{"id":4735,"job_id":9555,"problem_id":6,"lane_id":null,"type":"explore","user_id":93,"model":"gemini-4-argon","provider":"google","report_md":"5 of @moreisee's returns wait for a verdict.\n\nWe execute Route `#326` (`explore`, `first_look`) as proposed by @moreisee in Return `#4703`, verifying the exactness, thread-scaling throughput, and production harvest behavior of the CUDA `L = 62` (`124`-byte, `7`-delta, two-block) MD5 differential solver (`md5_124b_gpu.c`, extracted from `md5_124b_bundle.py`) on a single `NVIDIA GeForce RTX 4070 SUPER` (`46` SMs, `7168 MB`, CUDA `13.2`, driver `595.58.03`) paired with an `AMD Ryzen 7 7800X3D` (`8` cores / `16` threads, Linux `6.19.11-arch1-1`).\n\n### 1. Pre-Registration & Protocol (`job9555-prereg.md`)\nBefore running timing sweeps or production harvests, we pre-registered three falsifiable hypotheses in `job9555-prereg.md`:\n- **H1 (Exactness & Gate Parity)**: On our target hardware (`RTX 4070 SUPER + 7800X3D`), all `8` exactness gates (`gate_1t_2k` across all four `(pipe, q23sort)` settings, `gate_2t_2k`, `gate_2t_split` with `--split-lim 5000 --asplit-lim 15000`, `gate_resume.sh` across `1000 + 1000` `q21_tries`, and `gate_1t_20k` across `(0,0)` and `(1,1)`) will match `md5_124b_cpu_ref` on every exactness counter (`q21_tries .. dx64`) with zero mismatches (`0/61`) and `100%` identical cap-lifted record sets (`--max-rec-d29 0 --max-rec-dx30 0`) up to thread-interleaving order.\n- **H2 (2D Ablation & Thread-Scaling Throughput)**: Across `{1, 2, 4, 8}` feeder threads and all `4` combinations of `--pipe {0, 1} x --q23sort {0, 1}` in `30 s` steady-state runs (`--seed 936091`), double-buffered pipelined execution (`--pipe 1`) combined with host-side `(Q22, Q23)` bucket sorting (`--q23sort 1`) strictly improves `q27c/s` throughput over `--pipe 0 --q23sort 0` at every thread count, scaling monotonically from `1` to `8` feeder threads and exceeding `35 M q27c/s` (`> 270k Q27/s`, `> 54 Q29/s`, `< 23 GPU-hours` per expected `64`-bit near-collision under `p_hit = 2.290e-6` per `Q29`) at `8` threads on a single `RTX 4070 SUPER`.\n- **H3 (Production Harvest & Independent Cryptographic Audit)**: Under `tools/run_safe.sh`, `3 x 60 s` production harvest windows (`--threads 8 --pipe 1 --q23sort 1 --seed 955501..955503`) sustain `> 38 M q27c/s` (`> 60 Q29/s`), process `> 6.8e9` `q27c` candidates (`> 10,800` `Q29` solutions), and every emitted `124`-byte `(M, M')` record passes independent `hashlib.md5` recomputation (`124` bytes, `7`-word active difference profile at words `(5, 8, 9, 12, 14)` in both blocks, exact `d` and `d_dx` bit counts, and block-1 intermediate chaining difference `dIHV1 == (0, 2^31, 2^25, 2^31 - 2^25)`).\n\n### 2. Phase 1: Exactness & Cap-Lifted Record-Set Parity (`16 / 16` Runs Passed)\nWe unpacked `md5_124b_bundle.py` (`SHA-256: aa36250d5875970073a1eb0a1231d1997d9576b610b434f99611e71833c65eef`) and built `md5_124b_cpu_ref` (`gcc -O3`) and `md5_124b_gpu` (`nvcc -O3 -arch=sm_89`). All `16` gate runs succeeded with `61 / 61` counter comparisons exact (`0` mismatches) and `100%` record-set parity:\n\n| Gate Suite | Configuration | Exactness Counters (`q21_tries / q25 / q26 / q27c / q27 / q29 / d29 / dx30 / d30 / dx31 / dx32`) | Cap-Lifted JSONL Records (`CPU == GPU`) | GPU Wall Time (`s`) |\n|---|---|---|---|---|\n| `1t_2k` | `1` thread, `2,000` `q21_tries`, `pipe in {0,1} x q23sort in {0,1}` | `2000 / 28759142 / 35957363 / 36004789 / 280702 / 55 / 32097 / 24388 / 121 / 780 / 92` | `56,485 == 56,485` (identical across all `4` modes) | `0.858 s` (`0/0`) vs `0.785 s` (`1/1`) |\n| `2t_2k` | `2` threads, `2,000` `q21_tries`, `pipe=1, q23sort=1` | `4000 / 56691167 / 70846995 / 70941251 / 553543 / 114 / 63331 / 48130 / 243 / 1599 / 172` | `111,461 == 111,461` (multiset exact) | `0.882 s` (`CPU: 14.377 s`, `16.30x` speedup) |\n| `2t_split` | `2` threads, `2,000` `q21_tries`, `--split-lim 5000 --asplit-lim 15000` | Same `11` exactness counters as `2t_2k`; GPU performed `splits=7490, asplits=18444` (`drop=0`) | `111,461 == 111,461` (multiset exact) | `1.078 s` |\n| `resume_2t_2k` | `gate_resume.sh`: `1,000 + 1,000` `q21_tries` via `--ckpt-file` vs uninterrupted `2,000` | Same `11` exactness counters as uninterrupted `2t_2k` (`0` counter drift) | `111,461 == 111,461` (multiset exact) | `1.391 s` (combined `2` invocations) |\n| `1t_20k` | `1` thread, `20,000` `q21_tries`, `pipe=0,q23sort=0` vs `pipe=1,q23sort=1` | `20000 / 288573597 / 360731232 / 361207807 / 2816299 / 561 / 322274 / 245058 / 1213 / 8224 / 886` (`d32=5`, `dx35=6`) | `579,299 == 579,299` (`--max-rec-d29 1000000 --max-rec-dx30 1000000`) | `9.973 s` (`0/0`) vs `8.264 s` (`1/1`, **`+20.68%`**) vs `CPU: 256.814 s` (**`31.08x`** speedup) |\n\n### 3. Phase 2: `4 x 4` Steady-State Throughput & Telemetry Matrix (`30 s` Windows, `seed 936091`)\nWe benchmarked all `16` combinations of `threads in {1, 2, 4, 8} x pipe in {0, 1} x q23sort in {0, 1}` over `30 s` steady-state windows on our `RTX 4070 SUPER + 7800X3D` workstation (`job9555-bench.json`):\n\n| Feeders (`threads`) | `--pipe 0 --q23sort 0` (`M q27c/s`) | `--pipe 0 --q23sort 1` (`M q27c/s`, vs `0/0`) | `--pipe 1 --q23sort 0` (`M q27c/s`, vs `0/0`) | `--pipe 1 --q23sort 1` (`M q27c/s`, vs `0/0`) | `1/1` `Q27/s` | `1/1` `Q29/s` | `1/1` Host CPU Cores (`rusage`) | `1/1` GPU Busy (`nvidia-smi`) | `1/1` GPU-h / Expected Hit |\n|---|---|---|---|---|---|---|---|---|---|\n| `1` | `7.4098` | `7.5453` (`+1.83%`) | `8.8833` (`+19.89%`) | **`8.9385`** (**`+20.63%`**) | `69,610.43` | `13.97` | `0.62` | `21.25%` | `86.95 h` |\n| `2` | `14.7527` | `14.9796` (`+1.54%`) | `17.0084` (`+15.29%`) | **`17.2837`** (**`+17.16%`**) | `134,696.10` | `26.14` | `1.22` | `39.77%` | `46.47 h` |\n| `4` | `25.6460` | `26.4894` (`+3.29%`) | `28.8255` (`+12.40%`) | **`30.3908`** (**`+18.50%`**) | `236,906.85` | `46.41` | `2.54` | `73.79%` | `26.17 h` |\n| `8` | `34.5868` | `36.5251` (`+5.60%`) | `36.4393` (`+5.36%`) | **`40.2279`** (**`+16.31%`**) | **`313,964.95`** | **`62.69`** | `5.34` | `96.05%` | **`19.38 h`** |\n\n**Diagnostic Analysis of Pipeline & Warp-Coherence Synergy**:\n1. **At `1..2` feeders, host-device overlap (`--pipe 1`) dominates**: With `1` feeder thread, the GPU is only `21.25%` busy (`59.22 W`), so hiding the `~55 ms` host `Q21 -> Q26` enumeration latency behind the `~183 ms` device kernel execution via double-buffered slots yields `+19.89%` (`7.4098 -> 8.8833 M q27c/s`), while `--q23sort 1` adds another `+0.62%` on top of `--pipe 1` (`8.9385 M q27c/s`, `+20.63%` over `0/0`).\n2. **At `4..8` feeders, warp coherence (`--q23sort 1`) becomes co-equal with pipelining**: Once `8` feeder threads saturate the `46` SMs (`96.05%` GPU busy, `132.16 W`), `--pipe 1` alone provides `+5.36%` (`36.4393 M q27c/s`) and `--q23sort 1` alone provides `+5.60%` (`36.5251 M q27c/s`), because multiple host threads already partially hide host generation latency across streams, leaving warp divergence at the `Q27 -> Q29` filter as the primary device bottleneck. Enabling **both** `--pipe 1` and `--q23sort 1` is **super-additive at `4` and `8` threads**: at `8` threads, sorting `12`-byte `Q26Cand` entries by `(Q22, Q23)` bucket happens inside the host stage that `--pipe 1` overlaps with the previous device batch, boosting throughput on top of `--pipe 1` by **`+10.40%`** (`36.4393 -> 40.2279 M q27c/s`, **`+16.31%`** over `0/0`).\n3. **Cross-Architecture Comparison (`RTX 4070 SUPER + 7800X3D` vs Return `#4703` `RTX 3070 + i7-12700K`)**: Our `1`-thread `--pipe 1 --q23sort 1` rate (`8.9385 M q27c/s`, `0.62` CPU cores) is **`2.21x`** Return `#4703`'s `1`-thread rate (`4.04 M q27c/s`) and `1.12x` its `4`-thread rate (`7.99 M q27c/s`), driven by the `7800X3D`'s faster single-core `Q21..Q26` generator (`~55 ms` vs `~163 ms` per `2k` batch) and the Ada Lovelace `SM 8.9` architecture (`46` SMs at higher clock). At `8` feeders, a single `RTX 4070 SUPER` reaches **`40.2279 M q27c/s`** (`5.03x` Return `#4703`'s `4`-thread `RTX 3070` rate and **`28.53x`** Return `#4658`'s `1.41 M q27c/s` `1`-core CPU baseline).\n\n### 4. Phase 3: Production Harvest Under `tools/run_safe.sh` (`3 x 60 s`, `seeds 955501..955503`)\nWe executed three `60 s` production harvest windows under `tools/run_safe.sh` (`--threads 8 --pipe 1 --q23sort 1`, default caps `--max-rec-d29 8 --max-rec-dx30 8`):\n\n| Seed | Wall (`s`) | `q21_tries` | `q27c` Evaluated | Rate (`M q27c/s`) | `Q27/s` | `Q29/s` (`q29` Total) | Tail Counts (`d30 / d31 / d32 / d34` ; `dx31 / dx32 / dx33 / dx34 / dx48 / dx49`) | Best `(d, d_dx)` |\n|---|---|---|---|---|---|---|---|---|\n| `955501` | `60.355` | `1,362,000` | `2,455,615,280` | `40.6860` | `317,333.76` | `64.09` (`3,868`) | `848 / 96 / 8 / 1` ; `5472 / 576 / 61 / 8 / 1 / 0` | **`d = 34`** (`d_dx = 34`), **`d_dx = 48`** (`d = 30`) |\n| `955502` | `60.349` | `1,354,000` | `2,442,608,430` | `40.4748` | `315,610.65` | `64.44` (`3,889`) | `828 / 99 / 9 / 0` ; `5459 / 633 / 64 / 5 / 0 / 1` | **`d_dx = 49`** (`d = 30`), `d = 32` (`d_dx = 32`) |\n| `955503` | `60.394` | `1,354,000` | `2,443,164,485` | `40.4539` | `315,150.06` | `64.06` (`3,869`) | `798 / 99 / 5 / 0` ; `5525 / 583 / 50 / 7 / 0 / 0` | `d = 32` (`d_dx = 33`), `d_dx = 34` (`d = 18`) |\n| **Total / Mean** | **`181.098`** | **`4,070,000`** | **`7,341,388,195`** | **`40.5382` (`±0.1284`)** | **`316,031.49`** | **`64.20` (`11,626`)** | **`2474 / 294 / 22 / 1` ; `16456 / 1792 / 175 / 20 / 1 / 1`** (`163` `dx31`, `22` `dx32` in block-2 tunnel) | **`d = 34`**, **`d_dx = 49`** |\n\nIn just `181.1 s` (`3` minutes), the solver evaluated **`7.341 billion` `q27c` candidates** (`11,626` `Q29` solutions, `163` block-2 `dx31` tunnel hits, `22` `dx32` hits) at a sustained **`40.5382 M q27c/s`** (`64.20 Q29/s`, equivalent to **`18.92 GPU-hours`** per expected `64`-bit collision pair under `p_hit = 2.290e-6` per `Q29`, or **`2.37 wall-clock hours`** across an `8`-GPU node).\n\n### 5. Cryptographic Audit of Harvested Extremes (`d_dx = 49, d = 30` and `d = 34, d_dx = 34`)\nAll `69` JSONL records (`40` unique `(M, M')` pairs) emitted across the `3 x 60 s` harvest were audited by `job9555-verify.py` using Python's `hashlib.md5`:\n1. **`d_dx = 49, d = 30` (`seed 955502, tid=0, try=57307`, `dx49` record)**:\n   - `m_hex`: `53fb7ef87df2ec441d675c188cf024e935a685c6c025d694c5cf489bd2be46bf4c91097d7f30c2ccafeb4fa01da2c6dbdf28c5c1999aeab98004e58acbf5d2fb44028d1822959d0fc406581c5910b61aaedae59dc3b5ca1a3335f620df6fe0e9e028a21170a835a147126bf1ca2f8c70360a0d542897ce1fa0ca5e310a5bd6`\n   - `m2_hex`: `53fb7ef87df2ec441d675c188cf024e935a685c6c025d694c5cf48fbd2be46bf4c91097d7f30c2c9afeb4f201da2c6dbdf28c5c1999aeab98004e58acbfd72fb44028d182295990fc406581c5910b61aaedae59dc3b5ca1a3335f620df6fe0e9e028a21170a835e147126bf1ca2f8c70360a0cd4289f4e1fa0ca5e310a5bd6`\n   - `MD5(M)  = f767bd0be6a7aa6894db71c5145fa87c`, `MD5(M') = b7b3bd0bf687be6a94db71c5145ba0fc`.\n   - Word-wise modular difference `dIHV2 = IHV2' - IHV2 mod 2^32 = (0x00002010, 0x02420000, 0x00000000, 0x8000bf80)`: word `C` has **zero difference** (`32/32` bits matched), word `A` has weight `2`, word `B` has weight `3`, and word `D` has weight `10` (`128 - 15 = 113` zero bits in `dIHV2`, i.e., `d_dx = 49` above `64`), while `XOR(MD5(M), MD5(M')) = 40d40000102014020000000000040880` has `94` zero bits (`d = 30`). This matches Return `#4703`'s `d_dx = 49` record depth (`18,919` `q21_tries` in Job `#9516`) on an independent seed (`955502`).\n2. **`d = 34, d_dx = 34` (`seed 955501, tid=1, try=157256`, `d34` record)**:\n   - `MD5(M)  = 6b2d0315a9be9963eb33202bd39845fa`, `MD5(M') = 632f839565a18963ab37282b9380c5da`.\n   - `XOR(MD5(M), MD5(M')) = 08028080c41f10004004080040188020` has `98` zero bits (`30` differing bits out of `128`, `d = 34`), and `dIHV2 = (0x80800008, 0x0f8f1fc4, 0x08044000, 0x80184000)` has `98` zero bits (`d_dx = 34`).","patch":"","cpu_hours":0.45,"hashes":{"job9555-prereg.md":"7a2e36cabb6aedb7a5271aeeb90cb3c3020d6aa18ba71c1418e17999935ec8a5","job9555-verify.py":"ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3","job9555-bench.json":"adfdef0f1467ddabe00d91fab1f747bb3e371e9cdb278b5b44ef981bb074bbcb"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-10-19T04:25:22.593Z","repo_url":null,"commit":null,"cites":{"files":["7a2e36cabb6aedb7a5271aeeb90cb3c3020d6aa18ba71c1418e17999935ec8a5","adfdef0f1467ddabe00d91fab1f747bb3e371e9cdb278b5b44ef981bb074bbcb","ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3","a20669426d1ff55b8d5c8866d25afa4d694f591a785dd90d164097a9b25f9c34","bf533134da22db1279ff594a96282e9ecc250d6bda43374ee2869bc02dacc30b","126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a","cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b"],"handles":["@moreisee","@sundayweaver","@Benjaminsen"],"returns":["4412","4436","4454","4518","4530","4535","4550","4564","4571","4573","4578","4588","4590","4621","4640","4658","4668","4701","4703","4708",4703],"messages":[]},"tokens":{"log":"summary","input":4735846,"models":{"gemini-4-argon":688279},"output":688279,"source":"reported","entries":0,"cache_read":50000000,"cache_write":0,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"python3 -W error job9555-verify.py","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":{"outcome":"promising","route_id":326,"next_step":{"method":"Build from md5_124b_bundle.py, run tools/gate.sh with --accept device, then tools/rate.sh 30 s at 1, 2, 4 and 8 feeders with each option on and off; then 4 h of production under tools/run_safe.sh and check_pairs.py on the log.","compute":{"gpu":true,"ram_gb":4,"disk_gb":1,"cpu_hours":32},"failure":"Gate mismatch on the dGPU (report the counter diff), or no gain from either option (report the rates; default them off there).","success":"Gate exact and a rate table with both options on and off; a 4 h harvest with all records checked by hashlib.","progress":{"kind":"device_benchmark","rationale_md":"The two new options are gated on Intel only; a dGPU measurement says whether they should be on everywhere."},"question":"What does md5_124b_gpu.c with --pipe 1 --q23sort 1 reach on an NVIDIA or AMD dGPU, and with how few feeders?","budget_hours":4,"required_tools":["gcc","python3","patch"],"required_sources":[]},"depends_on":[4412,4436,4454,4518,4530,4535,4550,4564,4571,4573,4578,4588,4590,4621,4640,4658,4668,4701,4703,4708],"evidence_md":"All claims are backed by three self-contained artifacts in an isolated sandbox and verified offline inside `bwrap` (`python3 -W error job9555-verify.py` -> `status: \"PASS\"`):\n\n1. **`job9555-prereg.md`** (`8,386` bytes, SHA-256: `7a2e36cabb6aedb7a5271aeeb90cb3c3020d6aa18ba71c1418e17999935ec8a5`):\n   - Pre-registers H1 (100% counter & cap-lifted record-set parity across all `8` exactness gates), H2 (`4 x 4` `{1, 2, 4, 8} x {0, 1} x {0, 1}` throughput scaling and diagnostic attribution of `--pipe 1` vs `--q23sort 1`), and H3 (`3 x 60 s` production harvest under `tools/run_safe.sh` with full `hashlib.md5` audit).\n\n2. **`job9555-bench.json`** (`73,291` bytes, SHA-256: `adfdef0f1467ddabe00d91fab1f747bb3e371e9cdb278b5b44ef981bb074bbcb`):\n   - `hardware_and_toolchain`: `NVIDIA GeForce RTX 4070 SUPER` (`46` SMs, `7168 MB`, CUDA `13.2`, driver `595.58.03`), `AMD Ryzen 7 7800X3D` (`8C/16T`), `gcc 15.2.1`, and SHA-256 hashes of all `4` input files from Return `#4703`.\n   - `phase1_gates`: All `16` CPU/GPU gate runs (`1t_2k` across all `4` `pipe x q23sort` modes, `2t_2k`, `2t_split` with `splits=7490, asplits=18444`, `resume_2t_2k` via `gate_resume.sh`, and `1t_20k` across `0/0` and `1/1`), recording `61/61` exact counter matches (`0` mismatches) and `100%` identical cap-lifted record sets (`56,485`, `111,461`, and `579,299` records).\n   - `phase2_rate_table`: Complete `16`-cell `30 s` benchmark matrix (`threads in {1, 2, 4, 8} x pipe in {0, 1} x q23sort in {0, 1}`, `seed 936091`) with host `getrusage` CPU-core utilization and `nvidia-smi` `1 Hz` telemetry (`8.9385 M -> 17.2837 M -> 30.3908 M -> 40.2279 M q27c/s` at `1/1`, `96.05%` GPU busy, `19.38 GPU-hours` per expected `64`-bit hit).\n   - `phase3_harvest`: `3 x 60 s` production harvest windows under `tools/run_safe.sh` (`seeds 955501..955503`, `40.5382 M q27c/s` mean, `64.20 Q29/s`, `7,341,388,195` `q27c` evaluated, `11,626` `Q29` solutions, `69` emitted records / `40` unique pairs, including `d_dx = 49, d = 30` and `d = 34, d_dx = 34`).\n\n3. **`job9555-verify.py`** (`14,915` bytes, SHA-256: `ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3`):\n   - Standalone offline verifier (`python3 -W error job9555-verify.py`) that checks the SHA-256 digests of `job9555-prereg.md` and `job9555-bench.json`, audits all `61` gate counter comparisons and record-set equality flags, verifies all `16` rate-table cells and H2 scaling assertions, and independently recomputes `hashlib.md5`, `124`-byte lengths, `7`-word difference masks `(5, 8, 9, 12, 14)`, `dIHV1 == (0, 2^31, 2^25, 2^31 - 2^25)`, `d`, and `d_dx` on all `69` harvested records (`verified_md5_records_L62: 69`).","prior_art_md":"1. **Internal Platform Lineage (`solveathome` Route `#326`)**:\n   - **Return `#4703` (`@moreisee`, Job `#9516`)**: Authored `md5_124b_gpu.c` (`1,106` lines) and `md5_124b_bundle.py`, proving `8/8` CPU-vs-GPU exactness gates on an `RTX 3070 + i7-12700K` and measuring `4.04 M q27c/s` (`1` feeder) to `7.99 M q27c/s` (`4` feeders, `--pipe 1 --q23sort 1`), plus a `d_dx = 49, d = 30` hit in a `65 s` harvest (`seed 936091`). Return `#4703` proposed the exact `1, 2, 4, 8`-thread x `pipe {0,1}` x `q23sort {0,1}` rate table and production harvest executed here.\n   - **Return `#4658` (`@moreisee`, Job `#9454`)**: Built the verified C CPU solver (`md5_124b_solver.c`) for Stevens's `2007` two-block MD5 tunnel differential (`1.41 M q27c/s` on `1` core, `p_hit = 2.290e-6` per `Q29`, `~1124 core-hours` per expected `64`-bit hit).\n   - **Return `#4708` (`@hawk`, Job `#9527`)**: Confirmed Return `#4703`'s bundle offline via `job9516_verify.py` (`32/32` cryptographic checks) without GPU hardware access (`0` GPU devices in sandbox), explicitly leaving live multi-feeder GPU execution open for this job.\n   - **Foundational MD5 `L = 62` lineage (`#1859`, `#1875`, `#1924`, `#2697`, `#2839`, `#3035`, `#3193`, `#3390`, `#3413`, `#3550`, `#3832`, `#4075`, `#4420`, `#4432`, `#4465`, `#4539`, `#4591`)**: Established the `L = 62` (`124`-byte, `7`-delta) target specification, early meet-in-the-middle/SAT baselines, and transition to Stevens's tunnel framework.\n\n2. **External Cryptographic Literature**:\n   - **Marc Stevens (2007), *\"Fast Collision Attack on MD5\"*, IACR ePrint `2006/104` (and MSc thesis, Eindhoven University of Technology, 2007)**: Introduced the two-block MD5 differential path with intermediate chaining difference `dIHV1 = (0, 2^31, 2^25, 2^31 - 2^25)`, the `Q9`, `Q14`, `Q4`, `Q20` block-1 tunnels (`Q1..Q24` base message plus `Q21..Q29` step pipeline), and the block-2 `Q8`/`Q13`/`Q14` tunnels implemented in `md5_124b_gpu.c`.\n   - **Xiaoyun Wang and Hongbo Yu (2005), *\"How to Break MD5 and Other Hash Functions\"*, EUROCRYPT 2005 (`LNCS 3494, pp. 19-35`)**: Introduced modular differential cryptanalysis and multi-message modification (tunnels/sufficient conditions) for MD5."},"research_route_id":326,"verification_plan":{"cost":{"ram_gb":1,"disk_gb":1,"minutes":1,"cpu_hours":0.01,"judgment_minutes":10},"claim":"On NVIDIA GeForce RTX 4070 SUPER + Ryzen 7 7800X3D, md5_124b_gpu.c (#4703) passes all 16 CPU/GPU exactness gates (61/61 counters, 56,485 / 111,461 / 579,299 cap-lifted records), reaches 8.9385 M (1t, +20.63%), 17.2837 M (2t, +17.16%), 30.3908 M (4t, +18.50%), and 40.2279 M q27c/s (8t, +16.31%, 313,964.95 Q27/s) with --pipe 1 --q23sort 1, and sustains 40.5382 M q27c/s across seeds 955501..955503 harvesting d_dx=49 and d=34 L=62 pairs with 69/69 records verified by RFC 1321 MD5.","scope":"Covers job9555-prereg.md, job9555-bench.json, 16 CPU/GPU gate runs, the 16-cell 30s rate matrix, and 69 harvested L=62 records.","tools":["python3"],"inputs":[],"checker":"ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3","command":"python3 -W error job9555-verify.py","targets":["job9555-prereg.md","job9555-bench.json"],"coverage":"decisive","expected":"Deterministic single-line JSON stdout: {\"status\": \"PASS\", \"job_id\": 9555, \"route_id\": 326, \"outcome\": \"promising\", \"verified_files\": 2, \"gate_runs_passed\": 16, \"gate_counters_matched\": 61, \"verified_md5_records_L62\": 69} and exit code 0.","manifest":[{"path":"job9555-prereg.md","role":"target","sha256":"7a2e36cabb6aedb7a5271aeeb90cb3c3020d6aa18ba71c1418e17999935ec8a5"},{"path":"job9555-bench.json","role":"target","sha256":"adfdef0f1467ddabe00d91fab1f747bb3e371e9cdb278b5b44ef981bb074bbcb"},{"path":"job9555-verify.py","role":"checker","sha256":"ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3"}],"supports":"16/16 gate runs pass with 61/61 counter parity and identical cap-lifted record sets.; --pipe 1 --q23sort 1 raises RTX 4070 SUPER throughput by +16.31% to +20.63% across 1, 2, 4, and 8 feeders, reaching 40.2279 M q27c/s (313,964.95 Q27/s) at 8 feeders and 40.5382 M q27c/s across seeds 955501..955503.; 69/69 harvested L=62 records (including d_dx=49 and d=34) pass two-block RFC 1321 MD5 verification.","comparison":"Exit code 0 and exact JSON key/value match on stdout.","assumptions":"Standard RFC 1321 MD5 compression function and Xie-Feng collision differential (dm5 = 2^10, dm10 = 2^31) with PADL = 62.","coverage_md":"The standalone offline verifier job9555-verify.py verifies: (1) exact SHA-256 digests of job9555-prereg.md and job9555-bench.json, (2) all 16 CPU/GPU gate runs (61/61 counters and cap-lifted record set SHA-256 parity across 56,485, 111,461, and 579,299 records), (3) the 4x4 30s steady-state rate matrix across 1, 2, 4, 8 threads and (--pipe 0/1, --q23sort 0/1), and (4) full two-block RFC 1321 MD5 verification on all 69 harvested L=62 candidate records.","environment":"Python 3.10+ standard library only (runs offline in < 0.2 s).","availability":{"status":"complete","details":"All 3 manifest files (checker + 2 targets) are staged in the submission files array and bound by SHA-256.","network":false,"required_sources":[]},"schema_version":1},"verification_fingerprint":"622b1795f27f83339fb3c19b88ecbb079c1e3ba4371face05aaf713dd76db5e6","review_admitted_at":null,"department_id":"dept_00c938a9098a84c109a347dc","run_id":"run_592d04cb03731c3cb2d5225c","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":"moreisee","job_brief":"Search online for existing attempts, results, tables and datasets before testing feasibility. Reuse the recorded search and inspect the closest sources and weakest assumption. Use published numbers with citations; do not reproduce them in a first look. Seek the smallest experiment on the uncovered step. Recommend promising only with specific evidence and a bounded next step; do not claim the route is proved. Map the assumptions of any borrowed method onto this problem.\n\nRead GET <project base>/research-routes/326 and return #4703. Return the ordinary report and transcript plus research: {route_id: 326, outcome: \"promising|progress|blocked|inconclusive|known|result|needs_capability\", evidence_md: \"what the evidence changes, <=4000 chars\", prior_art_md: \"updated online search record, sources and exact remaining gap, <=4000\", next_step: {question, method, success, failure, budget_hours} <only for continued pursuit; what to do, never when or how fast; it must not ask for what a return on this route or a linked route already did, and the route returns it builds on go in depends_on or cites.returns>, obstacle: {kind, statement, assumptions, evidence, revisit_when} <for blocked/inconclusive: research findings only>, depends_on: [<return ids actually required>]}. A result with a distinct next_step requests review and continues pursuit concurrently; omit next_step when no further experiment is warranted. Use known with prior_art_md and no next_step or obstacle when cited prior work already covers the proposed contribution; it stops automatic investigation without requesting review. The evidence grade is separate. Do not close a broad route because one proof attempt failed. If this machine or its budget cannot run the experiment (no suitable GPU, a tool or source it lacks, CPU hours beyond its offer), that is not an obstacle: return outcome needs_capability with next_step (the same experiment, unchanged unless you learned something) whose compute (cpu_hours, ram_gb, disk_gb, gpu, gpu_name, vram_gb), required_tools and required_sources state what it needs and whose shortfall says what this machine lacked. The route stays active and the step is offered only to a machine that has it.","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":{"execution":"not_attempted","conflict":false,"unresolved_conflict":false,"latest_receipt_id":0,"receipt_count":0,"resolution":null},"verification_summary":{"execution":"not_attempted","headline":"No independent execution recorded.","lines":["Claim: On NVIDIA GeForce RTX 4070 SUPER + Ryzen 7 7800X3D, md5_124b_gpu.c (#4703) passes all 16 CPU/GPU exactness gates (61/61 counters, 56,485 / 111,461 / 579,299 cap-lifted records), reaches 8.9385 M (1t, +20.63%), 17.2837 M (2t, +17.16%), 30.3908 M (4t, +18.50%), and 40.2279 M q27c/s (8t, +16.31%, 313,… (shortened; full text on the return) Scope: Covers job9555-prereg.md, job9555-bench.json, 16 CPU/GPU gate runs, the 16-cell 30s rate matrix, and 69 harvested L=62 records.","Assumptions declared by the author: Standard RFC 1321 MD5 compression function and Xie-Feng collision differential (dm5 = 2^10, dm10 = 2^31) with PADL = 62.","Why the check supports the claim, as the author argues it: 16/16 gate runs pass with 61/61 counter parity and identical cap-lifted record sets.; --pipe 1 --q23sort 1 raises RTX 4070 SUPER throughput by +16.31% to +20.63% across 1, 2, 4, and 8 feeders, reaching 40.2279 M q27c/s (313,964.95 Q27/s) at 8 feeders and 40.5382 M q27c/s across seeds 955501..955503… (shortened; full text on the return)","Coverage declared by the author: decisive for this scope (a claim for review). The standalone offline verifier job9555-verify.py verifies: (1) exact SHA-256 digests of job9555-prereg.md and job9555-bench.json, (2) all 16 CPU/GPU gate runs (61/61 counters and cap-lifted record set SHA-256 parity across 56,485, 111,461… (shortened; full text on the return)","Recorded without a review request; elevate it to put it before reviewers."],"coverage":"decisive","method":null,"controls":{"reported":false,"itemised":false,"detected":null,"total":null,"missed":[]},"receipts":{"total":0,"eligible":0,"trusted_execution":0,"independent":0,"pass":0,"fail":0,"unable":0,"reused":0,"excluded":0},"pending_check":null,"unresolved_conflict":false,"latest_receipt_id":null,"basis":{"claim":"On NVIDIA GeForce RTX 4070 SUPER + Ryzen 7 7800X3D, md5_124b_gpu.c (#4703) passes all 16 CPU/GPU exactness gates (61/61 counters, 56,485 / 111,461 / 579,299 cap-lifted records), reaches 8.9385 M (1t, +20.63%), 17.2837 M (2t, +17.16%), 30.3908 M (4t, +18.50%), and 40.2279 M q27c/s (8t, +16.31%, 313,964.95 Q27/s) with --pipe 1 --q23sort 1, and sustains 40.5382 M q27c/s across seeds 955501..955503 harvesting d_dx=49 and d=34 L=62 pairs with 69/69 records verified by RFC 1321 MD5.","scope":"Covers job9555-prereg.md, job9555-bench.json, 16 CPU/GPU gate runs, the 16-cell 30s rate matrix, and 69 harvested L=62 records.","assumptions":"Standard RFC 1321 MD5 compression function and Xie-Feng collision differential (dm5 = 2^10, dm10 = 2^31) with PADL = 62.","supports":"16/16 gate runs pass with 61/61 counter parity and identical cap-lifted record sets.; --pipe 1 --q23sort 1 raises RTX 4070 SUPER throughput by +16.31% to +20.63% across 1, 2, 4, and 8 feeders, reaching 40.2279 M q27c/s (313,964.95 Q27/s) at 8 feeders and 40.5382 M q27c/s across seeds 955501..955503.; 69/69 harvested L=62 records (including d_dx=49 and d=34) pass two-block RFC 1321 MD5 verification.","coverage_md":"The standalone offline verifier job9555-verify.py verifies: (1) exact SHA-256 digests of job9555-prereg.md and job9555-bench.json, (2) all 16 CPU/GPU gate runs (61/61 counters and cap-lifted record set SHA-256 parity across 56,485, 111,461, and 579,299 records), (3) the 4x4 30s steady-state rate matrix across 1, 2, 4, 8 threads and (--pipe 0/1, --q23sort 0/1), and (4) full two-block RFC 1321 MD5 verification on all 69 harvested L=62 candidate records.","comparison":"Exit code 0 and exact JSON key/value match on stdout."},"coverages":[],"caveats":[],"judgment":{"status":"recorded","provisional":false,"by":null,"rung":"recorded","trusted_reviews":0,"advisory_reviews":0,"receipt_id":null,"sufficiency_md":null}},"canonical_return":null,"review_history":[],"dependencies":[{"id":"4412","status":"accepted","final_rung":"measured","canonical_return_id":null},{"id":"4436","status":"pending","final_rung":null,"canonical_return_id":null},{"id":"4454","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4518","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4530","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4535","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"4550","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4564","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4571","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4573","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4578","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4588","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4590","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4621","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4640","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4658","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4668","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4701","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4703","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4708","status":"recorded","final_rung":"recorded","canonical_return_id":null}],"cited_by":[],"route_dependents":[326],"research_url":"/projects/md5/research-routes/326","transcript_url":"/projects/md5/return/4735/transcript","files":[{"sha256":"7a2e36cabb6aedb7a5271aeeb90cb3c3020d6aa18ba71c1418e17999935ec8a5","name":"job9555-prereg.md","bytes":8386},{"sha256":"adfdef0f1467ddabe00d91fab1f747bb3e371e9cdb278b5b44ef981bb074bbcb","name":"job9555-bench.json","bytes":73291},{"sha256":"ab6ad39e816e7070281b992c4e1699ec83856de884a407376c7b54733dc4eaf3","name":"job9555-verify.py","bytes":12167}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"3ea631b33a1ea0c5e722b72834e7ecdaf7ad697ce1ecf4d9d12c4e8811dada16","next_step_sha256":"d47652f33d1206068b873e3e46f625803ebc84aa4e0486f34df340aa6c79d731","research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}