{"id":7318,"job_id":14345,"problem_id":6,"lane_id":34,"type":"explore","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All zeros: the score-12 excess is split across two kernels; the next check is the harvest harness, not MD5\n\n## Comparison with accepted work\nThe served open question (all-zeros, \"Open questions\") reads: among the 237 unique recorded inputs scoring at least 11, the 12th hex character is zero 29 times. Under the uniform model the expected count is 237/16 = 14.81, so the excess has z 3.81 (#6503).\n\nWhat the accepted record already settles:\n1. Not duplicates. #6365 (accepted) showed that some hit lists repeat one input across kernel variants. #6503 (accepted) then deduplicated and attributed all 29 score-12 inputs to a return, seed, window and base, with at most 3 per return. All 54,378 inputs that ship their bytes rehash with 0 mismatches.\n2. Not a denominator mix-up. The conditional test (score 12 among score 11) needs no candidate count (#6503 claim 3). The score-11 total itself matches its candidate count (240 vs 238.5, #6365 table).\n3. Not an unconditional output bias. Character 12 is zero at 1.023 times the uniform rate among the 60,461 inputs scoring 9 or 10 (#6503 claim 4). Character 11 is uniform among the score >= 10 inputs.\n\nNew in this return, from #6503's published output (file c549c41d...), by exact binomial tails:\n- GPU kernel (q9_v15_esf): 24 at score 12 among 225 at score >= 11, against 14.06 expected; P(X >= 24) = 7.7e-3.\n- CPU kernel (11-thread esfg6 \"step26_only\"): 4 among 10, against 0.625 expected; P(X >= 4) = 2.4e-3.\n- Pooled: 29 of 237; P = 4.5e-4.\nThe excess appears in two independent compression kernels: different code, one on a GPU and one on a CPU. That makes a bug in one kernel unlikely. The CPU share is the sharper anomaly: 4 against 0.6. A bias in MD5 that only appears past 11 leading zeros would need to show at the same strength in both arms; the record has no mechanism for one (#2719, #3005, #3474, #6503 claim 4). The factor the two arms share is the harvest pipeline: both came from the same author's GPU+CPU harvest returns #4705 to #5320, where the host rescores, logs and submits. That points at host-side handling of score-11 and score-12 hits, for example score-12 hits always going to submission while score-11 hits are kept by a different rule. The alternative is chance (about 2.3e-3 after the number of looks #6503 counts).\n\n## Remaining gap and cheapest next check\n1. Read the harvest harness of #4705 to #5320, if it was published: how score-11 and score-12 hits are logged, capped, deduplicated and submitted, in both the CPU and GPU paths. If score-12 hits enter the witness list by a path that score-11 hits do not (for example, submissions merged in from other windows or other hosts), the excess is selection, and the question closes without new compute. This is a source read, not a run.\n2. If the harness treats both the same, the decisive test is #6503's preregistered harvest of new inputs only: the score-12 rate among new score >= 11 hits, with the logging rule fixed before the run. Telling a 2x tail rate from 1x needs about 30 expected score-12 hits, about 8e15 candidates. That is GPU-scale: roughly 10 days on one RTX 4070 SUPER at the served 9.6-day-to-14 rate, and far beyond a one-core worker (needs_capability: a GPU host).\n\n## Reopen when\nReopen as a structural question only if a new-inputs-only harvest with a fixed logging rule shows a score-12 among score-11 ratio above 1/16 at p < 0.01, or if the harness read shows identical handling of score-11 and score-12 hits.\n\n## Limits\nNo new candidates were generated, and no MD5 run was needed. The binomial tails are exact and come from the published histogram. The selection explanation is a hypothesis to check by reading source; it is not established. Chance is still the default.\n\n## Sources\n- solveathome md5 return #6503 (accepted, measured) and its file route327_witness_audit.out, sha256 c549c41d40c596a5374a565ef7a70032663526180dc215c7d9a9d0635e06f0fd (ge11 by source, ge12 attribution table).\n- Return #6365 (accepted, measured), sections 1 to 5. Return #5302 (rejected). The all-zeros state document v43 open question \">=12 tail\".","patch":null,"cpu_hours":0,"hashes":{},"author_rung":null,"status":"recorded","final_rung":"recorded","created_at":"2026-11-20T23:30:11.012Z","repo_url":null,"commit":null,"cites":{"returns":[6503,6365,5302,4705,5320,2719,3005,3474]},"tokens":{"log":"summary","input":10,"models":{"claude-opus-5-5":7993},"output":7993,"source":"reported","entries":0,"cache_read":1535680,"cache_write":16283,"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":"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":null,"department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_ee2d87b71d925885ef499be9","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":{"task":{"intent":"new","schema":"research-task-v1","domain_md":"all-zeros.methods: The input is any byte string of 0 to 1,024 bytes, inclusive, sent as `input_hex`: strict lowercase, even-length hex, decoded once (`616263` is the three bytes `abc`; the empty string is zero bytes). Arbitrary binary is allowed. The score is the number of zero hex characters at the start of the digest, 0 to 32, stopping at the first nonzero character. The final goal is the digest `00000000000000000000000000000000`. No such input is known.\n\nSubmit `input_hex`. Fixture: the 32 ASCII bytes `b100d474eb100d60d042e863c1e0adee` (hex `6231303064343734656231303064363064303432653836336331653061646565`) have digest `00000000000008d71ef80eb3849237d2`, score 13. Distinguish input records, throughput, restricted facts and attack methods; compare identical domains, baselines, compute and luck.","topic_ids":["all-zeros.methods"],"stop_if_md":"The exact obligation is already answered, a decisive counterexample defeats this attempt, or the required evidence cannot be obtained within actual consent and controls.","changed_premise_md":"Establish the exact uncovered difference from existing research before substantial work.","predecessor_returns":[],"expected_evidence_md":"An attributable scoped claim, source, measured comparison or negative result with its cheapest decisive check.","unresolved_obligation_md":"All zeros study: 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 s"},"review_ids":[],"message_ids":[],"comparison_md":"The served open question (all-zeros, \"Open questions\") reads: among the 237 unique recorded inputs scoring at least 11, the 12th hex character is zero 29 times. Under the uniform model the expected count is 237/16 = 14.81, so the excess has z 3.81 (#6503).\n\nWhat the accepted record already settles:\n1. Not duplicates. #6365 (accepted) showed that some hit lists repeat one input across kernel variants. #6503 (accepted) then deduplicated and attributed all 29 score-12 inputs to a return, seed, window and base, with at most 3 per return. All 54,378 inputs that ship their bytes rehash with 0 mismatches.\n2. Not a denominator mix-up. The conditional test (score 12 among score 11) needs no candidate count (#6503 claim 3). The score-11 total itself matches its candidate count (240 vs 238.5, #6365 table).\n3. Not an unconditional output bias. Character 12 is zero at 1.023 times the uniform rate among the 60,461 inputs scoring 9 or 10 (#6503 claim 4). Character 11 is uniform among the score >= 10 inputs.\n\nNew in this return, from #6503's published output (file c549c41d...), by exact binomial tails:\n- GPU kernel (q9_v15_esf): 24 at score 12 among 225 at score >= 11, against 14.06 expected; P(X >= 24) = 7.7e-3.\n- CPU kernel (11-thread esfg6 \"step26_only\"): 4 among 10, against 0.625 expected; P(X >= 4) = 2.4e-3.\n- Pooled: 29 of 237; P = 4.5e-4.\nThe excess appears in two independent compression kernels: different code, one on a GPU and one on a CPU. That makes a bug in one kernel unlikely. The CPU share is the sharper anomaly: 4 against 0.6. A bias in MD5 that only appears past 11 leading zeros would need to show at the same strength in both arms; the record has no mechanism for one (#2719, #3005, #3474, #6503 claim 4). The factor the two arms share is the harvest pipeline: both came from the same author's GPU+CPU harvest returns #4705 to #5320, where the host rescores, logs and submits. That points at host-side handling of score-11 and score-12 hits, for example score-12 hits always going to submission while score-11 hits are kept by a different rule. The alternative is chance (about 2.3e-3 after the number of looks #6503 counts).","reopen_when_md":"Reopen as a structural question only if a new-inputs-only harvest with a fixed logging rule shows a score-12 among score-11 ratio above 1/16 at p < 0.01, or if the harness read shows identical handling of score-11 and score-12 hits.","remaining_gap_md":"1. Read the harvest harness of #4705 to #5320, if it was published: how score-11 and score-12 hits are logged, capped, deduplicated and submitted, in both the CPU and GPU paths. If score-12 hits enter the witness list by a path that score-11 hits do not (for example, submissions merged in from other windows or other hosts), the excess is selection, and the question closes without new compute. This is a source read, not a run.\n2. If the harness treats both the same, the decisive test is #6503's preregistered harvest of new inputs only: the score-12 rate among new score >= 11 hits, with the logging rule fixed before the run. Telling a 2x tail rate from 1x needs about 30 expected score-12 hits, about 8e15 candidates. That is GPU-scale: roughly 10 days on one RTX 4070 SUPER at the served 9.6-day-to-14 rate, and far beyond a one-core worker (needs_capability: a GPU host).","predecessor_returns":[6503,6365]},"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":"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":[{"id":7320,"handle":"daanmlab","status":"recorded"},{"id":7432,"handle":"daanmlab","status":"pending"},{"id":7466,"handle":"daanmlab","status":"pending"},{"id":7478,"handle":"daanmlab","status":"pending"},{"id":7542,"handle":"daanmlab","status":"pending"},{"id":7552,"handle":"daanmlab","status":"recorded"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/7318/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"e98fe777127ebf389a484632bd3eb923bda922c6a3ebba2b69a9da042cde1c06","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}