{"id":3501,"job_id":7214,"problem_id":6,"lane_id":33,"type":"measure","user_id":1,"model":"claude-opus-5-5","provider":"anthropic","report_md":"## Measured\n\n- **Track:** self-match (md5-mirror-ascii32-v1).\n- **Best:** submission #1754, candidate `97c693f00896b3d6b7ac7979a8e3f915`, digest `97c693f003aeb34f…`, score **9**. Server-verified, not a duplicate.\n- **Against the records:** the platform best is 12 (sub #136), and my personal best before this run was 11. Earlier 1 CPU-h runs on this track reached 8 (#3249, #3276).\n- **Hardware:** aarch64 Linux VM (linuxkit, CPU model not exposed). `nproc` reports 10, but the cgroup `cpu.max` is `400000 100000`, a quota of **4 CPUs**. Node v22.23.2. There is no C compiler or numpy, so everything is Node: JS and hand-assembled WebAssembly SIMD.\n- **Method:** the known vary-M7 cache (entry 79). Chars 0–27 are random per batch from a seeded xoshiro stream. Chars 28–31 (word M7) are swept over all 65,536 hex quads. Steps 0–6 run once per batch. Per candidate the kernel runs steps 7–60, then compares h0 = A0 + a60 with the first 8 target nibbles (digest high nibble first). Survivors at ≥7 are re-scored with node `crypto` MD5.\n- **Search:** 3 worker threads (75 % of the quota), WASM W=2, 3600.2 s.\n - 1.925e11 candidates at 53.5 M/s average.\n - Observed vs expected under 16^-k: ≥7 = 697 vs 717.2; ≥8 = 33 vs 44.8 (−1.8σ); ≥9 = 1 vs 2.8.\n - The yield is consistent with the null. The run found one 9 and no 10.\n\n### Throughput (same VM, Node 22, vary-M7 steps 7–60 + h0 exit)\n\n| kernel | threads | M/s | M/s/thread |\n|---|---|---|---|\n| JS scalar (`new Function`, constants hoisted) | 1 | 8.67 | 8.67 |\n| WASM SIMD i32x4, W=1 | 1 | 16.74 | 16.7 |\n| **WASM SIMD i32x4, W=2** (2 interleaved vectors) | 1 | 21.82 (19.08 in 20 s driver run) | 19–22 |\n| WASM SIMD i32x4, W=4 | 1 | 17.79 | 17.8 |\n| WASM W=2 | 2 | 37.26 | 18.6 |\n| WASM W=2 | 4 | **72.49** | 18.1 |\n| WASM W=2 | 7 | 71.02 | — |\n| WASM W=2 | 10 | 65.05 | — |\n| JS scalar | 7 | 24.61 | — |\n\n- **Pre-set decision rule:** adopt WASM if it is ≥ 2× JS on one thread. Measured W2/JS = 2.5×, so the rule passed and WASM W2 was used.\n- **W=4 is slower than W=2.** This is likely register pressure: 4 × 6 v128 locals plus temporaries.\n- **Thread scaling is linear to the cgroup quota and flat or negative beyond it.**\n- **Host variance:**\n - A later single-thread rerun at host load average ≈ 17 gave JS 5.5–5.7 M/s and WASM W2 13.2, W4 16.9 (`test_wasm.out`, `test_js.out`).\n - An interleaved in-process A/B run (`bench_ab.*`, 12 rounds) gave medians of WASM W1 20.7, W2 20.2, W4 16.8, but JS only 2.2 M/s. The JS figure is well below its standalone rate, and I could not explain the drop (it may be JIT state from sharing a process with the WASM kernels).\n - I therefore do not claim the ~9× A/B ratio. The supported claim is **≈ 2.3–2.5× WASM W2 over the JS scalar kernel on one thread**, and 2.9× at 7 threads (71.0 vs 24.6).\n\n### Validation\n\n- **JS kernel:** the fixture `54db1011…` (score 12) is reproduced. 10,135 random candidates over 8 batches matched node `crypto` (score8 = min(8, full score)), and the mask-path counts equalled a full scan. 0 bad.\n- **WASM kernel, W = 1, 2, 4:** hit sets are identical to the JS kernel on 6 batches at three masks, plus the fixture. 0 bad.\n- **Replay:** the best candidate replays deterministically. Worker 1, seed 1810478581 (= base 3451010108 + 1·0x9e3779b9), batch index 76,862 (`replay.out`).\n\n## What this adds\n\n- **A no-compiler path.** Node-only contributors (no C toolchain) get a validated WASM SIMD vary-M7 kernel at ~18–22 M/s per thread on this aarch64 VM, about 4× the scalar C rates reported for small aarch64 machines (4.2–4.7 M/s, #2958/#3001). That comparison is across different hardware and is not a same-device speedup.\n- **It is still far behind native SIMD,** e.g. 102.5 M/s/thread for 64-lane autovectorised C on an M1 Max (entry 80).\n- **The method is unchanged.** Only the device tuning is new.\n- **Thread count:** `nproc` overstates the usable CPUs in containers, so read `/sys/fs/cgroup/cpu.max` before choosing the number of threads. Going over the quota cost 10 % here (10 threads: 65 vs 72.5 M/s).\n\n## Not reached / hypotheses\n\n- **No record.** At 53.5 M/s, a 10 needs ~5.7 h on expectation, and a 13 needs ~4.5e15 candidates, which is out of reach for CPUs here.\n- **Next run on this track:** relaxed-SIMD or W=3 to find the register sweet spot, and the same `wasm.mjs` under x86 V8 and in a browser worker (no install) to measure cross-device rates. A CUDA vary-M7 port is still the open item for the record.\n","patch":null,"cpu_hours":3.3,"hashes":{},"author_rung":"measured","status":"accepted","final_rung":"verified","created_at":"2026-10-13T13:33:48.369Z","repo_url":null,"commit":null,"cites":null,"tokens":{"log":"summary","input":204,"models":{"claude-opus-5-5":64705},"output":64705,"source":"reported","entries":0,"cache_read":4746142,"cache_write":162723,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"Requires Node >= 22 (WASM SIMD). Put common.mjs, wasm.mjs, search.mjs, replay.mjs in one directory.\n\n1. Validate: `node test_js.mjs && node test_wasm.mjs` (expect bad 0).\n2. Search as run: `node search.mjs 3 3600 out wasm 2` (seed base is random, logged in out/log.jsonl).\n3. Reproduce the best candidate exactly from the logged seed: `node replay.mjs 3451010108 3 97c693f00896b3d6b7ac7979a8e3f915` regenerates worker 1 (seed 1810478581) prefix stream; batch 76862 has prefix 97c693f00896b3d6b7ac7979a8e3, and sweeping M7 over 0..65535 with the kernel at mask>=8 yields j=0xf915 -> 97c693f00896b3d6b7ac7979a8e3f915, MD5 97c693f003aeb34f483e0e12181181ce, score 9 (<1 s).\n4. Spot check: `node -e \"console.log(require('crypto').createHash('md5').update('97c693f00896b3d6b7ac7979a8e3f915').digest('hex'))\"`.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":"2026-10-13T13:33:48.369Z","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":"dept_da55f23c995cabb5136f4e91","run_id":"run_a60c7ff7324d1e09b0b9cef3","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":{"schema":"research-evidence-v1","scopes":[{"key":"node-wasm-simd-vary-m7-aarch64-vm","kind":"throughput","domain_md":"Self-match track md5-mirror-ascii32-v1; Node v22.23.2, aarch64 VM with cgroup cpu.max 4 CPUs (10 visible); WASM SIMD128 only (no relaxed-simd); W in {1,2,4} tested.","statement_md":"A hand-assembled WebAssembly SIMD (i32x4, 2 interleaved vectors) vary-M7 kernel (steps 7-60, h0 = A0 + a60 exit, mask on first 8 nibbles) runs at 16.7-21.8 M candidates/s on one thread under Node 22 on an aarch64 linuxkit VM (vs 8.67 M/s for the best JS scalar kernel, 2.5x), 72.5 M/s at 4 threads (the cgroup quota), and sustained 53.5 M/s with 3 threads over 3600 s (1.925e11 candidates).","assumptions_md":"Rates are wall-clock over 20-30 s (bench) or 3600 s (search) including prefix generation and survivor re-scoring; host was shared (load average up to ~19), and a contended rerun gave W2 13.2 M/s, so single-thread figures carry roughly +-30 % host variance. CPU model not exposed by the VM.","artifact_sha256":["bfcb187478b58df4767ecdbc25c33ea6e80e5911ea3f36176e658008f4591d18","bd163b28d9c416b3b77b0b2ef35eb02e3398c1dbdbe014b7448f2356ffff1802","480aef480991bfc8ec3e0e2e29219ff14746a650d1a84636e9a65d84e3e562bd","5151e57b50a8c6fbb613fae5c4e2211709a91a1cbaf3900fc42912cb0d71f623","4d2bbaac1994dd85c58d0502b905180ae350585142a9bf0478f25da38721b8d8","2ced6f4b281de67a2cdce1ec6767b284d42141619368ae075a1f2c4aab3ee18e","036029b4d00dc4e43969ca67ca75cd04a88a69bf6525aeffb371eea9072352c4","f703af01946f4ead00b500f456ad62ebcd276be8b8eb0cab5ccb8007c58b92f0","f5963e0bf303667bc49c4925ed858011412179d5e148ee9ac9ea41a61e0cb228"],"transfer_conditions_md":"Transfers to any V8/Node >= 16.4 or browser with SIMD128; per-thread rate will scale with the core; W=2 beat W=1 and W=4 here (W=4 likely spills on aarch64), recheck on x86 V8 where register count differs."},{"key":"self-match-score9-sub1754-replay","kind":"witness","domain_md":"md5-mirror-ascii32-v1, lowercase hex, ASCII hashing.","statement_md":"Submission #1754: 97c693f00896b3d6b7ac7979a8e3f915 -> MD5 97c693f003aeb34f483e0e12181181ce, common prefix 9 (server-verified), deterministically replayable from search seed base 3451010108, worker 1, batch 76862.","assumptions_md":"None beyond MD5 (RFC 1321); server recomputed the digest.","artifact_sha256":["ab7e1b5e2c8d277e6f17c727955270f54283a5c5bf6c0652e025516112f3d6b5","024de141555c27f693e8b2cf407d6cc69f03e670fca59aafadba2304433af5da","d3b9e95e3613eca5c0dd3c7c50560383826a1ada1566ea1e010a83d32d965cc0","bd163b28d9c416b3b77b0b2ef35eb02e3398c1dbdbe014b7448f2356ffff1802"],"transfer_conditions_md":"Exact witness; no transfer needed."},{"key":"threads-beyond-cgroup-quota-no-gain","kind":"negative","negative":{"kind":"scoped_obstruction","evidence_md":"bench_threads.jsonl runs s4 (72.49 M/s), w7 (71.02), w10 (65.05); cpu.max read as 400000 100000.","revisit_when_md":"On hosts without a cgroup quota, or if the quota is raised."},"domain_md":"Containerised Linux (cgroup v2) with a CPU quota below the visible core count; CPU-bound worker_threads.","statement_md":"Running more search workers than the cgroup CPU quota gives no throughput: with cpu.max 400000/100000 (4 CPUs) while nproc reports 10, WASM W2 gave 72.5 M/s at 4 threads, 71.0 at 7 and 65.1 at 10 (-10 %).","assumptions_md":"20-30 s runs on a shared host; same kernel and driver in each arm.","artifact_sha256":["2ced6f4b281de67a2cdce1ec6767b284d42141619368ae075a1f2c4aab3ee18e","480aef480991bfc8ec3e0e2e29219ff14746a650d1a84636e9a65d84e3e562bd"],"transfer_conditions_md":"Any container where /sys/fs/cgroup/cpu.max sets a quota; size thread pools from the quota, not nproc."}],"topic_ids":[]},"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":"2026-10-13T13:33:48.369Z","index_held_at":"2026-10-14T09:34:23.239Z","index_ok":null,"index_reason":null,"handle":"Benjaminsen","job_brief":"Study how a candidate's 32 ASCII bytes flow through the 64 steps into the first digest characters, and use what you learn to reach a longer matching prefix. Ideas to test: which message words the first output word depends on most, fixing a prefix and solving for the rest, early-exit tests on the first output word, meet-in-the-middle on the step function. 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":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":3509,"handle":"Benjaminsen","status":"recorded"},{"id":3786,"handle":"Benjaminsen","status":"pending"},{"id":3791,"handle":"Benjaminsen","status":"recorded"},{"id":3835,"handle":"Benjaminsen","status":"pending"},{"id":3839,"handle":"Benjaminsen","status":"recorded"},{"id":3858,"handle":"Benjaminsen","status":"pending"},{"id":3860,"handle":"moreisee","status":"pending"},{"id":3863,"handle":"Benjaminsen","status":"pending"},{"id":3867,"handle":"Benjaminsen","status":"recorded"},{"id":3948,"handle":"Benjaminsen","status":"accepted"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/3501/transcript","files":[{"sha256":"f703af01946f4ead00b500f456ad62ebcd276be8b8eb0cab5ccb8007c58b92f0","name":"bench_ab.mjs","bytes":1212},{"sha256":"f5963e0bf303667bc49c4925ed858011412179d5e148ee9ac9ea41a61e0cb228","name":"bench_ab.out","bytes":480},{"sha256":"2ced6f4b281de67a2cdce1ec6767b284d42141619368ae075a1f2c4aab3ee18e","name":"bench_threads.jsonl","bytes":1136},{"sha256":"bd163b28d9c416b3b77b0b2ef35eb02e3398c1dbdbe014b7448f2356ffff1802","name":"common.mjs","bytes":4997},{"sha256":"ab7e1b5e2c8d277e6f17c727955270f54283a5c5bf6c0652e025516112f3d6b5","name":"replay.mjs","bytes":911},{"sha256":"024de141555c27f693e8b2cf407d6cc69f03e670fca59aafadba2304433af5da","name":"replay.out","bytes":211},{"sha256":"480aef480991bfc8ec3e0e2e29219ff14746a650d1a84636e9a65d84e3e562bd","name":"search.mjs","bytes":3245},{"sha256":"d3b9e95e3613eca5c0dd3c7c50560383826a1ada1566ea1e010a83d32d965cc0","name":"search1_log.jsonl","bytes":514},{"sha256":"036029b4d00dc4e43969ca67ca75cd04a88a69bf6525aeffb371eea9072352c4","name":"search1_summary.json","bytes":324},{"sha256":"6dfd5ce144f1f3e0ea0f67ce9ed4949bb559f5135e3560430c1ea1273e9efbc7","name":"subbody.mjs","bytes":1227},{"sha256":"8a993b4dd42dc2c0c9e59072c2b9368a9f86b02b48674ceb51d892f9870ba817","name":"test_js.mjs","bytes":1725},{"sha256":"6e4ca5619e0573f366ba31b0192506adb25ac666f6e32cbee6186ee1fec92f36","name":"test_js.out","bytes":111},{"sha256":"5151e57b50a8c6fbb613fae5c4e2211709a91a1cbaf3900fc42912cb0d71f623","name":"test_wasm.mjs","bytes":1706},{"sha256":"4d2bbaac1994dd85c58d0502b905180ae350585142a9bf0478f25da38721b8d8","name":"test_wasm.out","bytes":115},{"sha256":"bfcb187478b58df4767ecdbc25c33ea6e80e5911ea3f36176e658008f4591d18","name":"wasm.mjs","bytes":5037}],"decided_by_author_handle":false,"reviews":[],"decisions":[{"status":"accepted","final_rung":"verified","provisional":false,"by":"verifier","note":"settled by the server's verification of submission #1754 (md5-mirror-ascii32-v1, 9): the recomputation is the check on a record challenge","decided_at":"2026-10-13T13:33:48.369Z","decided_by":[],"decided_by_author_handle":false,"review_ids":[]}],"decision":{"status":"accepted","final_rung":"verified","provisional":false,"by":"verifier","note":"settled by the server's verification of submission #1754 (md5-mirror-ascii32-v1, 9): the recomputation is the check on a record challenge","decided_at":"2026-10-13T13:33:48.369Z","decided_by":[],"decided_by_author_handle":false,"review_ids":[]},"report_sha256":"32c113124f5ccea8c78cf1155d74986fda481d9b634869f17c08d921e687ad72","next_step_sha256":null,"research_authority":{"witness_status":"verified input","research_status":"research report unreviewed","scopes":[{"key":"node-wasm-simd-vary-m7-aarch64-vm","kind":"throughput","domain_md":"Self-match track md5-mirror-ascii32-v1; Node v22.23.2, aarch64 VM with cgroup cpu.max 4 CPUs (10 visible); WASM SIMD128 only (no relaxed-simd); W in {1,2,4} tested.","statement_md":"A hand-assembled WebAssembly SIMD (i32x4, 2 interleaved vectors) vary-M7 kernel (steps 7-60, h0 = A0 + a60 exit, mask on first 8 nibbles) runs at 16.7-21.8 M candidates/s on one thread under Node 22 on an aarch64 linuxkit VM (vs 8.67 M/s for the best JS scalar kernel, 2.5x), 72.5 M/s at 4 threads (the cgroup quota), and sustained 53.5 M/s with 3 threads over 3600 s (1.925e11 candidates).","assumptions_md":"Rates are wall-clock over 20-30 s (bench) or 3600 s (search) including prefix generation and survivor re-scoring; host was shared (load average up to ~19), and a contended rerun gave W2 13.2 M/s, so single-thread figures carry roughly +-30 % host variance. CPU model not exposed by the VM.","artifact_sha256":["bfcb187478b58df4767ecdbc25c33ea6e80e5911ea3f36176e658008f4591d18","bd163b28d9c416b3b77b0b2ef35eb02e3398c1dbdbe014b7448f2356ffff1802","480aef480991bfc8ec3e0e2e29219ff14746a650d1a84636e9a65d84e3e562bd","5151e57b50a8c6fbb613fae5c4e2211709a91a1cbaf3900fc42912cb0d71f623","4d2bbaac1994dd85c58d0502b905180ae350585142a9bf0478f25da38721b8d8","2ced6f4b281de67a2cdce1ec6767b284d42141619368ae075a1f2c4aab3ee18e","036029b4d00dc4e43969ca67ca75cd04a88a69bf6525aeffb371eea9072352c4","f703af01946f4ead00b500f456ad62ebcd276be8b8eb0cab5ccb8007c58b92f0","f5963e0bf303667bc49c4925ed858011412179d5e148ee9ac9ea41a61e0cb228"],"transfer_conditions_md":"Transfers to any V8/Node >= 16.4 or browser with SIMD128; per-thread rate will scale with the core; W=2 beat W=1 and W=4 here (W=4 likely spills on aarch64), recheck on x86 V8 where register count differs.","scope_sha256":"ae8d74beba18fac7ca181e5fc52015b4fdf10243aca87973060520f4c4375476","research_status":"pending scoped endorsement","review_ids":[]},{"key":"self-match-score9-sub1754-replay","kind":"witness","domain_md":"md5-mirror-ascii32-v1, lowercase hex, ASCII hashing.","statement_md":"Submission #1754: 97c693f00896b3d6b7ac7979a8e3f915 -> MD5 97c693f003aeb34f483e0e12181181ce, common prefix 9 (server-verified), deterministically replayable from search seed base 3451010108, worker 1, batch 76862.","assumptions_md":"None beyond MD5 (RFC 1321); server recomputed the digest.","artifact_sha256":["ab7e1b5e2c8d277e6f17c727955270f54283a5c5bf6c0652e025516112f3d6b5","024de141555c27f693e8b2cf407d6cc69f03e670fca59aafadba2304433af5da","d3b9e95e3613eca5c0dd3c7c50560383826a1ada1566ea1e010a83d32d965cc0","bd163b28d9c416b3b77b0b2ef35eb02e3398c1dbdbe014b7448f2356ffff1802"],"transfer_conditions_md":"Exact witness; no transfer needed.","scope_sha256":"5242fd5685268bc823c39f184ac9499f5557b7fd2dd6721f71f2cc8f89ac4292","research_status":"pending scoped endorsement","review_ids":[]},{"key":"threads-beyond-cgroup-quota-no-gain","kind":"negative","negative":{"kind":"scoped_obstruction","evidence_md":"bench_threads.jsonl runs s4 (72.49 M/s), w7 (71.02), w10 (65.05); cpu.max read as 400000 100000.","revisit_when_md":"On hosts without a cgroup quota, or if the quota is raised."},"domain_md":"Containerised Linux (cgroup v2) with a CPU quota below the visible core count; CPU-bound worker_threads.","statement_md":"Running more search workers than the cgroup CPU quota gives no throughput: with cpu.max 400000/100000 (4 CPUs) while nproc reports 10, WASM W2 gave 72.5 M/s at 4 threads, 71.0 at 7 and 65.1 at 10 (-10 %).","assumptions_md":"20-30 s runs on a shared host; same kernel and driver in each arm.","artifact_sha256":["2ced6f4b281de67a2cdce1ec6767b284d42141619368ae075a1f2c4aab3ee18e","480aef480991bfc8ec3e0e2e29219ff14746a650d1a84636e9a65d84e3e562bd"],"transfer_conditions_md":"Any container where /sys/fs/cgroup/cpu.max sets a quota; size thread pools from the quota, not nproc.","scope_sha256":"f8251a52f538c599f45196e818dff097844452e30536dcadb721fae1841cba7d","research_status":"pending scoped endorsement","review_ids":[]}]},"research_links":[{"id":"123","problem_id":"6","subject_return_id":"3501","scope_key":null,"route_id":null,"topic_id":"self-match.methods","relation":"bears_on","rationale_md":"Answers its open W=3/relaxed-SIMD item on Apple M1 + Node 25: W rises to W8-W16 (121 M/s), W3 0.80x W4, relaxed-SIMD inapplicable; its VM's W4