{"id":7179,"job_id":14136,"problem_id":6,"lane_id":34,"type":"explore","user_id":90,"model":"muse-spark","provider":"unknown","report_md":"# All zeros explore: L=128 tested at power for the first time — 1-billion-candidate smoke cell, uniform yield, and two corrections to the standing proposal\n\nJob 14136, explore, all-zeros. Improvement claim: none. No input scoring 13 or above was found, and none is claimed. Nothing goes to the results page. Best input scores 8.\n\n## Independence objective\n\nDeliberate replication needs a stated independence objective, so here it is: this return reuses no prior tally, candidate, or binary output. It re-reads the track state page snapshot and the filed returns for recent all-zeros jobs, and it runs a fresh 128-byte searcher built for this job from the plain full-round vector kernel, with every reported candidate re-verified with hashlib by parsing the cell log programmatically, not from hand-typed strings.\n\n## Comparison of the accepted scoped answers\n\nState page snapshot from the recent run dir (38,877 bytes, md5 20ded3e2aaa983b0416ac308c65b6c12), byte-identical to the snapshots in jobs 14100, 14108 and 14112. All-zeros lane unchanged since v43 (2026-11-03), through return #6503. Records unchanged: platform best 13 of 32 (submission #180, 2026-10-11); best published result 14 of 32 (Beneri Hash Game, record #209).\n\nWhat the recent filed returns already establish, re-read this cycle:\n\n- Return 7155 (job 14100): the il8 edge over il4 is a solo-core figure (~1.04x under two-core load, not 1.31x); the state page is stale on three items (s2 deficit closed by 7113, many-core total closed by 7133, fastest-way line still lists il4 at 235.8 M/s against il8 at 321.48 M/s solo). First proposal of the L=128 experiment.\n- Return 7159 (job 14105): fresh 400B il8 exact-tally cell, best 10, uniform.\n- Return 7161 (job 14108): fresh 400B il8 exact-tally cell, best 9, uniform; repeats the L=128 proposal.\n- Return 7164 (job 14112): fresh 400B il8 exact-tally cell, best 9, uniform; repeats the L=128 proposal, and notes this VM instance runs slower than prior ones.\n\nThe standing lead verified, not taken on faith: L=128 was still uncovered. Jobs 14100, 14108 and 14112 each proposed it and none built it; no 128-byte cell exists in any run dir, and the loop state still lists it as open. This job closes the \"untested\" half of it with a measured cell. Two corrections to the proposal itself follow; both change the cost of the full experiment.\n\n## Correction 1: a 128-byte input costs three compressions, not two\n\nA 128-byte message is two message blocks, but MD5 padding needs 0x80 plus 8 length bytes, and 128 is already 0 mod 64, so the padded message is 192 bytes: two message blocks plus one padding block, three compressions per candidate. The prior returns call this a \"two-block\" cell, counting message blocks only. The cost per candidate is 3x a full-round single-block searcher, not 2x.\n\n## Correction 2: the il8 tunnel does not mechanically extend to L=128\n\nThe il8 binary is a single-block 52-byte tunnel searcher: it fixes tunnel conditions on Q10/Q11, derives message words m[0..12] from fixed Q values, pads with m[13]=0x80, m[14]=416, varies a single 32-bit tunnel variable, and runs only steps 24..60 before scoring h0. Every one of those steps is single-block structure. A 128-byte input needs full 64-step compressions for blocks 1 and 2 (the intermediate chaining values are full state; nothing can be scored until the final block), with only the last block able to reuse the step-60 h0 exit. So the full build is a new kernel: 192 vector steps per candidate at plain-kernel efficiency (or about 164 if a block-1 tunnel variant is built, cutting block 1 to steps 24..63), not a mechanical extension of the il8 source in the style of gen_il8.py. The generator script only widens the interleave of the same single-block kernel.\n\n## The cheap decisive check: a 1-billion-candidate 128-byte smoke cell\n\nInstead of building the full kernel first, this job built the smallest searcher that answers the premise: a standalone driver reusing the plain full-round 16-lane vector kernel from tools/md5zeros_v16.c verbatim, run three times per candidate (block 1: word 0 is the candidate index, words 1..15 fixed; block 2: all 16 words fixed; block 3: padding 0x80, zeros, length 1024 bits; fixed words from splitmix64, seed 20261109). Source and log: runs/job-14136/work/md5_128_smoke.c and runs/job-14136/work/cell128.log.\n\nCorrectness first: the driver prints a self-test line for candidate 0, and its digest matches hashlib exactly; a second spot check (candidate 9) also matches. 1,000,000,000 fresh candidates (index 0..999,999,999, a space no prior cell has touched, since no 128-byte cell ever ran), pinned to core 1 while the sibling il8 production cell ran on core 0. Exit 0, 88 seconds, 11.37 M/s loaded.\n\nResult: best 8. Best-8 input is candidate index 564353088 (block-1 word 0 = 0x21a35840, remaining words fixed from seed 20261109; full 256-hex input in the cell log), digest 00000000c30ad405d2107770b8fdc091. Exact tallies:\n\n| rung | observed | expected | z |\n|---|---|---|---|\n| >=2 | 3900851 | 3906250 | -2.74 |\n| >=4 | 15376 | 15258.8 | +0.95 |\n| >=6 | 50 | 59.6 | -1.24 |\n| >=7 | 5 | 3.73 | +0.66 |\n| >=8 | 1 | 0.23 | +1.59 |\n| >=9 | 0 | 0.015 | P(no 9) 0.986 |\n| >=10 | 0 | 0.0009 | P(no 10) 0.999 |\n\nAll 10 scored lines in the log (self-test plus new-best lines at scores 0..8, including the score-7 and score-8 hits) were re-verified with hashlib by parsing the log programmatically: 10 of 10 match, 0 mismatches. Yield per candidate is consistent with uniform. The largest deviation is rung 2 at -2.74 sigma; across 7 rungs the chance of a deviation at least that large under uniform is about 4%, and rung 2 is the least informative rung (first digest byte only), so this is reported as noise, not a signal, and it is not a basis for any structural claim. No 9 was expected in 1B candidates (0.015 expected), so best 8 is the expected outcome.\n\n## Cost: what a full L=128 cell would take\n\nMeasured: 11.37 M/s with the plain 16-lane 3-block kernel, about 20x slower per candidate than the production il8 tunnel cell (~225 M/s loaded). Step-count analysis says most of that gap is lane width, not steps: the tunnel does 37 steps at 128 lanes, this smoke does 192 steps at 16 lanes. A 128-lane 3-block build at tunnel-like per-step efficiency would run roughly 40-90 M/s, so a full 400B L=128 cell would take about 1.2-2.6 hours on one core: it fits inside a 4-hour job window but needs chunking across cycles, unlike the 30-minute 400B tunnel cells. At this smoke kernel's own rate a 400B cell would take 9.8 hours and is not worth running.\n\n## Cheapest new experiment (exact specification)\n\nBuild the 128-lane 3-block searcher: reuse the il8 interleave-8 step machinery for the final block (h0 exit at step 60), run full 64-step compressions for blocks 1 and 2, vary block-1 word 0 over a fresh 2^32 index space with fixed words from seed 20261109 (same fixed words as this smoke cell, so ranges stay disjoint by index), self-test against hashlib on sampled candidates, then a 20B cell comparing rate and per-rung yield against this smoke cell's 11.37 M/s and uniform tallies. Refutation/confirmation: if yield per candidate stays uniform and rate lands in the projected 40-90 M/s band, L=128 is confirmed as ~2-5x the tunnel cost with no gain and the dead-end line can be closed as tested-at-power; if yield differs at any rung beyond 3 sigma on 20B candidates, the length question reopens. The block-1 tunnel variant (about 164 steps instead of 192) is a follow-up optimization for that build, not a prerequisite.\n\n## Other notes\n\n- The VM reset since the 01:28 cycle close (boot id changed); no work lost for this job, which held nothing beforehand, and the sibling production cell was already running on core 0 throughout this check. The VM now reports 4 vCPUs; this check used core 1 only, as instructed.\n- Nothing scored 13 or above; nothing to the results page.\n- Next disjoint all-zeros start nonce for 52-byte cells remains 18280403100000 (in use by the sibling cell this cycle); next 128-byte candidate index after this smoke cell is 1000000000.\n","patch":null,"cpu_hours":0,"hashes":{},"author_rung":null,"status":"recorded","final_rung":"recorded","created_at":"2026-11-09T06:45:48.153Z","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 (job 14136, explore, all-zeros)\n\nReproduce the 128-byte smoke cell:\n\n1. Build the driver from the run dir:\n   gcc -O3 -march=znver4 -mtune=znver4 -o runs/job-14136/work/md5_128_smoke runs/job-14136/work/md5_128_smoke.c -lm\n2. Self-test 16 candidates and check candidate 0 against hashlib:\n   runs/job-14136/work/md5_128_smoke 16 5\n   python3 -c \"import hashlib; print(hashlib.md5(bytes.fromhex('<256-hex input from the [selftest] line>')).hexdigest())\"\n   The printed digest must equal the hashlib output.\n3. Run the 1-billion-candidate cell pinned to core 1:\n   taskset -c 1 runs/job-14136/work/md5_128_smoke 1000000000 10 > runs/job-14136/work/cell128.log 2>&1\n   Expected: about 88 s, about 11.37 M/s loaded, best 8, tallies t2=3900851 t4=15376 t6=50 t7=5 t8=1 t9=0 t10=0.\n4. Verify every scored line with hashlib by parsing the log (match input hex to digest and score); expected 10 of 10, 0 mismatches.\n\nCheapest next experiment: build the 128-lane 3-block searcher (full 64-step compressions for blocks 1 and 2, final block with the step-60 h0 exit; vary block-1 word 0 over a fresh index space starting at 1000000000, same fixed words from seed 20261109), self-test against hashlib, then a 20B cell comparing rate (projected 40-90 M/s) and per-rung yield against uniform.","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/7179/transcript","files":[],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"a13a0535e7b1c8905c0bf097810cd6d7a1d1e67126351e7266feccd589b75132","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}