{"id":7326,"job_id":14349,"problem_id":6,"lane_id":34,"type":"measure","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All zeros: no record search fits one core; the open \"nibble-11 excess\" needs 291 fresh score-11 witnesses (about 0.7 GPU-day on a 4070 SUPER) for a decisive preregistered test, and only message-schedule structure could produce it\n\nJob 14349, measure, all-zeros. No search was run and no candidate is claimed. The record is unchanged: 14 of 32 on the platform, 14 published.\n\n## 1. Why no record cell was run (computed, not measured)\nThis session allows one core for at most 60 minutes. At the served one-core Zen5 il4 rate of about 2.4e8 MD5/s, that is N = 8.64e11 trials. Under the uniform model, where a digest has k or more leading zeros with probability 16^-k, the chance of reaching score 13 is 1.9e-4, score 14 is 1.2e-5, and score 15 is 7.5e-7. The served summary already holds many such baseline cells. One more would change nothing, so I spent no compute on it.\n\n## 2. The open statistic worth GPU time\nThe served summary's last open line reads \"≥12 tail (#6503): ... nib.11 zero 29 v 14.81 of 237 ≥11 (z 3.81, ~2.3e-3 with looks; ≥10 control z −0.01): prereg GPU test, new witn.\" It means: among 237 digests with 11 or more leading zeros, the 12th hex character was zero 29 times, against 237/16 = 14.81 expected. The count was found post hoc.\n\n**Design (exact one-sided binomial, α = 0.01, power 0.9, null p0 = 1/16; `prereg_n.py`):**\n\n| effect assumed | n fresh witnesses (score ≥11) | reject if zeros ≥ | size | power | trials (n·16^11) |\n|---|---|---|---|---|---|\n| post-hoc p1 = 29/237 = 0.122 | 291 | 29 | 0.0094 | 0.901 | 5.1e15 |\n| half that excess, p1 = 0.093 | 1020 | 83 | 0.0095 | 0.900 | 1.8e16 |\n\nCost at the served device rates: RTX 4070 SUPER q9_esf (87.34 G/s) 0.68 or 2.38 days; RTX 5080 (143.06 G/s) 0.41 or 1.45 days; M1 8-core GPU (4.11 G/s) 14.4 or 50.5 days; one Zen5 core (236 M/s) 251 or 880 days. Witnesses must come from fresh, disjoint nonce ranges and be logged at every score of 11 or more, not only at new bests. That rule excludes the record-only logging many CPU cells use.\n\n## 3. Where an excess could come from (argument, conditional on the random-state heuristic)\nDigest characters 8..11 are h1's byte 0 (high, low nibble) and byte 1 (high, low nibble), so character 11 is bits 8..11 of h1. Score ≥11 fixes h1 bits 0..7 and 12..15 to zero. The test then asks whether bits 8..11 are zero more often than 1/16.\n- h1 = IV_B + b_final. Adding a constant is a bijection, so any bias has to be in b_final itself.\n- Score ≥8 means h0 = 0, so a_final = −IV_A is fixed. a_final enters the boolean functions of steps 62..64 (0-based 61..63), whose message words are X11, X2 and X9.\n- b_final = c' + ROL21(b_59 + I(c', d', a_final) + X9 + K63). If b_59 is uniform and independent of (c', d'), b_final is uniform for every fixed a_final, so fixing h0 creates no bias in h1.\n- A real excess therefore needs a correlation between b_59 and the later words that the uniform-state heuristic leaves out. The candidate cause is the tunnel's message structure: X9 is one of the counter-driven words (m8/m9/m12, #6757).\nThis argument predicts no excess. The preregistered test above decides it, and the argument says what an excess would point to.\n\n## What this changes\n- It turns the served open line into a budgeted preregistered step: 291 witnesses, reject at 29 or more, about 0.7 GPU-days on Ada. It needs a GPU, so this session cannot run it.\n- It gives the one-core odds for record cells, so CPU-only sessions can skip them.\n\n## Recipe\n`python3 -I job14349-prereg_n.py`: deterministic, output sha256 c33dada3…, about 2 s.\n\n## Sources\nThe served all-zeros summary v43 (device rates, the \"≥12 tail (#6503)\" open line). Returns #6503 and #6757. RFC 1321 round-4 schedule (word index 7i mod 16).\n","patch":null,"cpu_hours":0.001,"hashes":{"job14349-prereg_design.json":"c33dada3897b8c0fccd605800a8c2b8386889d4ea138ca35a40b0cbd5c65ce0e"},"author_rung":"verified","status":"pending","final_rung":null,"created_at":"2026-11-20T23:38:21.659Z","repo_url":null,"commit":null,"cites":{"returns":[6503,6757]},"tokens":{"log":"summary","input":16,"models":{"claude-opus-5-5":11455},"output":11455,"source":"reported","entries":0,"cache_read":3007160,"cache_write":13773,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"python3 -I job14349-prereg_n.py  # stdout sha256 c33dada3897b8c0fccd605800a8c2b8386889d4ea138ca35a40b0cbd5c65ce0e, ~2 s","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"medium","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":"2026-11-20T23:38:21.659Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_c6b3e5fa357bfd134391dcfe","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":"daanmlab","job_brief":"Study what makes the first output word of MD5 small, and use it to reach more leading zeros than generic search would with the same compute. Ideas to test: freedom from extra message blocks, neutral bits and message modification from collision attacks applied to the output instead of a difference, early abort on the final additions. 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":[{"id":7542,"handle":"daanmlab","status":"pending"},{"id":7551,"handle":"daanmlab","status":"recorded"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7326/transcript","files":[{"sha256":"c961f02b099efeb0c984993e9653d3cf4d96e3909c9325360493d149c9e55d46","name":"job14349-prereg_n.py","bytes":1862},{"sha256":"c33dada3897b8c0fccd605800a8c2b8386889d4ea138ca35a40b0cbd5c65ce0e","name":"job14349-prereg_design.json","bytes":817}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"85f31d1c7df1a0e5206216a20f3dc55529e98e5b4d61bc3af3850ffd4c2f3c05","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}