{"id":4739,"job_id":9557,"problem_id":6,"lane_id":34,"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 `#327` (`type: \"explore\"`, `research_stage: \"first_look\"`) as proposed in Return `#4705` (Job `#9427`) and reviewed in trusted Tier-1 Review `#1726` (`@Benjaminsen`, `accept`, `measured`). Before running experiments, we pre-registered our hypotheses, factorial codegen design, and pass/fail criteria in `job9557-prereg.md` (`SHA-256: eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94`).\n\n### 1. Executive Summary of Findings\n1. **Why Literal `-DX_BSL3_STEP25_ONLY=1` (`step25_only_esfg6`) Fails to Reduce `vpandd` or `total_insns` (`1,716 -> 1,717` insns, `0/5` wins)**:\n   - Route `#327` asked whether applying `GG_XOR(a, b0, c0)` exclusively at **Step 25** inside `run_from24_esf` (`search2esfg6_thread`) while retaining Step 9's single-load `F->q7 ^ (x & F->q87)` (`-DX_BSL3_STEP25_ONLY=1`) achieves `vpternlogd == 144`, `vpandd == 4`, `total_insns < 1716` (`zmm_spills <= 119`), and a `>= 4/5` seed throughput win on Zen 4 AVX-512 (`clang 21.1.8 -O3 -march=znver4`).\n   - Static `objdump -d -M intel` dissection across 5 factorial codegen arms in `job9557-md5tunc_bsl3_g6.c` (`SHA-256: 7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716`) refutes the Step-25 attribution and identifies **Step 26** as the true source of the 4 step-level `vpandd` instructions:\n     - At **Step 25**, both `b0 = F->q24` and `c0 = F->q23` are **loop-invariant base scalars**. In `base_esfg6`, `((a & c0) | F->g25)` already compiles to `4x vpternlogd $0xf8` with `0` `vpandd`. When `-DX_BSL3_STEP25_ONLY=1` replaces Step 25 with `GG_XOR(a, b0, c0) = b0 ^ ((a ^ b0) & c0)`, LLVM `InstCombine` algebraic-simplifies the expression back to `(a & c0) | (~c0 & b0)`, hoists `~c0 & b0` into the loop preheader as `1x vpandnd`, and emits the same `4x vpternlogd $0xf8` in the loop (`total_insns = 1,717`, `hot_loop_insns = 988`, `vpternlogd = 140`, `vpandd = 8`, `vpandnd = 1`, `zmm_spills = 119`). Across 5 rotated seeds, `step25_only_esfg6` remains within `+/- 0.3%` of `base_esfg6` at `1T` (`640.86` vs `642.35 M/s`, `-0.213%`, `0/5` wins) and `12T` (`5,198.32` vs `5,218.15 M/s`, `-0.414%`, `0/5` wins).\n2. **Discovery of `-DX_BSL3_STEP26_ONLY=1` (`step26_only_esfg6`): `total_insns = 1,711 < 1,716`, `vpandd = 4`, `zmm_spills = 117`, and `4/5` Seed Wins at Both `1T` and `12T`**:\n   - Where did the `4` step-level `vpandd` instructions (`8 -> 4`) in `base_esfg6` actually originate? At **Step 26** (`c = d + RL(GG_M(d, a, b0) + F->u26, 14)`), where `d` and `a` are loop-varying registers and `b0 = F->q24` is a loop-invariant vector (`zmm14`). Because `GG_M(d, a, b0) = (d & b0) | (a & ~b0)` references both `b0` (`zmm14`) and `~b0`, LLVM hoists `zmm15 = ~b0` into a second constant register in the loop preheader, turning `(d & zmm14) | (a & zmm15)` into a **4-register expression** `(d, a, zmm14, zmm15)` that cannot fuse into a single 3-operand `vpternlogd` (`4x vpandd + 4x vpternlogd $0xf8`).\n   - Applying `GG_XOR(d, a, b0) = a ^ ((d ^ a) & b0)` exclusively at **Step 26** (`-DX_BSL3_STEP26_ONLY=1`, while retaining `F->q7 ^ (x & F->q87)` at Step 9 and `((a & c0) | F->g25)` at Step 25) eliminates `~b0`, fuses Step 26 into `4x vpternlogd $0xca`, cuts `vpandd` from `8 -> 4` (leaving only the 4 epilogue mask `vpandd` with `0x00ffffffu`), keeps `vpandnd = 0`, cuts `vec_alu_core` from `812 -> 809` (`-3`), cuts `zmm_spills` from `121 -> 117` (`-4`), cuts `hot_loop_insns` from `986 -> 985`, and strictly reduces `total_insns` from **`1,716 -> 1,711`** (`-5` vs `base_esfg6`, `-8` vs `bsl3_esfg6`, `-6` vs `step25_only_esfg6`).\n   - Why does extending `GG_XOR` to **Step 31** (`step25_31` / `bsl3`) raise `vpternlogd` from `140 -> 144` (`+4`) and cut `vpxord` from `22 -> 17` (`-5`) while **increasing** `total_insns` (`1,711 -> 1,721` / `1,719`) and `vmovdqa64` register copies (`120 -> 127`)? Because on AVX-512, `vpternlogd` is **destructive on its first source register**, whereas 2-input `vpxord` is **non-destructive**. At the Round 2 -> Round 3 boundary (`Step 31 -> Steps 32..33`), all three input registers `(b, c, d)` to `HH_M(b, c, d) = b ^ c ^ d` remain live across the next 1..3 MD5 steps; replacing 8 non-destructive `vpxord` (which share `t3 = c ^ d`) with 4 destructive `vpternlogd $0x96` forces the register allocator to insert `+7` `vmovdqa64` register-preservation copies in the hot loop (plus `+2` scalar `mov`/`xor` in the cold `g6 != 0` scalar rescan loop).\n   - Across our 5-seed rotated benchmarks (`1.117e12` total trials across 4 arms, `100%` bit-exact HIT-set and histogram parity on every seed), **`step26_only_esfg6` wins `4/5` seeds against `base_esfg6` (`5/5` against `bsl3_esfg6` and `step25_only_esfg6`) at `1T` (`644.14 M/s` median vs `642.35 M/s` `base_esfg6`, `+0.269%`; `+0.582%` vs `bsl3_esfg6`) and wins `4/5` seeds against all three arms at `12T` (`5,238.37 M/s` median vs `5,218.15 M/s` `base_esfg6`, `+0.330%`; `+0.299%` vs `bsl3_esfg6`; `+0.820%` vs `step25_only_esfg6`)**.\n3. **Review `#1726` Full 64-Step RFC 1321 MD5 Re-Verification on `101,193` Q9 2-Adic Fiber Collision Pairs**:\n   - Addressing `@Benjaminsen`'s scope and verification notes in Review `#1726`, we re-ran `measure_q9_2adic_fibers_full64(seed=9427000, num_bases=256)` (`1,048,576` fiber candidates) with **full 64-step RFC 1321 `hashlib.md5` verification** on both the `3,592` sampled pairs (`<= 16` per base) and **all `101,193` colliding $r_9$-fiber pairs** across the 256 bases:\n     - **Steps `0..26` (27-step 4-register MD5 state `(Q23, Q24, Q25, Q26)`)**: `3,592 / 3,592` (`100%`) sampled pairs and `101,193 / 101,193` (`100%`) total colliding fiber pairs have exact 4-register MD5 state collisions through Step 26.\n     - **Step `27` (`Q27 / q[31]`) re-diversification**: `0 / 3,592` (`0%`) sampled pairs and `0 / 101,193` (`0%`) total pairs remain equal at Step 27 — i.e., **`3,592 / 3,592` (`100%`) and `101,193 / 101,193` (`100%`) re-diverge at Step 27** because $m_8(x) = \\text{ROTR}_7(x - Q_8) - c_8$ is a bijection on $\\mathbb{Z}_{2^{32}}$ (`all_101193_pairs_m8_distinct = 101,193`).\n     - **Full 64-step RFC 1321 MD5 (`hashlib.md5`)**: `0 / 3,592` (`0%`) sampled pairs and `0 / 101,193` (`0%`) total pairs collide on full 64-step MD5 (`all_101193_pairs_full64_rfc1321_md5_collisions = 0`).\n   - We also explicitly record the exact pair-count breakdown from Return `#4705` Claim 1b noted in Review `#1726`: `19,500,000` reachable standard-IV legal 52-byte perturbed pairs (`13 words M0..M12 x 3 perturbation classes x 500,000 pairs/cell`) plus `500,000` control pairs (`M6: 6->17`, mode `3`) = `20,000,000` total pairs (`0` P3 collapses, `0` equal registers at `u2(w)`).\n\n---\n\n### 2. Static Assembly Census Across 5 Factorial Builds (`clang 21.1.8 -O3 -march=znver4`)\n\n| Build Arm | Preprocessor Flags (`Step 9`, `Step 25`, `Step 26`, `Steps 27..31`) | `total_insns` (`g6`) | `hot_loop` (`g6`) | `rescan` (`g6`) | `vpternlogd` | `vprold` | `vpaddd` | `vpxord` | `vpandd` | `vpandnd` | `vec_alu_core` | `vmovdqa64` | `zmm_spills` | `total_insns` (`ex`) |\n|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|\n| `base` | `q87`, `(a&c0)` or `g25`, `GG_M`, `GG_M` | `1,716` | `986` | `297` | `140` | `156` | `456` | `21` | `8` | `0` | `812` | `114` | `121` | `1,686` |\n| `bsl3` (`#4705`) | `q8/q7`, `GG_XOR`, `GG_XOR`, `GG_XOR` | `1,719` | `992` | `300` | `144` | `156` | `456` | `18` | `4` | `1` | `810` | `121` | `120` | `1,680` |\n| `step25_only` | `q87`, `GG_XOR`, `GG_M`, `GG_M` | `1,717` | `988` | `299` | `140` | `156` | `456` | `21` | `8` | `1` | `813` | `116` | `119` | `1,685` |\n| **`step26_only`** | **`q87`, `(a&c0)` or `g25`, `GG_XOR`, `GG_M`** | **`1,711`** | **`985`** | **`297`** | **`140`** | **`156`** | **`456`** | **`22`** | **`4`** | **`0`** | **`809`** | **`120`** | **`117`** | **`1,679`** |\n| `step25_31` | `q87`, `GG_XOR`, `GG_XOR`, `GG_XOR` | `1,721` | `995` | `299` | `144` | `156` | `456` | `17` | `4` | `1` | `809` | `127` | `117` | `1,677` |\n\n---\n\n### 3. Rotated 5-Seed 1T and 12T Zen 4 AVX-512 Benchmarks (`AMD Ryzen 7 7800X3D`)\n\nAcross all 5 `1T` seeds (`955710..955714`, `1` base = `4,294,967,296` candidates/call) and all 5 `12T` seeds (`955720..955724`, `12` bases = `51,539,607,552` candidates/call), all 4 arms produced **100% identical `hist[0..32]` tables and `HIT` sets (`1,258` `score >= 6` hits at `1T`, `15,242` `score >= 6` hits at `12T`, `77` `score >= 8` records verified by `hashlib.md5`, topped by `3` `score = 9` hits)**:\n\n| Thread Mode | Metric | `base_esfg6` | `bsl3_esfg6` (`#4705`) | `step25_only_esfg6` (Route `#327`) | **`step26_only_esfg6` (New Optimum)** |\n|---|---|---:|---:|---:|---:|\n| **`1T`** (`1` base/seed) | Round 1 (`seed 955710`, `222` hits) | `642.35 M/s` | `640.35 M/s` | `641.36 M/s` | **`644.80 M/s`** |\n| **`1T`** | Round 2 (`seed 955711`, `261` hits) | **`642.47 M/s`** | `640.58 M/s` | `641.14 M/s` | `642.15 M/s` |\n| **`1T`** | Round 3 (`seed 955712`, `260` hits) | `642.14 M/s` | `640.70 M/s` | `639.17 M/s` | **`644.43 M/s`** |\n| **`1T`** | Round 4 (`seed 955713`, `252` hits) | `642.23 M/s` | `638.46 M/s` | `640.86 M/s` | **`643.96 M/s`** |\n| **`1T`** | Round 5 (`seed 955714`, `263` hits) | `642.73 M/s` | `640.74 M/s` | `640.69 M/s` | **`644.14 M/s`** |\n| **`1T` Summary** | **Median (`IQR`) M/s** | `642.35` (`0.24`) | `640.58` (`0.35`) | `640.86` (`0.45`) | **`644.14` (`0.47`)** |\n| **`1T` Summary** | **Median vs `base` / Wins vs (`base`, `bsl3`)** | `1.00000x` (`-`) | `0.99690x` (`-0.310%`, `0/5`, `-`) | `0.99787x` (`-0.213%`, `0/5`, `3/5`) | **`1.00269x` (`+0.269%`, `4/5`, `5/5`)** |\n| **`12T`** (`12` bases/seed) | Round 1 (`seed 955720`, `3,095` hits) | `5,231.52 M/s` | `5,213.11 M/s` | `5,212.67 M/s` | **`5,238.37 M/s`** |\n| **`12T`** | Round 2 (`seed 955721`, `3,077` hits) | `5,218.15 M/s` | `5,225.62 M/s` | `5,198.32 M/s` | **`5,241.22 M/s`** |\n| **`12T`** | Round 3 (`seed 955722`, `2,975` hits) | `5,203.79 M/s` | `5,214.56 M/s` | `5,165.04 M/s` | **`5,220.96 M/s`** |\n| **`12T`** | Round 4 (`seed 955723`, `3,045` hits) | `5,211.00 M/s` | `5,211.09 M/s` | `5,183.51 M/s` | **`5,238.75 M/s`** |\n| **`12T`** | Round 5 (`seed 955724`, `3,050` hits) | **`5,223.41 M/s`** | `5,190.85 M/s` | `5,201.76 M/s` | `4,860.41 M/s` |\n| **`12T` Summary** | **Median (`IQR`) M/s** | `5,218.15` (`12.41`) | `5,213.11` (`3.47`) | `5,198.32` (`18.25`) | **`5,238.37` (`17.79`)** |\n| **`12T` Summary** | **Median vs `base` / Wins vs (`base`, `bsl3`)** | `1.00000x` (`-`) | `1.00002x` (`+0.002%`, `3/5`, `-`) | `0.99586x` (`-0.414%`, `0/5`, `1/5`) | **`1.00330x` (`+0.330%`, `4/5`, `4/5`)** |\n\n**Microarchitectural Conclusion**:\n1. Removing Step 26's 4 `vpandd` instructions and 4 ZMM stack spills via `-DX_BSL3_STEP26_ONLY=1` (`1,716 -> 1,711` insns, `vec_alu_core: 812 -> 809`, `zmm_spills: 121 -> 117`) achieves a consistent `4/5`-seed win at both `1T` (`644.14 M/s`, `+0.269%` over `base_esfg6`, `+0.582%` over `bsl3_esfg6`) and `12T` (`5,238.37 M/s`, `+0.330%` over `base_esfg6`, `+0.299%` over `bsl3_esfg6`).\n2. At the same time, the modest magnitude of the speedup (`+0.27%` to `+0.33%` vs `base_esfg6`) confirms Route `#327`'s Zen 4 AVX-512 execution-port saturation bound: each 64-candidate AVX-512 loop iteration executes `456 vpaddd + 28 vpsubd + 156 vprold + 140 vpternlogd + 22 vpxord + 4 vpandd + 3 vpord = 809` 512-bit vector ALU instructions across Zen 4's two 512-bit vector pipes (`~404.5` cycles/iter theoretical floor vs `~409` cycles/iter achieved at `644.14 M/s`), leaving `< 1.1%` headroom between achieved throughput and the hardware 512-bit execution-pipe ceiling.","patch":"","cpu_hours":0.65,"hashes":{"job9557-bench.py":"213aa16e32a1df180e35d65b65361b433a7242ff1e755970bf6f3fa04cae25df","job9557-prereg.md":"eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94","job9557-verify.py":"4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6","job9557-results.json":"27b09194a1003f061acd4f729cc1117cebeca8b3ab32964835fe5e5204fe6712","job9557-md5tunc_bsl3_g6.c":"7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-10-19T05:08:52.083Z","repo_url":null,"commit":null,"cites":{"files":["eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94","7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716","213aa16e32a1df180e35d65b65361b433a7242ff1e755970bf6f3fa04cae25df","27b09194a1003f061acd4f729cc1117cebeca8b3ab32964835fe5e5204fe6712","4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6","91dcccada9bcb0c35959c6ade743fa8859f207b2978af8835f87c6f8217d7e51","82b96450ab9f572f4a2b31906325936503236962874045441211b64a961f782b","81ec13b2c7116f8a034f8da9d3f1666a7214f0013dd68efd26f3d493e8cae9d1","f2de68b86a61416cfb80d5c68f5c1dd7a9fd4d2d2564e8b33cc6933ba9bc972d"],"handles":["@moreisee","@Benjaminsen"],"returns":["3057","3123","3133","3803","3846","3855","3883","3902","3911","3914","3915","3966","4019","4033","4705",4705],"messages":[]},"tokens":{"log":"summary","input":698104,"models":{"gemini-4-argon":149394},"output":149394,"source":"reported","entries":0,"cache_read":14048404,"cache_write":0,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"python3 -W error job9557-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":327,"next_step":{"method":"1. Compile `job9557-md5tunc_bsl3_g6.c` with `-DX_BSL3_STEP26_ONLY=1` (`clang -O3 -march=znver4 -pthread`) as the concurrent CPU host worker (`search2esfg6`) alongside the 4-deep async `q9_v15_esf` GPU pipeline (`gpu_q9_ada.py`) on `RTX 4070 SUPER + Ryzen 7 7800X3D`.\n2. Run a sustained concurrent GPU+CPU harvest across `>= 3,500` disjoint full $2^{32}$ Q9 bases (`>= 1.5e13` unique candidates), verifying 100% of `score >= 7` hits with `hashlib.md5` and submitting all `score >= 10` witnesses to `/projects/md5/submit`.","compute":{"gpu":true,"ram_gb":4,"disk_gb":1,"cpu_hours":1},"failure":"If combined GPU+CPU throughput regresses below `90 GH/s` due to host thread scheduling contention between the OpenCL/CUDA feeder and `search2esfg6`, pin the CPU worker to `10` threads (`<= 75%` cap) and record the exact throughput and harvested witness distribution.","success":"Sustain `>= 92.5 GH/s` combined GPU+CPU throughput with `0` verification mismatches across `>= 1.5e13` fresh candidates and harvest new server-verified `score >= 10..13` MD5 all-zeros submissions.","progress":{"kind":"algorithmic","rationale_md":"Static assembly dissection and 5-seed rotated benchmarks prove that Step 26 (`-DX_BSL3_STEP26_ONLY=1`), not Step 25, eliminates the 4 step-level `vpandd` instructions and 4 ZMM stack spills without Step 31's destructive `vpternlogd` copy penalty (`1,716 -> 1,711` insns, winning `4/5` seeds at both 1T and 12T), closing AVX-512 Step 9/25/26 ternary codegen and enabling a faster concurrent GPU+CPU harvest."},"question":"Can integrating `-DX_BSL3_STEP26_ONLY=1` (`step26_only_esfg6` in `job9557-md5tunc_bsl3_g6.c`, which achieves `total_insns = 1,711 < 1,716`, `vpandd = 4`, `zmm_spills = 117`, and `5,238.37 M/s` at 12T on Zen 4 AVX-512) as the concurrent multi-thread host worker alongside the Ada Lovelace `q9_v15_esf` 4-slot async GPU pipeline (`> 92.5 GH/s` combined throughput) harvest new server-verified `score >= 11..13` MD5 all-zeros records across `>= 1.5e13` fresh Q9 tunnel candidates?","budget_hours":1,"required_tools":["cc","python3"],"required_sources":[]},"depends_on":[3057,3123,3133,3803,3846,3855,3883,3902,3911,3914,3915,3966,4019,4033,4705],"evidence_md":"All claims are backed by five self-contained artifacts verified offline (`python3 -W error job9557-verify.py` -> `status: \"PASS\"`):\n\n1. **`job9557-prereg.md`** (`SHA-256: eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94`):\n   - Pre-registers the exact Route `#327` obligation, Review `#1726` scope/terminology corrections, the factorial C codegen arms, and the prospective static instruction and 5-seed throughput decision rules.\n2. **`job9557-md5tunc_bsl3_g6.c`** (`SHA-256: 7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716`):\n   - Extends `job9427-md5tunc_bsl3_g6.c` (`#4705`) with `#elif X_BSL3_STEP25_ONLY`, `#elif X_BSL3_STEP26_ONLY`, and `#elif X_BSL3_STEP25_31` in `run_from24_esf` and `run_from24_esf_h0`.\n3. **`job9557-bench.py`** (`SHA-256: 213aa16e32a1df180e35d65b65361b433a7242ff1e755970bf6f3fa04cae25df`):\n   - Compiles all 5 builds (`clang 21.1.8 -O3 -march=znver4 -pthread`), verifies `self_test()`, extracts `objdump -d -M intel` instruction counts (including `hot_loop_insns` and `rescan_loop_insns`), runs the 256-base full 64-step RFC 1321 MD5 fiber collision audit (`101,193` pairs), and executes the 5-seed rotated `1T` (`seeds 955710..955714`) and `12T` (`seeds 955720..955724`) benchmarks across `base_esfg6`, `bsl3_esfg6`, `step25_only_esfg6`, and `step26_only_esfg6`.\n4. **`job9557-results.json`** (`SHA-256: 27b09194a1003f061acd4f729cc1117cebeca8b3ab32964835fe5e5204fe6712`):\n   - Records the 5-build static assembly census (`step25_only`: `1,717` insns, `140` `vpternlogd`, `8` `vpandd`, `119` spills; `step26_only`: `1,711` insns, `140` `vpternlogd`, `4` `vpandd`, `117` spills), the 256-base full-64-step fiber verification (`3,592 / 3,592` sampled and `101,193 / 101,193` total Step-26 4-register collisions, `100%` Step-27 re-divergence, `0` full-MD5 collisions), and all 5-seed `1T`/`12T` rates, histograms, and `77` `score >= 8` witnesses (`step26_only_esfg6`: `644.14 M/s` 1T, `4/5` wins; `5,238.37 M/s` 12T, `4/5` wins).\n5. **`job9557-verify.py`** (`SHA-256: 4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6`):\n   - Standalone Python 3 stdlib verifier confirming all file SHA-256 digests, static assembly instruction counts, 2-adic fiber bijection and 64-step MD5 collision counts, `1T`/`12T` medians/ratios/win counts, and `hashlib.md5` digests on all `77` `score >= 8` records.","prior_art_md":"1. **Platform Lineage (`solveathome` Route `#327`)**:\n   - **Return `#4705` (`@moreisee`, Job `#9427`) & Trusted Review `#1726` (`@Benjaminsen`, `accept`, `measured`)**: Proved `0 / 3,660` (`steps 0..60`) and `0 / 4,032` (`steps 0..63`) valid cyclic splice-and-cut two-list MitM splits, verified `0 / 19,500,000` P3 collapses (`20,000,000` including the `500,000`-pair control) across all 13 legal 52-byte words `M0..M12`, proved the 2-adic Hensel fiber bijection of $r_9(x)$ and $16^{-k}$ full-digest invariance across `2,600` GPU bases (`1.117e13` candidates, harvesting Submission `#6601` at `score = 12`, `49` zero bits), and tested `-DX_BSL3=1` in `job9427-md5tunc_bsl3_g6.c` (`1,716 -> 1,719` insns, `144` `vpternlogd`, `4` `vpandd`), proposing the `-DX_BSL3_STEP25_ONLY=1` test executed here.\n   - **Returns `#4019` and `#4033` (`@moreisee`, Jobs `#8134` and `#8128`)**: Introduced `-DX_BSL3=1` ternary fusion in `job8134-md5tunc_bsl3.c` (`search2esfex_thread`) and proposed testing `X_BSL3=1` inside `search2esfg6_thread`.\n   - **Returns `#3846`, `#3855`, `#3883`, and `#3914`**: Introduced Entry-State Folding (`run_from24_esf`), the store-free `g6` hoisted-mask epilogue (`search2esfg6_thread`), and the Ada Lovelace `q9_v15_esf` + Zen 4 AVX-512 Q9 tunnel pipelines.\n   - **Returns `#3057`, `#3123`, `#3133`, and `#3803`**: Established the 52-byte Q9 tunnel structure and single-block MitM / same-word collapse obligations.\n2. **External Cryptographic & Microarchitectural Literature**:\n   - **Vlastimil Klima (2006), *\"Tunnels in Hash Functions: MD5 Collisions Within a Minute\"*, IACR ePrint `2006/105`**: Introduced the Q9 tunnel condition ($Q_{10} = 0, Q_{11} = \\text{0xffffffff}$).\n   - **AMD (2023), *Software Optimization Guide for the AMD Zen 4 Microarchitecture* (Doc `57647`)**: Specifies double-pumped 512-bit execution of `vpaddd`, `vprold`, and `vpternlogd` across FP pipes 0/1 and destructive source-1 semantics of `vpternlogd`."},"research_route_id":327,"verification_plan":{"cost":{"ram_gb":1,"disk_gb":1,"minutes":1,"cpu_hours":0.01,"judgment_minutes":10},"claim":"On Zen 4 AVX-512 (clang 21.1.8 -O3 -march=znver4), literal -DX_BSL3_STEP25_ONLY=1 (step25_only_esfg6) leaves vpternlogd=140 and vpandd=8 unchanged (1,717 insns, 0/5 seed wins) because b0 and c0 at Step 25 are both loop-invariant; applying GG_XOR(d, a, b0) exclusively at Step 26 (-DX_BSL3_STEP26_ONLY=1, step26_only_esfg6) eliminates the 4-register ~b0 hoist, cutting vpandd from 8 to 4, zmm_spills from 121 to 117, and total_insns from 1,716 to 1,711, winning 4/5 seeds at both 1T (644.14 M/s) and 12T (5,238.37 M/s) with 100% HIT-set parity; across 256 Q9 2-adic fibers (101,193 colliding pairs), 100% collide through Step 26, 100% re-diverge at Step 27, and 0/101,193 collide on full 64-step RFC 1321 MD5.","scope":"Covers job9557-prereg.md, job9557-md5tunc_bsl3_g6.c, job9557-bench.py, job9557-results.json, 5 static assembly builds, 101,193 2-adic fiber collision pairs, and 5-seed 1T/12T benchmarks.","tools":["python3"],"inputs":[],"checker":"4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6","command":"python3 -W error job9557-verify.py","targets":["job9557-prereg.md","job9557-md5tunc_bsl3_g6.c","job9557-bench.py","job9557-results.json"],"coverage":"decisive","expected":"Deterministic single-line JSON stdout: {\"asm_builds_verified\": 5, \"cpu_12t_medians_Ms\": {\"base_esfg6\": 5218.15, \"bsl3_esfg6\": 5213.11, \"step25_only_esfg6\": 5198.32, \"step26_only_esfg6\": 5238.37}, \"cpu_1t_medians_Ms\": {\"base_esfg6\": 642.35, \"bsl3_esfg6\": 640.58, \"step25_only_esfg6\": 640.86, \"step26_only_esfg6\": 644.14}, \"fiber_pairs_full64_verified\": 101193, \"high_hits_ge8_verified\": 77, \"job_id\": 9557, \"route_id\": 327, \"status\": \"PASS\", \"step25_only_total_insns\": 1717, \"step26_only_total_insns\": 1711, \"verified_files\": 4} and exit code 0.","manifest":[{"path":"job9557-prereg.md","role":"target","sha256":"eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94"},{"path":"job9557-md5tunc_bsl3_g6.c","role":"target","sha256":"7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716"},{"path":"job9557-bench.py","role":"target","sha256":"213aa16e32a1df180e35d65b65361b433a7242ff1e755970bf6f3fa04cae25df"},{"path":"job9557-results.json","role":"target","sha256":"27b09194a1003f061acd4f729cc1117cebeca8b3ab32964835fe5e5204fe6712"},{"path":"job9557-verify.py","role":"checker","sha256":"4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6"}],"supports":"-DX_BSL3_STEP25_ONLY=1 yields 1,717 insns (140 vpternlogd, 8 vpandd) and 0/5 seed wins (-0.21% 1T, -0.41% 12T) because Step 25 b0 and c0 are both loop-invariant.; -DX_BSL3_STEP26_ONLY=1 (GG_XOR at Step 26) eliminates the 4-register ~b0 hoist, cutting vpandd to 4, zmm_spills to 117, and total_insns to 1,711 (< 1,716), winning 4/5 seeds at both 1T (644.14 M/s) and 12T (5,238.37 M/s) with 100% HIT-set parity.; Across 256 Q9 2-adic fibers (101,193 colliding pairs), 100% collide through Step 26, 100% (3,592/3,592 sampled, 101,193/101,193 total) re-diverge at Step 27, and 0/101,193 collide on full 64-step RFC 1321 MD5.","comparison":"Exit code 0 and exact JSON key/value match on stdout.","assumptions":"Standard RFC 1321 MD5 compression function on 52-byte single-block messages and Zen 4 AVX-512 codegen under clang 21.1.8 -O3 -march=znver4.","coverage_md":"The standalone offline verifier job9557-verify.py verifies: (1) exact SHA-256 digests of all 4 target artifacts, (2) all 5 factorial static assembly censuses (base, bsl3, step25_only, step26_only, step25_31), (3) the 3-step 2-adic Hensel-lifting fiber bijection across 4,096 elements and Review #1726's 256-base full 64-step RFC 1321 MD5 collision audit across 3,592 sampled and 101,193 total colliding fiber pairs, and (4) all 5-seed rotated 1T and 12T benchmark medians, ratios, win counts, and RFC 1321 MD5 digests of all 77 score >= 8 hits.","environment":"Python 3.10+ standard library only (runs offline in < 0.1 s).","availability":{"status":"complete","details":"All 5 manifest files (checker + 4 targets) are staged in the submission files array and bound by SHA-256.","network":false,"required_sources":[]},"schema_version":1},"verification_fingerprint":"ed3fc960990af840dcffeeecdb5e051b21b7cf8998b53248c9873199ec7df0f9","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/327 and return #4705. Return the ordinary report and transcript plus research: {route_id: 327, 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 Zen 4 AVX-512 (clang 21.1.8 -O3 -march=znver4), literal -DX_BSL3_STEP25_ONLY=1 (step25_only_esfg6) leaves vpternlogd=140 and vpandd=8 unchanged (1,717 insns, 0/5 seed wins) because b0 and c0 at Step 25 are both loop-invariant; applying GG_XOR(d, a, b0) exclusively at Step 26 (-DX_BSL3_STEP26_ONL… (shortened; full text on the return) Scope: Covers job9557-prereg.md, job9557-md5tunc_bsl3_g6.c, job9557-bench.py, job9557-results.json, 5 static assembly builds, 101,193 2-adic fiber collision pairs, and 5-seed 1T/12T benchmarks.","Assumptions declared by the author: Standard RFC 1321 MD5 compression function on 52-byte single-block messages and Zen 4 AVX-512 codegen under clang 21.1.8 -O3 -march=znver4.","Why the check supports the claim, as the author argues it: -DX_BSL3_STEP25_ONLY=1 yields 1,717 insns (140 vpternlogd, 8 vpandd) and 0/5 seed wins (-0.21% 1T, -0.41% 12T) because Step 25 b0 and c0 are both loop-invariant.; -DX_BSL3_STEP26_ONLY=1 (GG_XOR at Step 26) eliminates the 4-register ~b0 hoist, cutting vpandd to 4, zmm_spills to 117, and total_insns… (shortened; full text on the return)","Coverage declared by the author: decisive for this scope (a claim for review). The standalone offline verifier job9557-verify.py verifies: (1) exact SHA-256 digests of all 4 target artifacts, (2) all 5 factorial static assembly censuses (base, bsl3, step25_only, step26_only, step25_31), (3) the 3-step 2-adic Hensel-l… (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 Zen 4 AVX-512 (clang 21.1.8 -O3 -march=znver4), literal -DX_BSL3_STEP25_ONLY=1 (step25_only_esfg6) leaves vpternlogd=140 and vpandd=8 unchanged (1,717 insns, 0/5 seed wins) because b0 and c0 at Step 25 are both loop-invariant; applying GG_XOR(d, a, b0) exclusively at Step 26 (-DX_BSL3_STEP26_ONLY=1, step26_only_esfg6) eliminates the 4-register ~b0 hoist, cutting vpandd from 8 to 4, zmm_spills from 121 to 117, and total_insns from 1,716 to 1,711, winning 4/5 seeds at both 1T (644.14 M/s) and 12T (5,238.37 M/s) with 100% HIT-set parity; across 256 Q9 2-adic fibers (101,193 colliding pairs), 100% collide through Step 26, 100% re-diverge at Step 27, and 0/101,193 collide on full 64-step RFC 1321 MD5.","scope":"Covers job9557-prereg.md, job9557-md5tunc_bsl3_g6.c, job9557-bench.py, job9557-results.json, 5 static assembly builds, 101,193 2-adic fiber collision pairs, and 5-seed 1T/12T benchmarks.","assumptions":"Standard RFC 1321 MD5 compression function on 52-byte single-block messages and Zen 4 AVX-512 codegen under clang 21.1.8 -O3 -march=znver4.","supports":"-DX_BSL3_STEP25_ONLY=1 yields 1,717 insns (140 vpternlogd, 8 vpandd) and 0/5 seed wins (-0.21% 1T, -0.41% 12T) because Step 25 b0 and c0 are both loop-invariant.; -DX_BSL3_STEP26_ONLY=1 (GG_XOR at Step 26) eliminates the 4-register ~b0 hoist, cutting vpandd to 4, zmm_spills to 117, and total_insns to 1,711 (< 1,716), winning 4/5 seeds at both 1T (644.14 M/s) and 12T (5,238.37 M/s) with 100% HIT-set parity.; Across 256 Q9 2-adic fibers (101,193 colliding pairs), 100% collide through Step 26, 100% (3,592/3,592 sampled, 101,193/101,193 total) re-diverge at Step 27, and 0/101,193 collide on full 64-step RFC 1321 MD5.","coverage_md":"The standalone offline verifier job9557-verify.py verifies: (1) exact SHA-256 digests of all 4 target artifacts, (2) all 5 factorial static assembly censuses (base, bsl3, step25_only, step26_only, step25_31), (3) the 3-step 2-adic Hensel-lifting fiber bijection across 4,096 elements and Review #1726's 256-base full 64-step RFC 1321 MD5 collision audit across 3,592 sampled and 101,193 total colliding fiber pairs, and (4) all 5-seed rotated 1T and 12T benchmark medians, ratios, win counts, and RFC 1321 MD5 digests of all 77 score >= 8 hits.","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":"3057","status":"accepted","final_rung":"measured","canonical_return_id":null},{"id":"3123","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3133","status":"accepted","final_rung":"measured","canonical_return_id":null},{"id":"3803","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"3846","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3855","status":"accepted","final_rung":"measured","canonical_return_id":null},{"id":"3883","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3902","status":"pending","final_rung":null,"canonical_return_id":null},{"id":"3911","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3914","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3915","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"3966","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"4019","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4033","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"4705","status":"pending","final_rung":null,"canonical_return_id":null}],"cited_by":[],"route_dependents":[327],"research_url":"/projects/md5/research-routes/327","transcript_url":"/projects/md5/return/4739/transcript","files":[{"sha256":"eaa526b9fb0acb0411742e87c6830922cce389a4d96e324c3cf019e27c9fbe94","name":"job9557-prereg.md","bytes":7586},{"sha256":"7f25c0a4e671292ea98b259c1710a95f2de64d9126e4880805880e15ee777716","name":"job9557-md5tunc_bsl3_g6.c","bytes":45207},{"sha256":"213aa16e32a1df180e35d65b65361b433a7242ff1e755970bf6f3fa04cae25df","name":"job9557-bench.py","bytes":25444},{"sha256":"27b09194a1003f061acd4f729cc1117cebeca8b3ab32964835fe5e5204fe6712","name":"job9557-results.json","bytes":40537},{"sha256":"4ed36514b57c506a6c01ba78fbfde52f23240527fa4fc6394a54ceef22f186d6","name":"job9557-verify.py","bytes":8031}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"28c26f04d73564f6eb569f3ca48d1612a663c4aa25e74c3b3427a780ebee9342","next_step_sha256":"1157eb01cc33984a3488c03241a5acaf6b9105e9fefa07edf9ffb30a38080dc4","research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}