{"id":2978,"job_id":6242,"problem_id":6,"lane_id":null,"type":"measure","user_id":73,"model":"claude-opus-5-5","provider":"anthropic","report_md":"# Smallest collision: long-run seeded md5sbc with stall-triggered witness jumps sustains 138 L=63-class Q29/s (0.70x unconstrained), about 2x #2976. A 126-byte single-block pair now prices at ~1.4x Stevens (~2^50.3)\n\n**Measured** (step following #2976). No collision; the record stays 248 and the published reference 128.\n\n## Method\n- Same instances as #2976.\n- Z3 produced **30** distinct L=63 class + Q23 witnesses per instance (all replay-verified).\n- md5sbc (#2974's build plus private changes, not redistributed) loads a witness list, starts at the first witness, and when no class Q23 pass has occurred for **3 s** unwinds its Q12/Q11/Q10 loops and jumps to the next witness, keeping the table.\n- Class-filtered, 120 s of search time per instance, sandboxed (525 CPU-s including table builds).\n\n## Results\n| instance | class Q23 | class Q29 | search s | jumps | class Q29/s | #2976 (20 s bursts) |\n|---|---|---|---|---|---|---|\n| 0x6163002f | 9 | 32,768 | 115 | 1 | **284** | 98 |\n| 0x61630021 | 5 | 16,384 | 113 | 1 | **145** | 132 |\n| 0x61630027 | 27 | 16,384 | 127 | 1 | **129** | 57 |\n| 0x61630023 | 41 | 0 | 120 | 29 | 0 | 0 |\n| **total** | | **65,536** | **475** | | **138** | 67.5 |\n\n- **Sustained class Q29 rate: 138/s, which is 0.70x the unconstrained median (197/s) and 2.0x #2976.**\n- **Honest attribution:** the three productive instances jumped only once each, so most of the gain comes from running uninterrupted for 120 s rather than in 20 s bursts. Witness jumping mattered only for 0x61630023, where 29 jumps found 41 class Q23 passes but **none reached Q29**: a tunnel-stage failure specific to that instance.\n- Q29 counts are taken at md5sbc's 4,096-step reports, so they slightly understate the true count at the time limit.\n\n## Pricing (estimate)\nWith downstream odds unchanged (#2972), a 126-byte (63 + 63) pair costs about Stevens' ~2^49.8 / 0.70, i.e. **~2^50.3 compression-equivalents (~1.4x Stevens)**, plus the cost of choosing instances (Z3 seconds; about half of instances are sat, #2968, and some sat instances fail in the tunnels, like 0x61630023). This revises #2976's ~3x. It is still multi-CPU-year scale on this hardware, so no pair is produced here.\n\n## Next step\nDetect tunnel-dead instances (Q23 passes but no Q29 within N passes) and skip them; then run a long multi-instance measurement or a GPU port. The latter is compute-heavy, so it is the person's decision.\n\n## OUTCOMES.md entry (proposed)\n| Track | Method | Budget and hardware | Best reached | What it shows |\n|---|---|---|---|---|\n| Smallest collision | md5sbc seeded at Z3 L=63 witnesses, long runs + 3 s stall jumps (4 instances) | ~9 CPU-min | 248 (unchanged) | Sustained class Q29 138/s = 0.70x unconstrained: 126-byte pair ~1.4x Stevens (~2^50.3) |","patch":null,"cpu_hours":0.15,"hashes":{"stall_restart_results.json":"e920fdc8fd6775fa4b3d8acd071657058196fba511fa941b48575a8e7239043a"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-10-11T11:29:20.408Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[2976,2974,2972,2968,2891,2661],"messages":[]},"tokens":{"log":"summary","input":10,"models":{"claude-opus-5-5":9717},"output":9717,"source":"reported","entries":0,"cache_read":4501012,"cache_write":11925,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"1. `python multi_witness_z3.py tables.json instances.txt 0x61630021,0x61630023,0x61630027,0x6163002f 30`, then write wl_<seed>.txt (one line per witness: Q9 Q10 Q11 Q12 hex). Supplied.\n2. md5sbc as #2974. Add: load WFILE into a list and seed from its first entry. In the per-sample block, if STALL > 0 and no class Q23 pass for STALL seconds, set a jump flag, set Q12cur = 0 and break. Before the Q11/Q10 loop ends, if jumping set cur = 0. Before the main Q9 loop end, if jumping, load the next witness into Q9cur and the first-pass seeds.\n3. Run `WFILE=wl_<seed>.txt STALL=3 FILTER63=1 ./md5sbc_jump 120 <seed>` for each seed. Logs are in stall_restart_logs.txt and the aggregate in stall_restart_results.json.","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":"2026-10-11T11:29:20.408Z","department_id":"dept_ef09d64fbbd7ddb34ab67f81","run_id":"run_c2ccb63b450f473296a41c24","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,"handle":"danieljmt","job_brief":"If the seeded class-filtered md5sbc search is restarted from a fresh Z3 witness whenever it produces no class Q23 pass within a short window (e.g. 3 s), does the sustained class Q29 rate per search-CPU-second rise above #2976's 67.5/s (0.34x unconstrained) on the same instances?\n\nWhy this step: #2976 showed productivity is lumpy: many witnesses stall while a few expand to thousands of Q29 states; cutting stalls early should raise the average and lower the ~3x Stevens price of a 126-byte pair.\n\nStop when: Sustained class Q29/s measured with stall-restart on >= 3 instances over >= 300 search-CPU-s total, or 20 CPU-minutes.","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":2979,"handle":"danieljmt","status":"pending"},{"id":2981,"handle":"danieljmt","status":"pending"},{"id":2984,"handle":"danieljmt","status":"pending"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/md5/return/2978/transcript","files":[{"sha256":"b4dbdaedb4f1896e02aba069500d3072af79c7c314c510b8ce56b1ea2a41fc3e","name":"multi_witnesses.json","bytes":22882},{"sha256":"f884e759a41ce14b4c59ca16a16b22b0edbabe13fcb0a2c300925bbc359141ac","name":"stall_restart_logs.txt","bytes":3499},{"sha256":"e920fdc8fd6775fa4b3d8acd071657058196fba511fa941b48575a8e7239043a","name":"stall_restart_results.json","bytes":818},{"sha256":"675667956fc5959575a7643f5e38056b52b0f4d7b6f03fa2eb243564a78f7a41","name":"wl_0x61630021.txt","bytes":1080},{"sha256":"5a2f2882669e6ee82e2ac21845be75db8aff4204ca50ef89f446f9c4ed8f16e9","name":"wl_0x61630023.txt","bytes":1080},{"sha256":"1f438487985a645ab8c46eda3c67cdb691707d2d8b6a3deb8c3eeb6f7a1956da","name":"wl_0x61630027.txt","bytes":1080},{"sha256":"3761d88c0be18f69bf9482484b6a59dab5dbe7a28d107083745284309ab81602","name":"wl_0x6163002f.txt","bytes":1080}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"a2fc1fa9ded625737efda46cb309f305319758dca9281c0f0f29cfe5d4c80dc0","research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}