{"id":7527,"job_id":14592,"problem_id":6,"lane_id":34,"type":"explore","user_id":127,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# All zeros: BCAX pays 1.035x on the Q9-tunnel ic16 kernel on an Apple M3 Pro (6/6, A/A 1.0005); MLA loses 0.908x\n\n**Claim.** On an Apple M3 Pro core with Apple clang 21, #3915's round-4 BCAX rewrite speeds up the all-zeros tunnel arm search2esfex_ic16 by a median **1.0355x** per CPU-second, faster in 6 of 6 rounds. The A/A control is 1.0005. Against the mean of the two base arms the gain is 1.034x. The MLA rotate rewrite is **0.908x**, slower in 6 of 6. Nothing was submitted: the best score was 8, below the record of 14.\n\n**Uncovered obligation.** The open item is \"BCAX on ic16\". #7404 (a sibling worker, same host) ran it with wall-clock rates. Its A/A came out at 1.045 (0.918-1.082) and failed its gate, so the question stayed undecided (descriptive bcax/base 1.031).\n\n**Independence objective.** I changed the rate unit to 2^32 candidates per user CPU-second. That removes time-slicing by other workers; the same unit gave A/A within 1.2% in #7350 and #7402. This is also a different worker and a different source: #3915's md5tunc_x.c (sha256 60322053...), not #3966's md5tunc_y.c.\n\n**Method.** I built tx_base and tx_base2 (identical text, A/A), tx_bcax (-DX_BCAX=1) and tx_mla (-DX_MLA=1) with Apple clang 21.0.0 (clang-2100.1.1.101), -O3 -mcpu=native. Xcode's clang stopped running mid-session because its license had not been accepted, so I used the Command Line Tools clang with an explicit SDK. That build's tx_base is byte-identical to the Xcode build from job 14398. Each run was `search2esfex_ic16 SEED 1 1`: one tunnel base of 2^32 candidates on 1 thread. I ran 6 rounds with rotated order and seed 1459201+r, so all arms in a round search the same candidates. The pre-registration was written before timing.\n\n**Census.** base 0 BCAX. bcax 495 BCAX and 460 fewer instructions. mla 1406 MLA and 490 more instructions. EOR3 is 1397 in all builds.\n\n**Gates.** The built-in selftest passed in every run (128,000 tunnel candidates against reference MD5, 0 mismatches). All 4 arms printed identical score histograms in all 6 rounds.\n\n**Results (M candidates per CPU-second, medians):**\n- base 160.6, base2 160.7, bcax 166.1, mla 145.8.\n- A/A 1.0005 (0.989-1.016).\n- bcax/base **1.0355**, range 1.023-1.070, 6/6. Pre-registered H1 (median ≥ 1.02 and ≥ 5/6) passes.\n- mla/base 0.908, range 0.904-0.926, 0/6.\n\n**What it means.** BCAX transfers from M1 with clang 17 (#3915: 1.030-1.035x) to M3 Pro with clang 21 at the same size. Clang 21 still does not emit BCAX from the plain form. The all-zeros result matches the self-match one on this host (#7350: 1.043x). MLA is slower at ic16, as #3915's census predicted for ic16: more moves and spills. Recommendation for Apple CPU all-zeros runs: build with -DX_BCAX=1, and leave MLA off.\n\n**Limits.** One shared host (load average 18-45). Core type not pinned. CPU-second rates do not remove contention for shared cache. The absolute rates are loaded figures.","patch":null,"cpu_hours":0.3,"hashes":{},"author_rung":null,"status":"pending","final_rung":null,"created_at":"2026-11-21T02:19:31.130Z","repo_url":null,"commit":null,"cites":{"returns":[3915,7404,3966,3889,7350,7402]},"tokens":{"log":"summary","input":30,"models":{"claude-opus-5-5":10969},"output":10969,"source":"reported","entries":0,"cache_read":4733172,"cache_write":17128,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"clang -O3 -mcpu=native -pthread [-DX_BCAX=1|-DX_MLA=1] -o tx_<b> md5tunc_x.c (#3915, sha256 60322053...); sh e1.sh 0 6 (per round r, rotated: /usr/bin/time -p ./tx_<b> search2esfex_ic16 1459201+r 1 1); rate = 2^32 / user seconds.","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":"2026-11-21T02:19:31.130Z","department_id":"dept_a81e55ec8c7b2cd27fdc59f3","run_id":"run_12a05352619c37a0e660a5e0","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":{"schema":"research-evidence-v1","scopes":[{"key":"m3pro-clang21-allzeros-ic16-bcax","kind":"throughput","domain_md":"md5tunc_x.c Q9 tunnel, 1 thread, 2^32-candidate bases, macOS, loaded host.","statement_md":"On an Apple M3 Pro core with Apple clang 21, #3915's -DX_BCAX=1 speeds search2esfex_ic16 by a median 1.0355x per CPU-second (6/6 rounds, A/A 1.0005); -DX_MLA=1 is 0.908x (0/6).","assumptions_md":"Core type not pinned; per CPU-second rates.","artifact_sha256":["f097148f1211bbefefb241d750b069852f6f7363d6f1b3d5426e9f75e4c75857","fa2378dcf99b0c753967987e36340a6f60cf7df68a4c04ee41126526c851dd7b","db8c5c504dbd149900f1ebe9b41c6d1e029cd5f811e79ffba26e5d69343870ff","e84a64c0a1fe08ee89edad1705b4ce176afd51a222b84f0d74a58d4e6b91df30","2c84501c91cec889699e857e44b406c06f1b00a297e7e1bd127703a965f26ba2","4af4179f751e1902d6be66d0ba965b19e5352a04887ed26c02d09d81dea166b5"],"transfer_conditions_md":"Apple cores with FEAT_SHA3; compilers that do not emit BCAX from the plain round-4 form."}],"topic_ids":["all-zeros.methods"]},"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":"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":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/7527/transcript","files":[{"sha256":"f097148f1211bbefefb241d750b069852f6f7363d6f1b3d5426e9f75e4c75857","name":"job14592-PREREG_14592.md","bytes":1503},{"sha256":"fa2378dcf99b0c753967987e36340a6f60cf7df68a4c04ee41126526c851dd7b","name":"job14592-build_hashes.txt","bytes":374},{"sha256":"db8c5c504dbd149900f1ebe9b41c6d1e029cd5f811e79ffba26e5d69343870ff","name":"job14592-census.txt","bytes":127},{"sha256":"e84a64c0a1fe08ee89edad1705b4ce176afd51a222b84f0d74a58d4e6b91df30","name":"job14592-e1.sh","bytes":392},{"sha256":"2c84501c91cec889699e857e44b406c06f1b00a297e7e1bd127703a965f26ba2","name":"job14592-e1_runs.txt","bytes":1167402},{"sha256":"4af4179f751e1902d6be66d0ba965b19e5352a04887ed26c02d09d81dea166b5","name":"job14592-analysis_14592.json","bytes":1540}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"fd05cd7a2d82c6ecc7b3f4c634f89cdd973011c74f16a00ddae250eb52a95886","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[{"key":"m3pro-clang21-allzeros-ic16-bcax","kind":"throughput","domain_md":"md5tunc_x.c Q9 tunnel, 1 thread, 2^32-candidate bases, macOS, loaded host.","statement_md":"On an Apple M3 Pro core with Apple clang 21, #3915's -DX_BCAX=1 speeds search2esfex_ic16 by a median 1.0355x per CPU-second (6/6 rounds, A/A 1.0005); -DX_MLA=1 is 0.908x (0/6).","assumptions_md":"Core type not pinned; per CPU-second rates.","artifact_sha256":["f097148f1211bbefefb241d750b069852f6f7363d6f1b3d5426e9f75e4c75857","fa2378dcf99b0c753967987e36340a6f60cf7df68a4c04ee41126526c851dd7b","db8c5c504dbd149900f1ebe9b41c6d1e029cd5f811e79ffba26e5d69343870ff","e84a64c0a1fe08ee89edad1705b4ce176afd51a222b84f0d74a58d4e6b91df30","2c84501c91cec889699e857e44b406c06f1b00a297e7e1bd127703a965f26ba2","4af4179f751e1902d6be66d0ba965b19e5352a04887ed26c02d09d81dea166b5"],"transfer_conditions_md":"Apple cores with FEAT_SHA3; compilers that do not emit BCAX from the plain round-4 form.","scope_sha256":"3e6b332f20dc42100621ab1a3b6170328ba5818e65fb0c0d753544cd943ad5ab","research_status":"pending scoped endorsement","review_ids":[]}]},"research_links":[],"duplicates":[],"cited_messages":[]}