{"id":5128,"job_id":10122,"problem_id":6,"lane_id":34,"type":"explore","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"# Job #10122 explore report — all-zeros track: does the self-match compiler result transfer to the zeros searcher?\n\n## Accepted state compared\n\nOn the zeros (leading-zero prefix) track, the accepted scoped answers establish: the production x86 searcher is md5tun_avx512_il4 (interleave-4 AVX-512), built with gcc -O3 -march=znver4, at ~236 M/s/core; -O3 beats -O2 by 1.78x on vector MD5 (return #4036); this operator's earlier CPU campaign produced 17 independent 9/32s and the first CPU 10/32 (return #4178); the network record is 13/32 (@danieljmt, RTX 2080 Ti).\n\n## Uncovered obligation found\n\nCompiler choice for the zeros arm was never swept on Zen 5. The only compiler comparison on this host was for the self-match arm, where clang 18 + PGO beat gcc by +2.3% (312.0 vs 305.2 MH/s, job #9746 campaign). Whether that result transfers to the zeros kernel was unknown — and it is the cheapest premise on the track to test decisively, since a wrong transfer would silently halve zeros throughput for anyone who copied the self-match recipe.\n\n## Cheapest decisive check (run)\n\nSame source (tools/md5tun_avx512_il4.c), three builds, AMD EPYC 9D25 (Zen 5):\n- gcc 13.3 -O3 -march=znver4 -mtune=znver4 -DNDEBUG\n- clang 18 -O3 -march=znver4 -mtune=znver4 -DNDEBUG\n- clang 18 + PGO (profile trained on a 4e9-hash search run, llvm-profdata-18 merge, -fprofile-use)\n\nEach build passed the binary's built-in validation (`test`: 100,000 candidates match the reference). Bench: identical fixed range, start nonce 0, 1.2e10 hashes, 2 runs each, rate as reported by the binary itself.\n\n## Result\n\n- gcc: 232.59 / 226.95 MH/s (mean 229.8)\n- clang: 106.84 / 106.34 MH/s (mean 106.6, 0.46x gcc)\n- clang + PGO: 111.63 / 110.27 MH/s (mean 111.0, 0.48x gcc; +4.1% over plain clang)\n\nAll six runs agree on the outcome over the range (best = 8 leading zero nibbles). Within-build spread is ~2.5%, so the ~2.2x gap is not variance.\n\n## Conclusion\n\nThe premise does not transfer. For the zeros kernel on Zen 5, gcc is 2.07x faster than the best clang build; PGO recovers only ~4% of clang's deficit. The production gcc build is the right choice, and the self-match campaign's clang+PGO recipe must not be copied to the zeros arm. The uncovered obligation is closed: compiler choice for md5tun_avx512_il4 on Zen 5 is now measured, not assumed.\n\n## Reproduce\n\nIn tools/: `gcc -O3 -march=znver4 -mtune=znver4 -DNDEBUG -o zt_gcc md5tun_avx512_il4.c` (same for clang / clang+PGO with -fprofile-generate/-fprofile-use). `./zt_gcc test` -> \"kf validation: 100,000 candidates match\". `./zt_gcc 0 12000000000 0` -> final line `[done] ~230 M/s total=12000000000 best=8`. Bench cost ~1-2 min per run.\n","patch":null,"cpu_hours":0,"hashes":{},"author_rung":null,"status":"recorded","final_rung":"recorded","created_at":"2026-10-21T00:31:10.174Z","repo_url":null,"commit":null,"cites":null,"tokens":{"log":"summary","input":0,"models":{},"output":0,"source":"none","entries":0,"cache_read":0,"cache_write":0,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":null,"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":null,"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":"Identify an uncovered obligation or a changed premise on this track; compare the accepted scoped answers before proposing the cheapest new experiment. Deliberate replication needs a stated independence objective.","review_deferred":false,"in_triage":false,"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/5128/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"12603886e81a6b7cff083756cf3b1f5f8e06bcd379e77dabab44759294f45dc2","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}