{"id":6999,"job_id":13792,"problem_id":6,"lane_id":34,"type":"explore","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"# All zeros: return #5477 read in full — the CS25 CPU arm and the still-open same-machine head-to-head\n\nJob 13792, explore, all-zeros. Improvement claim: none. No input scoring 13 or above was found; nothing goes to the results page.\n\n## Comparison of the accepted scoped answers\n\nThe accepted CS25 line is returns #5420 (collapsed-base synthesis), #5473 (cs25 kernel), #5481 (Q16 gate decoupling and carry-free bit partition), and #5477 (Step-60 carry-gap theorem and the double-buffered OpenCL pipeline). Returns #5420, #5473 and #5481 were read in full in the prior cycle; this cycle return #5477 was read in full for the first time (its report is 20,245 characters; the full return JSON never completed after repeated truncated fetches, so the report was extracted from the longest partial). No accepted return newer than #5481 appears in the research context, so the accepted set is unchanged since the last comparison.\n\nWhat #5477 adds, beyond the summary known before:\n\n- Theorem 1, verified exhaustively over all 2^32 values of the step-60 word in 10.68 ms: at even thresholds (k = 2, 4, 6, 8) the gate mask is a contiguous low-bit mask, so pre-addition mask inversion is lossless (100% recall, 0 false positives). At odd thresholds (k = 1, 3, 5, 7) a 4-bit nibble gap breaks the fold: naive inversion keeps only 40 of 68 thr-7 hits (58.82%) in the reported data, but keeps 6 of 6 score >= 8 hits. For a record push at score 13, the even-threshold gates are the safe ones.\n- CPU arm: cs25 on Zen 4 AVX-512, clang 21.1.8 -O3 -march=znver4, 8 threads, 5,317.42 +- 6.16 M/s, with 15/15 exact hit-set parity across the cs25 variants tested.\n- GPU arm: 92.5122 +- 1.0930 GH/s end-to-end on an RTX 4070 SUPER at 99.9672% pipeline efficiency.\n- Five score-9 witnesses, hashlib-verified. The top score in the whole CS25 line stays 9, against the record of 13.\n\n## Uncovered obligation\n\nA CPU-only, same-machine head-to-head of the published cs25 tunnel against the production interleave-4 tunnel. The headline numbers invite a wrong conclusion: #5477's CPU arm at 5,317 M/s over 8 threads is about 665 M/s per thread, against 236 M/s per core for the production il4 binary measured on this VM. That looks like ~2.8x, but it is not a comparison: different compiler (clang 21.1.8 vs gcc 13.3), different machine, and cs25 is a tunnel that skips most rounds per candidate, so candidates per second do not compare directly. Earlier local work also found a scalar CS25 tunnel does not beat generic search on CPU-only, so the headline number needs the head-to-head rather than a quote.\n\n## Cheapest decisive check, and why it is still open\n\nFetch the published cs25 source (job11081-md5tunc_cs25.c), build it with gcc -O3 -march=znver4 on this VM, and bench it against il4 on identical nonce ranges. Attempted this cycle and blocked on transport: the file fetch truncates at 48-58 KB on every attempt, the server ignores Range headers, and the best partial prefix saved (58,255 bytes) ends mid-function, so it is not compilable. That partial is kept with this return's work files for the next cycle. clang 21.1.8 is also not installed on this VM (gcc 13.3 only), so an exact replication of the published claim would need a clang install as well.\n\n## Experiment run this cycle\n\nA fresh disjoint 20-billion-candidate control cell with the production binary tools/md5tun_avx512_il4 (verified by a 10M smoke run at 228.42 M/s solo), from nonce 13269850000000, pinned to core 1 while the parallel measure cell ran on core 0: 172.30 M/s under parallel load, best score 8. All 8 new-best hits were hashlib-verified from the log, 0 mismatches. The expected best for 20B candidates is about 8.3, so 8 is consistent with the uniform model. The best-8 input is 556c5af2e9c0ffb013e4a6251094cc829b0123f0a758255a98b31204575cd565c42e22f78213579279209e72f1b148977b9caf91, digest 000000003b123a5c91a8664e434fbcec.\n","patch":null,"cpu_hours":0,"hashes":{},"author_rung":null,"status":"recorded","final_rung":"recorded","created_at":"2026-11-05T23:44:22.992Z","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":"# Recipe\n\nTo repeat:\n\n1. Fetch return #5477 via GET /projects/md5/return/5477, retrying on IncompleteRead and extracting report_md from the longest partial if the full JSON never completes. Compare against the complete local copies of returns #5420, #5473 and #5481.\n2. Try fetching the cs25 source at GET /files/abf8ac7b6ebfa9ad1843c551333900f9fe0cdfc77465a2fddc5551771eabef09. If it truncates, keep the longest partial; it is not compilable unless it ends cleanly.\n3. Verify the production binary with a 10M smoke run: taskset -c 1 ./tools/md5tun_avx512_il4 0 10000000 5.\n4. Run a disjoint control cell: taskset -c 1 ./tools/md5tun_avx512_il4 13269850000000 20000000000 60, logging to the run dir.\n5. Verify every [new best] line in the log by recomputing md5 with hashlib and comparing digests.","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/6999/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"c2d88d0e6ff707d1b0626b5096ccbc6f0398f867a750f082db1d71dd5b03da75","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}