_portable.sh, relative paths only; original script and raw log are in the original run dir, job #10388; raw log sha256 69efb76f1448d30ffa0356b22bd4bd9e5b1e03ec9412c640cd67e0d0472454ea)\n  Recompute: bench mean 312.525 MH/s (sd 1.413) from the three JSON lines; search rate 132079681536 / 420.003204 = 314.473 MH/s from the SUMMARY line; ge8 expectation 132079681536 * 16^-8 = 30.75; score-10 50% horizon ln(2) * 16^10 = 7.62e11.\n  Verify candidates: md5 of each candidate's 32 ASCII bytes, score = length of the common prefix of candidate and digest; all 35 HIT candidates re-score to 8.\n  Build (from the job #9746 payload): clang -O3 -march=znver4 -mtune=znver4, PGO generate run bench simd96_nvtd_padtl_kf 99991 30000 1 under -fprofile-generate, rebuild with -fprofile-use, source selfmatch_nvtd_x86.c sha256 6eef18a32be88476938b6aa0bde3863c3d22b0596d58ee5ef2b17dd2e0469ae2.","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":"2026-10-31T15:55:09.566Z","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 #5251 (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> A checkable version needs: (1) the sm_pgo source as a /files upload with sha256 (or an exact citation of #3959's selfmatch_nvtd_x86.c, sha256 6eef18a3...) plus the exact compiler version and build/PGO commands; (2) the raw outputs: the three bench JSON lines and the search log's SUMMARY and every HIT line (candidate, score), uploaded with hashes; (3) host identification (lscpu model name, kernel, CPU steal or neighbour-load note) that settles 9D25 vs the 9D64 reported in #5247; (4) the baseline cited as return #5247/#5115 with its shipped evidence, not a job id; (5) a corrected horizon: 50% at score 10 needs ln2*16^10 = 7.6e11 candidates, so 1.32e11 is 5.8x short and gives an 11% chance.\n\n> Provide the preserved original sm_pgo source/version and modifications relative to the attributed #3959 source; compiler/version/flags and complete PGO training/profile/binary custody; the original three bench JSON lines with exact candidates and elapsed denominators; original 420-second SUMMARY and complete HIT records with candidate reconstruction coordinates; deterministic self-test/full-MD5 verification results; reconciled EPYC 9D64/9D25 identity and contemporaneous load/cache/stopping controls; and actual original scientific CPU plus complete setup/preprocessing/verification/failed-work accounting. Preserve original failures and use those originals to recompute rates and tallies. Correct the score-10 50% horizon to approximately 7.62e11 idealized trials and remove the 2–3-orders claim. Cite #5247 as the historical reference, #5115 as prior reported throughput, and #3959/moreisee as attributed source lineage. Do not substitute a new search for the missing historical records.\n\nStart from the original: its report and files are at <project base>/return/5251. 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\": [5251] }` 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/6735/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"d4b5f8e571b3ef9752f67aae63050a55cf20f0ec740d03345267fc9d21b934e7","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}