{
 "date": "2026-10-22",
 "change": "opt-in mixed-role directional hybrid (AFDEP_MIXED_HYBRID=1): FFN ranks [0, F), 0 < F < U, share NVLink domain 0 with Attention ranks; per-rank roles through generated copies of hybrid_dispatch.cuh and hybrid_combine.cuh; every rank's reply inbox in a slice before the staging slice (docs/design.md section 18)",
 "decision": "opt-in prototype (stage 1, dense); the section-17 sparse switches do not apply yet. Not yet run: fault injection for the reply-slice and staging orderings, two-buffer interleave, 8-GPU nodes, serving.",
 "native_source_sha256": {
  "control_default": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663",
  "mixed": [
   "81b71eb093c583acd6cd0972df9232769c02fb202f8662eab7b9553c1ee336b7"
  ],
  "mixed_first_screen": "d1c3c18a"
 },
 "setup": "kermit 6 pods x 4 H200, InfiniBand, NCCL GIN, FP8, hidden 2048, top-8, capacity 2048, 24 transport SMs, CUDA graphs; benchmarks.comm with benchmarks.screen's pairing, one torchrun per pod (2 fresh processes per arm, rotated, max over Attention ranks, round trip checked, per-arm digest checked)",
 "screens": {
  "f2_11to1_e128": {
   "arms": {
    "union": {
     "_label": "cf"
    },
    "mixed": {
     "_label": "xm",
     "_env": {
      "AFDEP_MIXED_HYBRID": "1"
     }
    },
    "masked": {
     "_label": "cf",
     "role_aware": false
    }
   },
   "table": [
    {
     "tokens": 1,
     "union": {
      "median_us": 212.11495876312256,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 162.9769229888916,
      "paired_change_pct": [
       -23.156460629547038,
       -23.17502545436262
      ],
      "geomean_change_pct": -23.165743602662925,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 211.25396251678467,
      "paired_change_pct": [
       -0.44344140827431433,
       -0.36849118031025974
      ],
      "geomean_change_pct": -0.40597334483620795,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 8,
     "union": {
      "median_us": 235.40435791015625,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 189.00795459747314,
      "paired_change_pct": [
       -19.784700680197666,
       -19.633854206862
      ],
      "geomean_change_pct": -19.709312868954488,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 235.73264122009277,
      "paired_change_pct": [
       -0.1182573547562571,
       0.3968927204495598
      ],
      "geomean_change_pct": 0.13898641930667477,
      "repeats_faster": 1
     }
    },
    {
     "tokens": 32,
     "union": {
      "median_us": 264.9659585952759,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 207.46880531311035,
      "paired_change_pct": [
       -22.00116015260507,
       -21.397281183591378
      ],
      "geomean_change_pct": -21.699802832107295,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 264.55219745635986,
      "paired_change_pct": [
       -0.27432303687419113,
       -0.03751449620433833
      ],
      "geomean_change_pct": -0.15598897388651167,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 128,
     "union": {
      "median_us": 333.4525489807129,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 284.0547561645508,
      "paired_change_pct": [
       -14.73787202993112,
       -14.890275054720414
      ],
      "geomean_change_pct": -14.81410762466715,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 332.4079132080078,
      "paired_change_pct": [
       -0.486612774297146,
       -0.13979034081249564
      ],
      "geomean_change_pct": -0.3133523873201982,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 512,
     "union": {
      "median_us": 763.7957763671875,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 701.6360282897949,
      "paired_change_pct": [
       -8.089239089285593,
       -8.187168601137051
      ],
      "geomean_change_pct": -8.13821689496429,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 760.4345703125,
      "paired_change_pct": [
       -0.3295431075050925,
       -0.5502993058012717
      ],
      "geomean_change_pct": -0.4399823924651214,
      "repeats_faster": 2
     }
    }
   ]
  },
  "f1_23to1_e64": {
   "arms": {
    "union": {
     "_label": "cf"
    },
    "mixed": {
     "_label": "xm",
     "_env": {
      "AFDEP_MIXED_HYBRID": "1"
     }
    },
    "masked": {
     "_label": "cf",
     "role_aware": false
    }
   },
   "table": [
    {
     "tokens": 1,
     "union": {
      "median_us": 280.3237581253052,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 222.24304676055908,
      "paired_change_pct": [
       -21.32081651206712,
       -20.111836695100493
      ],
      "geomean_change_pct": -20.718631071479756,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 279.23532485961914,
      "paired_change_pct": [
       -0.8796621215414024,
       0.10772616778573774
      ],
      "geomean_change_pct": -0.3871913758319301,
      "repeats_faster": 1
     }
    },
    {
     "tokens": 8,
     "union": {
      "median_us": 319.3677234649658,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 267.6208782196045,
      "paired_change_pct": [
       -16.175755450078533,
       -16.229962176034142
      ],
      "geomean_change_pct": -16.202863196214846,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 319.5514392852783,
      "paired_change_pct": [
       0.4448910630371339,
       -0.3285553830316035
      ],
      "geomean_change_pct": 0.0574204976668824,
      "repeats_faster": 1
     }
    },
    {
     "tokens": 32,
     "union": {
      "median_us": 353.41163635253906,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 293.9847993850708,
      "paired_change_pct": [
       -16.65545142445205,
       -16.97422742813225
      ],
      "geomean_change_pct": -16.81499212515257,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 352.86996841430664,
      "paired_change_pct": [
       0.10222718970640532,
       -0.4076433477740893
      ],
      "geomean_change_pct": -0.15303353651956453,
      "repeats_faster": 1
     }
    },
    {
     "tokens": 128,
     "union": {
      "median_us": 437.4517250061035,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 383.40487480163574,
      "paired_change_pct": [
       -12.379181888680058,
       -12.330678101164494
      ],
      "geomean_change_pct": -12.354933350241415,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 436.14283561706543,
      "paired_change_pct": [
       -0.30664294950349946,
       -0.2917731024535075
      ],
      "geomean_change_pct": -0.2992083031984061,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 512,
     "union": {
      "median_us": 1082.700366973877,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 1015.5105590820312,
      "paired_change_pct": [
       -6.139889318423608,
       -6.271536516870347
      ],
      "geomean_change_pct": -6.205736014718677,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 1080.2915573120117,
      "paired_change_pct": [
       -0.1451984771154935,
       -0.29964966995971176
      ],
      "geomean_change_pct": -0.22245395897819709,
      "repeats_faster": 2
     }
    }
   ]
  },
  "f2_first_screen_d1c3c18a": {
   "table": [
    {
     "tokens": 1,
     "union": {
      "median_us": 211.8291997909546,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 163.5779571533203,
      "paired_change_pct": [
       -22.657220951436607,
       -22.89991514690246
      ],
      "geomean_change_pct": -22.778663392695584,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 211.02980136871338,
      "paired_change_pct": [
       -0.5401650408152214,
       -0.21406501532086342
      ],
      "geomean_change_pct": -0.3772484578741908,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 8,
     "union": {
      "median_us": 235.64239978790283,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 188.9050006866455,
      "paired_change_pct": [
       -19.693961713149832,
       -19.973485123349988
      ],
      "geomean_change_pct": -19.833845248462612,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 234.68811988830566,
      "paired_change_pct": [
       -0.4862295466841249,
       -0.32407250931596465
      ],
      "geomean_change_pct": -0.4051840303457954,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 32,
     "union": {
      "median_us": 265.06792545318604,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 207.48535633087158,
      "paired_change_pct": [
       -21.584755443495172,
       -21.862323532066807
      ],
      "geomean_change_pct": -21.72366251982638,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 264.08055782318115,
      "paired_change_pct": [
       -0.2827716724466467,
       -0.46200919159818943
      ],
      "geomean_change_pct": -0.37243073974297936,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 128,
     "union": {
      "median_us": 333.9589214324951,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 284.3815612792969,
      "paired_change_pct": [
       -14.821450736131037,
       -14.869281083110375
      ],
      "geomean_change_pct": -14.845369267838027,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 332.3055171966553,
      "paired_change_pct": [
       -0.5678582505827312,
       -0.42220974380478005
      ],
      "geomean_change_pct": -0.4950606459772522,
      "repeats_faster": 2
     }
    },
    {
     "tokens": 512,
     "union": {
      "median_us": 762.7810287475586,
      "paired_change_pct": [
       0.0,
       0.0
      ],
      "geomean_change_pct": 0.0,
      "repeats_faster": 0
     },
     "mixed": {
      "median_us": 701.7377853393555,
      "paired_change_pct": [
       -7.951627391941862,
       -8.053752021463389
      ],
      "geomean_change_pct": -8.002703877552742,
      "repeats_faster": 2
     },
     "masked": {
      "median_us": 760.1337432861328,
      "paired_change_pct": [
       -0.21402378476860218,
       -0.4799234081724002
      ],
      "geomean_change_pct": -0.34706228248905235,
      "repeats_faster": 2
     }
    }
   ]
  },
  "union_profile_kernel_durations": "nsys kernel traces (not in the repository): hybrid_dispatch ~87-113 us on every GPU kind at 1 and 32 tokens; hybrid_combine 97-112 us (T=1) and 145-169 us (T=32)",
  "union_channel_sweep": "docs/validation-sparse-hybrid.json screens.kermit_mixed_role_11to1: 8 SMs instead of 24 give -17.1% at 1 token and +55.8% at 512"
 },
 "gpu_suite_81b71eb0": [
  "F=2 FP8 capacity 8",
  "F=2 BF16 capacity 8",
  "F=2 BF16 capacity 64 hidden 2048 24 SMs",
  "F=1 (8 experts) FP8 capacity 8",
  "F=3 (12 experts) FP8 capacity 8"
 ],
 "gpu_suite_d1c3c18a": [
  "F=2 FP8 capacity 64 hidden 2048 24 SMs"
 ],
 "gpu_suite_scope": "24 ranks per run, 16 eager cases, 32 graph replays, 832 checked graph exchanges, cases incl. empty, inactive, all-skipped, skew, partial prefix, one small source (spread routing), delayed replies and dispatches",
 "unit": "tests/unit 41/41; Ruff check and format clean",
 "design": "three designs (hub-domain, rail-role, relay composition) from a map of the union protocol, each with an independent adversarial review; stage 1 implements the dense core of the hub-domain / rail-role design",
 "scope_limits": "measured and suite-validated on 6 x 4 H200 only (4-GPU NVLink domains, one FFN-holding domain, F = 1..3); 8-GPU nodes, reply/staging fault injection, two-buffer interleave and serving not yet run",
 "published_trees": {
  "mixed_hybrid_commit": "c065385b1a3927a9c0398f07b99bb415f9d7a67731f90a6d1531ade4457720aa (measured as 81b71eb0)",
  "branch_head": "34fa5eac5093fd9a398edade5f44f88ad32c834ee17314e3770817e4ab4601c0 (comment fixes)",
  "_note": "each published tree differs from the measured one only in comments and text (renamed comments and the release's comment fixes)"
 }
}