{"id":7031,"job_id":13841,"problem_id":6,"lane_id":35,"type":"measure","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"# Smallest collision: 12121 advantage rebounds on fresh seeds (2.77x); 126 B projection recomputed to about 2^27.3\n\nJob 13841, measure, smallest-collision. Improvement claim: none. No\ncandidate scoring 13 or above was found, so nothing goes to the results\npage. No pair below 128 bytes was found; the best depth reached was 30,\nand no record is a collision.\n\n## Hypothesis\n\nThe narrowed ratios from #13826 (12121 1.94x, p79 1.26x vs served, down\nfrom 2.59x/1.59x in #13813) replicate on fresh seeds, or were seed\nnoise; the 126 B projection is recomputed from the fresh rates either\nway. If the narrowing was noise, the ratios should bounce back toward\nthe #13813 values.\n\n## Experiment\n\ngxr_fast pilots, 3 fresh seeds (13841-13843) x 100 s x 1 thread per arm\n(300 thread-s per arm), one core. Same binaries and headers as #13826 /\n#13813 (gcc -O3 -march=znver4 -mtune=znver4). Three arms:\n\n| Arm | Path | Q27/thread-s (seeds) | Q28 | Q29 | Best depth |\n|---|---|---|---|---|---|\n| 12121 | 126 B, T9 25 | 216.68 (196.84, 231.78, 221.42) | 223 | 9 | 30 |\n| p79 | 126 B, T9 24 | 127.59 (129.80, 130.11, 122.87) | 118 | 6 | 29 |\n| served | 128 B, T9 23 | 78.10 (84.85, 77.99, 71.47) | 68 | 3 | 29 |\n\nRatios vs served on Q27: 12121 2.77x, p79 1.63x. Prior cycles: #13813\n(seeds 13813-13815) measured 2.59x and 1.59x; #13765 measured 2.55x\nand 1.46x; #13826 measured 1.94x and 1.26x.\n\n## Finding\n\nThe narrowing in #13826 was seed noise, not a real decline: on fresh\nseeds the ratios rebound to 2.77x and 1.63x, within 0.10 bit (12121) and\n0.04 bit (p79) of the #13813 values and above them. The ordering\n12121 > p79 > served replicates for the fourth consecutive seed set.\nThe seed spread on 12121 is sd 14.7 (about 7%), so single-seed\ncomparisons still need roughly half a bit of margin; the #13826 dip\n(1.94x) sits about 0.4 bit below the other three sets and is the\noutlier, not the trend.\n\n## Projection\n\nSwapping the fresh-seed local Q27 ratio into 4373's projection method\nmoves the 126 B cost on 12121 from about 2^27.4 (#13813, #13765) to\nabout 2^27.3 thread-s per pair, still below the served 128 B cost of\n2^27.96. The change is the ratio move only (log2(2.77/2.59) = 0.10\nbit). The remaining gap is still the generator: even at the local rate,\na 126 B pair is on the order of 2^31 thread-s with gxr as it is, and\nthe route 305 generator step is what closes it.\n\nVerification: 417 records checked independently (best record per seed\nplus every d=28/d=29 log record, depth/padding/hashlib checks via\nverify_best.py), 417 OK, 0 bad. Q29 counts: 12121 9 vs served 3 (3.0x),\nconsistent with the Q27 ratio.\n","patch":null,"cpu_hours":0,"hashes":{},"author_rung":null,"status":"pending","final_rung":null,"created_at":"2026-11-06T14:54:37.595Z","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, on one core:\n\n1. From runs/job-13841/work/, run `bash ../run_arms.sh` (it expects the\n   gxr_12121, gxr_p79, and gxr_served binaries in that directory; they\n   are copies of the #13826 binaries, built gcc -O3 -march=znver4\n   -mtune=znver4). It runs 3 arms x 3 seeds (13841-13843) x 100 s and\n   writes gxr_<arm>.jsonl plus per-seed logs.\n2. Aggregate Q27/thread-s from the {\"prog\":...} summary line of each\n   log (q27 field divided by 100 s), pool per arm, and divide by the\n   served pool for the ratios.\n3. Verify with `python3 verify_best.py` in the work directory (checks\n   best-per-seed records plus every d=28/d=29 log record).\n4. To recompute the projection, swap the fresh 12121/served Q27 ratio\n   into 4373's projection method as #13813 did; the served 128 B\n   baseline is 2^27.96 thread-s per pair.","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":"2026-11-06T14:54:37.595Z","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":"Study how MD5 collisions are built (differential paths, message modification, the single-block attacks of Xie and Feng and Stevens) and what limits their length, and use it to find a shorter full collision. Running fastcoll gives 128 + 128 bytes from known techniques; it is the baseline to measure against. Ideas to test: where the single-block attacks spend their work, whether a shorter second member or a shared prefix can change the bound, what a 64 + 64 search would cost on your machine. 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/7031/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"9a4869febc677fa7b367f13080385bb1198097005a02c13942ec25ec24fc2a30","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}