{
 "date": "2026-09-27",
 "model": "DeepSeek V4.1 Flash Q4",
 "model_bytes": 518596067328,
 "model_sha256": "a5e2e2c3ada4b2e98d9f9e4b50f6d9c2a12c2c96f5da165c07e13aff9264984e",
 "engine": "argonautlabsai/ds4-argodrive @ 9ad4a610ad1a0c0ac5a10f6036af6f49434bbb5c, tag v41-stack-20260927 (fork of antirez/ds4)",
 "hardware": "Apple M5 Max, 128 GB; internal SSD plus two NVMe enclosures over Thunderbolt 5",
 "replicas": "three byte-identical verified copies; prefill split 10:5:5, decode split 10:6:6; Engram on the internal SSD",
 "prompt_tokens": 512,
 "generated_tokens": [
  512,
  200
 ],
 "mode": "raw completion, greedy, context allocation 4096",
 "control": {
  "binary": "built from tag v41-router-qualified-20260921",
  "profile": "v41-router-20260921",
  "ssd_streaming_cache_experts": 4200
 },
 "candidate": {
  "profile": "v41-stack-20260927",
  "adds": {
   "DS4_ARGODRIVE_FLAG_READBACK": "1",
   "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
   "DS4_ARGODRIVE_PSO_CACHE": "1"
  },
  "ssd_streaming_cache_experts": 4600,
  "engine_paths_are_opt_in": true
 },
 "metric": {
  "prompt_processing": "prefill tokens / prefill seconds, harness CSV",
  "generation": "generated tokens / generation seconds, excludes prefill and startup",
  "steady": "excludes the first decode step as well",
  "first_decode_step_ms": "duration of the first decode step"
 },
 "gates": [
  "output SHA-256 equal to the published reference for the prompt length",
  "requested expert cache honoured, no reduced allocation",
  "zero swap growth during the arm",
  "GPU active clock at least 1600 MHz during decode, from an independent powermetrics collector",
  "fresh process and empty expert cache per arm; no build and no second engine process during any arm"
 ],
 "reference_output_sha256": {
  "512_generated": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4",
  "200_generated": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
 },
 "groups": [
  {
   "id": "stack512",
   "role": "qualification",
   "order": "BAAB",
   "generated_tokens": 512,
   "A": "published binary, router profile, cache 4200",
   "B": "stack: cache 4600 + flag readback + bf16 epilogues + pipeline cache",
   "medians_A": {
    "steady_tok_s": 20.3,
    "generation_tok_s": 20.04,
    "prefill_tok_s": 44.95,
    "first_decode_step_ms": 370.693
   },
   "medians_B": {
    "steady_tok_s": 22.74,
    "generation_tok_s": 22.325000000000003,
    "prefill_tok_s": 44.019999999999996,
    "first_decode_step_ms": 452.4935
   },
   "steady_gain_pct": 12.019704433497536,
   "inclusive_gain_pct": 11.402195608782462,
   "pair_deltas_steady": [
    2.39,
    2.49
   ],
   "complete": true
  },
  {
   "id": "stack200",
   "role": "qualification",
   "order": "ABBA",
   "generated_tokens": 200,
   "A": "published binary, router profile, cache 4200",
   "B": "stack: cache 4600 + flag readback + bf16 epilogues + pipeline cache",
   "medians_A": {
    "steady_tok_s": 19.965,
    "generation_tok_s": 19.314999999999998,
    "prefill_tok_s": 45.209999999999994,
    "first_decode_step_ms": 384.972
   },
   "medians_B": {
    "steady_tok_s": 22.490000000000002,
    "generation_tok_s": 21.5,
    "prefill_tok_s": 44.25,
    "first_decode_step_ms": 449.812
   },
   "steady_gain_pct": 12.647132481843236,
   "inclusive_gain_pct": 11.312451462593852,
   "pair_deltas_steady": [
    2.56,
    2.49
   ],
   "complete": true
  },
  {
   "id": "flag-readback",
   "role": "single lever",
   "order": "ABAB",
   "generated_tokens": 200,
   "A": "candidate binary, router profile, cache 4200",
   "B": "+ DS4_ARGODRIVE_FLAG_READBACK=1",
   "medians_A": {
    "steady_tok_s": 20.045,
    "generation_tok_s": 19.425,
    "prefill_tok_s": 45.18,
    "first_decode_step_ms": 366.188
   },
   "medians_B": {
    "steady_tok_s": 21.369999999999997,
    "generation_tok_s": 20.625,
    "prefill_tok_s": 45.195,
    "first_decode_step_ms": 383.2155
   },
   "steady_gain_pct": 6.61012721376899,
   "inclusive_gain_pct": 6.177606177606165,
   "pair_deltas_steady": [
    1.28,
    1.37
   ],
   "complete": true
  },
  {
   "id": "cache4600-512",
   "role": "single lever",
   "order": "BAAB",
   "generated_tokens": 512,
   "A": "published binary, cache 4200",
   "B": "published binary, cache 4600",
   "medians_A": {
    "steady_tok_s": 20.245,
    "generation_tok_s": 19.990000000000002,
    "prefill_tok_s": 45.35,
    "first_decode_step_ms": 365.31899999999996
   },
   "medians_B": {
    "steady_tok_s": 20.845,
    "generation_tok_s": 20.505,
    "prefill_tok_s": 44.29,
    "first_decode_step_ms": 451.6245
   },
   "steady_gain_pct": 2.9636947394418245,
   "inclusive_gain_pct": 2.5762881440720253,
   "pair_deltas_steady": [
    0.6,
    0.6
   ],
   "complete": true
  },
  {
   "id": "cache4600-200",
   "role": "single lever",
   "order": "ABAB",
   "generated_tokens": 200,
   "A": "candidate binary, cache 4200",
   "B": "candidate binary, cache 4600",
   "medians_A": {
    "steady_tok_s": 19.89,
    "generation_tok_s": 19.2,
    "prefill_tok_s": 45.145,
    "first_decode_step_ms": 407.312
   },
   "medians_B": {
    "steady_tok_s": 20.555,
    "generation_tok_s": 19.685000000000002,
    "prefill_tok_s": 44.165,
    "first_decode_step_ms": 472.494
   },
   "steady_gain_pct": 3.3433886375062727,
   "inclusive_gain_pct": 2.5260416666666785,
   "pair_deltas_steady": [
    0.72,
    0.61
   ],
   "complete": true
  },
  {
   "id": "bf16-epilogues",
   "role": "single lever",
   "order": "ABAB",
   "generated_tokens": 200,
   "A": "candidate binary, router profile",
   "B": "+ DS4_ARGODRIVE_BF16_EPILOGUES=1",
   "medians_A": {
    "steady_tok_s": 20.035,
    "generation_tok_s": 19.405,
    "prefill_tok_s": 45.07,
    "first_decode_step_ms": 370.1375
   },
   "medians_B": {
    "steady_tok_s": 20.125,
    "generation_tok_s": 19.384999999999998,
    "prefill_tok_s": 45.225,
    "first_decode_step_ms": 425.04200000000003
   },
   "steady_gain_pct": 0.4492138757175024,
   "inclusive_gain_pct": -0.10306622004639943,
   "pair_deltas_steady": [
    0.09,
    0.09
   ],
   "complete": true
  },
  {
   "id": "pipeline-cache",
   "role": "single lever",
   "order": "ABAB",
   "generated_tokens": 200,
   "A": "candidate binary, router profile",
   "B": "+ DS4_ARGODRIVE_PSO_CACHE=1",
   "medians_A": {
    "steady_tok_s": 19.955,
    "generation_tok_s": 19.335,
    "prefill_tok_s": 45.33,
    "first_decode_step_ms": 368.6295
   },
   "medians_B": {
    "steady_tok_s": 20.119999999999997,
    "generation_tok_s": 19.5,
    "prefill_tok_s": 45.33,
    "first_decode_step_ms": 362.709
   },
   "steady_gain_pct": 0.8268604359809473,
   "inclusive_gain_pct": 0.853374709076804,
   "pair_deltas_steady": [
    0.16,
    0.17
   ],
   "complete": true
  },
  {
   "id": "env-cache",
   "role": "single lever",
   "order": "ABAB",
   "generated_tokens": 200,
   "A": "candidate binary, env cache off",
   "B": "candidate binary, env cache on (default)",
   "medians_A": {
    "steady_tok_s": 19.915,
    "generation_tok_s": 19.265,
    "prefill_tok_s": 45.150000000000006,
    "first_decode_step_ms": 387.122
   },
   "medians_B": {
    "steady_tok_s": 20.005000000000003,
    "generation_tok_s": 19.355,
    "prefill_tok_s": 45.144999999999996,
    "first_decode_step_ms": 382.8675
   },
   "steady_gain_pct": 0.45192066281698473,
   "inclusive_gain_pct": 0.46716844017649084,
   "pair_deltas_steady": [
    0.03,
    0.15
   ],
   "complete": true
  },
  {
   "id": "stack512-excluded",
   "role": "excluded: clock gate",
   "order": "BAAB",
   "generated_tokens": 512,
   "A": "published binary",
   "B": "stack",
   "medians_A": {
    "steady_tok_s": 14.0,
    "generation_tok_s": 13.86,
    "prefill_tok_s": 44.68,
    "first_decode_step_ms": 408.56
   },
   "medians_B": {
    "steady_tok_s": 22.77,
    "generation_tok_s": 22.32,
    "prefill_tok_s": 44.68,
    "first_decode_step_ms": 485.88
   },
   "steady_gain_pct": 62.64285714285713,
   "inclusive_gain_pct": 61.03896103896105,
   "pair_deltas_steady": [
    8.77
   ],
   "complete": true
  },
  {
   "id": "stack200-excluded",
   "role": "excluded: clock gate",
   "order": "ABBA",
   "generated_tokens": 200,
   "A": "published binary",
   "B": "stack",
   "medians_A": {
    "steady_tok_s": 15.14,
    "generation_tok_s": 14.8,
    "prefill_tok_s": 45.14,
    "first_decode_step_ms": 364.404
   },
   "medians_B": null,
   "steady_gain_pct": null,
   "inclusive_gain_pct": null,
   "pair_deltas_steady": null,
   "complete": false
  }
 ],
 "arms": [
  {
   "arm": "stack512-1B",
   "group": "stack512",
   "label": "B",
   "side": "stack: flag readback + 4,600 experts + BF16 epilogues + pipeline cache",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 43.9,
   "generation_tok_s": 22.23,
   "steady_tok_s": 22.65,
   "first_decode_step_ms": 461.1,
   "steady_misses_per_token": 11.06,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 35.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.8,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
    "DS4_ARGODRIVE_PSO_CACHE": "1",
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack512-2A",
   "group": "stack512",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.84,
   "generation_tok_s": 20.0,
   "steady_tok_s": 20.26,
   "first_decode_step_ms": 371.3,
   "steady_misses_per_token": 13.09,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.5,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.72,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack512-3A",
   "group": "stack512",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 45.06,
   "generation_tok_s": 20.08,
   "steady_tok_s": 20.34,
   "first_decode_step_ms": 370.1,
   "steady_misses_per_token": 13.09,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.65,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack512-4B",
   "group": "stack512",
   "label": "B",
   "side": "stack: flag readback + 4,600 experts + BF16 epilogues + pipeline cache",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.14,
   "generation_tok_s": 22.42,
   "steady_tok_s": 22.83,
   "first_decode_step_ms": 443.9,
   "steady_misses_per_token": 11.06,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 35.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.75,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
    "DS4_ARGODRIVE_PSO_CACHE": "1",
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack200-1A",
   "group": "stack200",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.26,
   "generation_tok_s": 19.32,
   "steady_tok_s": 19.94,
   "first_decode_step_ms": 369.4,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.3,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.61,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "stack200-2B",
   "group": "stack200",
   "label": "B",
   "side": "stack: flag readback + 4,600 experts + BF16 epilogues + pipeline cache",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.43,
   "generation_tok_s": 21.52,
   "steady_tok_s": 22.5,
   "first_decode_step_ms": 444.5,
   "steady_misses_per_token": 12.22,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 33.8,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.67,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
    "DS4_ARGODRIVE_PSO_CACHE": "1",
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "stack200-3B",
   "group": "stack200",
   "label": "B",
   "side": "stack: flag readback + 4,600 experts + BF16 epilogues + pipeline cache",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.07,
   "generation_tok_s": 21.48,
   "steady_tok_s": 22.48,
   "first_decode_step_ms": 455.1,
   "steady_misses_per_token": 12.22,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 33.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.79,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
    "DS4_ARGODRIVE_PSO_CACHE": "1",
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "stack200-4A",
   "group": "stack200",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.16,
   "generation_tok_s": 19.31,
   "steady_tok_s": 19.99,
   "first_decode_step_ms": 400.5,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 30.8,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.62,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "qualification, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "flag-readback-1A",
   "group": "flag-readback",
   "label": "A",
   "side": "candidate binary, flag readback off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.33,
   "generation_tok_s": 19.48,
   "steady_tok_s": 20.1,
   "first_decode_step_ms": 365.7,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 30.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 13.22,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "flag-readback-2B",
   "group": "flag-readback",
   "label": "B",
   "side": "candidate binary, flag readback on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.33,
   "generation_tok_s": 20.6,
   "steady_tok_s": 21.38,
   "first_decode_step_ms": 397.9,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.58,
   "environment_delta": {
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "flag-readback-3A",
   "group": "flag-readback",
   "label": "A",
   "side": "candidate binary, flag readback off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.03,
   "generation_tok_s": 19.37,
   "steady_tok_s": 19.99,
   "first_decode_step_ms": 366.7,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.3,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.66,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "flag-readback-4B",
   "group": "flag-readback",
   "label": "B",
   "side": "candidate binary, flag readback on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.06,
   "generation_tok_s": 20.65,
   "steady_tok_s": 21.36,
   "first_decode_step_ms": 368.5,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 33.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.67,
   "environment_delta": {
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "cache4600-512-1B",
   "group": "cache4600-512",
   "label": "B",
   "side": "published binary, 4,600 experts on",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.42,
   "generation_tok_s": 20.52,
   "steady_tok_s": 20.86,
   "first_decode_step_ms": 449.0,
   "steady_misses_per_token": 11.06,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 33.1,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.69,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "cache4600-512-2A",
   "group": "cache4600-512",
   "label": "A",
   "side": "published binary, 4,600 experts off",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 45.49,
   "generation_tok_s": 20.0,
   "steady_tok_s": 20.26,
   "first_decode_step_ms": 367.4,
   "steady_misses_per_token": 13.09,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.0,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.57,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "cache4600-512-3A",
   "group": "cache4600-512",
   "label": "A",
   "side": "published binary, 4,600 experts off",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 45.21,
   "generation_tok_s": 19.98,
   "steady_tok_s": 20.23,
   "first_decode_step_ms": 363.3,
   "steady_misses_per_token": 13.09,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.8,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.62,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "cache4600-512-4B",
   "group": "cache4600-512",
   "label": "B",
   "side": "published binary, 4,600 experts on",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.16,
   "generation_tok_s": 20.49,
   "steady_tok_s": 20.83,
   "first_decode_step_ms": 454.2,
   "steady_misses_per_token": 11.06,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.5,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.74,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "cache4600-200-1A",
   "group": "cache4600-200",
   "label": "A",
   "side": "candidate binary, 4,600 experts off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.34,
   "generation_tok_s": 19.16,
   "steady_tok_s": 19.85,
   "first_decode_step_ms": 405.4,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.59,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "cache4600-200-2B",
   "group": "cache4600-200",
   "label": "B",
   "side": "candidate binary, 4,600 experts on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.22,
   "generation_tok_s": 19.66,
   "steady_tok_s": 20.57,
   "first_decode_step_ms": 492.8,
   "steady_misses_per_token": 12.22,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.8,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.73,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "cache4600-200-3A",
   "group": "cache4600-200",
   "label": "A",
   "side": "candidate binary, 4,600 experts off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.95,
   "generation_tok_s": 19.24,
   "steady_tok_s": 19.93,
   "first_decode_step_ms": 409.3,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 30.9,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.69,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "cache4600-200-4B",
   "group": "cache4600-200",
   "label": "B",
   "side": "candidate binary, 4,600 experts on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.11,
   "generation_tok_s": 19.71,
   "steady_tok_s": 20.54,
   "first_decode_step_ms": 452.2,
   "steady_misses_per_token": 12.22,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.6,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.75,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "bf16-epilogues-1A",
   "group": "bf16-epilogues",
   "label": "A",
   "side": "candidate binary, BF16 epilogues off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.97,
   "generation_tok_s": 19.39,
   "steady_tok_s": 20.03,
   "first_decode_step_ms": 374.7,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 13.53,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "bf16-epilogues-2B",
   "group": "bf16-epilogues",
   "label": "B",
   "side": "candidate binary, BF16 epilogues on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.39,
   "generation_tok_s": 19.33,
   "steady_tok_s": 20.12,
   "first_decode_step_ms": 451.2,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.6,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.58,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "bf16-epilogues-3A",
   "group": "bf16-epilogues",
   "label": "A",
   "side": "candidate binary, BF16 epilogues off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.17,
   "generation_tok_s": 19.42,
   "steady_tok_s": 20.04,
   "first_decode_step_ms": 365.5,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.4,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.63,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "bf16-epilogues-4B",
   "group": "bf16-epilogues",
   "label": "B",
   "side": "candidate binary, BF16 epilogues on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.06,
   "generation_tok_s": 19.44,
   "steady_tok_s": 20.13,
   "first_decode_step_ms": 398.8,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.5,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.65,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "pipeline-cache-1A",
   "group": "pipeline-cache",
   "label": "A",
   "side": "candidate binary, pipeline cache off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.34,
   "generation_tok_s": 19.34,
   "steady_tok_s": 19.95,
   "first_decode_step_ms": 365.6,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.3,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.59,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "pipeline-cache-2B",
   "group": "pipeline-cache",
   "label": "B",
   "side": "candidate binary, pipeline cache on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.74,
   "generation_tok_s": 19.49,
   "steady_tok_s": 20.11,
   "first_decode_step_ms": 362.4,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.1,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.49,
   "environment_delta": {
    "DS4_ARGODRIVE_PSO_CACHE": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "pipeline-cache-3A",
   "group": "pipeline-cache",
   "label": "A",
   "side": "candidate binary, pipeline cache off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.32,
   "generation_tok_s": 19.33,
   "steady_tok_s": 19.96,
   "first_decode_step_ms": 371.7,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 30.8,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.59,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "pipeline-cache-4B",
   "group": "pipeline-cache",
   "label": "B",
   "side": "candidate binary, pipeline cache on",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 44.92,
   "generation_tok_s": 19.51,
   "steady_tok_s": 20.13,
   "first_decode_step_ms": 363.0,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.1,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.69,
   "environment_delta": {
    "DS4_ARGODRIVE_PSO_CACHE": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "env-cache-1A",
   "group": "env-cache",
   "label": "A",
   "side": "environment cache off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.06,
   "generation_tok_s": 19.22,
   "steady_tok_s": 19.91,
   "first_decode_step_ms": 407.8,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 32.2,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.66,
   "environment_delta": {
    "DS4_ARGODRIVE_ENV_CACHE": "0"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "env-cache-2B",
   "group": "env-cache",
   "label": "B",
   "side": "environment cache on (code default)",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.23,
   "generation_tok_s": 19.26,
   "steady_tok_s": 19.94,
   "first_decode_step_ms": 400.8,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.62,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "env-cache-3A",
   "group": "env-cache",
   "label": "A",
   "side": "environment cache off",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.24,
   "generation_tok_s": 19.31,
   "steady_tok_s": 19.92,
   "first_decode_step_ms": 366.4,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.6,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.62,
   "environment_delta": {
    "DS4_ARGODRIVE_ENV_CACHE": "0"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "env-cache-4B",
   "group": "env-cache",
   "label": "B",
   "side": "environment cache on (code default)",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.06,
   "generation_tok_s": 19.45,
   "steady_tok_s": 20.07,
   "first_decode_step_ms": 365.0,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 1620,
   "gpu_w": 31.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.66,
   "environment_delta": {},
   "binary": "candidate",
   "gate": "passed",
   "disposition": "single-lever screen, passed gates",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  },
  {
   "arm": "stack512-excluded-1B",
   "group": "stack512-excluded",
   "label": "B",
   "side": "stack: flag readback + 4,600 experts + BF16 epilogues + pipeline cache",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.68,
   "generation_tok_s": 22.32,
   "steady_tok_s": 22.77,
   "first_decode_step_ms": 485.9,
   "steady_misses_per_token": 11.06,
   "requested_cache_experts": 4600,
   "gpu_active_mhz": 1620,
   "gpu_w": 36.9,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.61,
   "environment_delta": {
    "DS4_ARGODRIVE_BF16_EPILOGUES": "1",
    "DS4_ARGODRIVE_PSO_CACHE": "1",
    "DS4_ARGODRIVE_FLAG_READBACK": "1"
   },
   "binary": "candidate",
   "gate": "passed",
   "disposition": "excluded with its group: clock gate",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack512-excluded-2A",
   "group": "stack512-excluded",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 512,
   "prompt_processing_tok_s": 44.68,
   "generation_tok_s": 13.86,
   "steady_tok_s": 14.0,
   "first_decode_step_ms": 408.6,
   "steady_misses_per_token": 13.09,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 762,
   "gpu_w": 7.7,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.76,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "failed",
   "disposition": "excluded: clock gate",
   "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
  },
  {
   "arm": "stack200-excluded-1A",
   "group": "stack200-excluded",
   "label": "A",
   "side": "published 21 Sep binary, router profile, 4,200 experts",
   "generated_tokens": 200,
   "prompt_processing_tok_s": 45.14,
   "generation_tok_s": 14.8,
   "steady_tok_s": 15.14,
   "first_decode_step_ms": 364.4,
   "steady_misses_per_token": 14.32,
   "requested_cache_experts": 4200,
   "gpu_active_mhz": 907,
   "gpu_w": 12.2,
   "swap_growth_mb": 0.0,
   "engine_launch_to_first_token_s": 11.64,
   "environment_delta": {},
   "binary": "published tag",
   "gate": "failed",
   "disposition": "excluded: clock gate",
   "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
  }
 ],
 "costs": {
  "prefill_512_512": "44.02 vs 44.95 tok/s (-2.1%)",
  "first_decode_step_512_512": "452 vs 371 ms",
  "gpu_power_w_512_512": "35.5 vs 32.4 W median during decode"
 },
 "did_not_work": {
  "cache_4750_4900": "exceeded the swap-growth guard, with and without deferring growth until the prefill staging pool is released",
  "cpu_keepalive_off_on_stack": "-0.09%, within noise; setting unchanged",
  "screened_and_rejected": [
   "memory-touching keep-alive kernels",
   "read-pool spinning",
   "completion polling",
   "Q8 row tiling and SIMD-group variants",
   "eighteen reader threads",
   "pipeline pre-warming",
   "expert pre-warm, staged release, static views and slab warm for the first decode step"
  ]
 },
 "limits": [
  "one machine, one model, one prompt at two generated lengths",
  "comparison against our own published 21 September binary in the same session, not against upstream ds4",
  "two runs per configuration give an observed range, not a confidence interval",
  "the 4,600-expert cache needs the headroom of a 128 GB machine; larger caches tripped the swap guard",
  "keepalive remains enabled and raises power draw",
  "output identity is regression evidence, not model-quality validation"
 ],
 "note": "Transcribed from argodrive/candidates/2026-09-27-stack/results.json in the engine repository, which carries the per-arm CSV hashes, phase accounting and source checksums. Machine paths, serials and prompts are excluded.",
 "ladder_vs_upstream": {
  "date": "2026-09-27",
  "prompt_tokens": 512,
  "generated_tokens": 200,
  "sequence": [
   "U",
   "S1",
   "U",
   "S2",
   "U",
   "S3",
   "S3",
   "U",
   "S2",
   "U",
   "S1",
   "U"
  ],
  "upstream": {
   "commit": "bd66c402070042bf0a79ad6ece8242de4c93680c",
   "patch": "argodrive/upstream-phase.patch (phase markers only)",
   "binary_sha256": "f7dcecfeebb3b044dbede4afd8eabdfae920c4a3338f4a5899bb94a829b5824a",
   "toolchain": "Apple clang 14.0.3 / macOS 26.4 SDK, same as the fork build",
   "cache": "automatic (3688 experts)"
  },
  "fork_profile": "v41-stack-20260927",
  "output_identical_on_all_arms": true,
  "rungs": [
   {
    "id": "upstream-internal",
    "label": "upstream ds4 bd66c40, internal SSD only",
    "binary": "pinned upstream ds4 (phase markers only)",
    "replicas": 0,
    "cache": "automatic (3688 experts)",
    "runs": 6,
    "medians": {
     "steady_tok_s": 10.175,
     "generation_tok_s": 9.9,
     "prefill_tok_s": 17.965,
     "first_decode_step_ms": 646.129
    },
    "range_steady": [
     10.06,
     10.24
    ],
    "vs_upstream": {
     "steady_tok_s": 1.0,
     "generation_tok_s": 1.0,
     "prefill_tok_s": 1.0
    }
   },
   {
    "id": "fork-internal",
    "label": "this fork, internal SSD only (no replicas)",
    "binary": "fork, profile v41-stack-20260927",
    "replicas": 0,
    "cache": 4600,
    "runs": 2,
    "medians": {
     "steady_tok_s": 18.665,
     "generation_tok_s": 17.45,
     "prefill_tok_s": 30.985,
     "first_decode_step_ms": 799.0115
    },
    "range_steady": [
     18.59,
     18.74
    ],
    "vs_upstream": {
     "steady_tok_s": 1.8343980343980342,
     "generation_tok_s": 1.7626262626262625,
     "prefill_tok_s": 1.7247425549679933
    }
   },
   {
    "id": "fork-plus-one",
    "label": "this fork, internal + one enclosure",
    "binary": "fork, profile v41-stack-20260927; prefill split 10:5, decode 10:6",
    "replicas": 1,
    "cache": 4600,
    "runs": 2,
    "medians": {
     "steady_tok_s": 21.055,
     "generation_tok_s": 19.54,
     "prefill_tok_s": 36.879999999999995,
     "first_decode_step_ms": 780.3975
    },
    "range_steady": [
     21.04,
     21.07
    ],
    "vs_upstream": {
     "steady_tok_s": 2.069287469287469,
     "generation_tok_s": 1.9737373737373736,
     "prefill_tok_s": 2.0528806011689396
    }
   },
   {
    "id": "fork-plus-two",
    "label": "this fork, internal + two enclosures",
    "binary": "fork, profile v41-stack-20260927; prefill split 10:5:5, decode 10:6:6",
    "replicas": 2,
    "cache": 4600,
    "runs": 2,
    "medians": {
     "steady_tok_s": 22.25,
     "generation_tok_s": 20.744999999999997,
     "prefill_tok_s": 42.415,
     "first_decode_step_ms": 694.418
    },
    "range_steady": [
     22.15,
     22.35
    ],
    "vs_upstream": {
     "steady_tok_s": 2.1867321867321867,
     "generation_tok_s": 2.0954545454545452,
     "prefill_tok_s": 2.360979682716393
    }
   }
  ],
  "arms": [
   {
    "id": "ladder-0927-1U",
    "rung": "upstream-internal",
    "order": 1,
    "prefill_tok_s": 17.76,
    "generation_tok_s": 9.92,
    "steady_tok_s": 10.2,
    "first_decode_step_ms": 645.408,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1523.0915770602214,
    "gpu_w": 16.549108928831753,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-2S1",
    "rung": "fork-internal",
    "order": 2,
    "prefill_tok_s": 31.1,
    "generation_tok_s": 17.36,
    "steady_tok_s": 18.59,
    "first_decode_step_ms": 812.497,
    "requested_cache_experts": 4600,
    "replicas": 0,
    "gpu_active_mhz": 1620.000000000001,
    "gpu_w": 29.312064846151298,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 16.66016300022602,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-3U",
    "rung": "upstream-internal",
    "order": 3,
    "prefill_tok_s": 17.91,
    "generation_tok_s": 9.77,
    "steady_tok_s": 10.06,
    "first_decode_step_ms": 675.128,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1491.3381543004398,
    "gpu_w": 16.440340998757748,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-4S2",
    "rung": "fork-plus-one",
    "order": 4,
    "prefill_tok_s": 36.89,
    "generation_tok_s": 19.52,
    "steady_tok_s": 21.04,
    "first_decode_step_ms": 783.695,
    "requested_cache_experts": 4600,
    "replicas": 1,
    "gpu_active_mhz": 1620.0000000000018,
    "gpu_w": 32.48599367571423,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 14.074193999869749,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-5U",
    "rung": "upstream-internal",
    "order": 5,
    "prefill_tok_s": 18.0,
    "generation_tok_s": 9.94,
    "steady_tok_s": 10.22,
    "first_decode_step_ms": 646.85,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1521.7330678685687,
    "gpu_w": 17.383352651302395,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-6S3",
    "rung": "fork-plus-two",
    "order": 6,
    "prefill_tok_s": 41.23,
    "generation_tok_s": 20.47,
    "steady_tok_s": 22.15,
    "first_decode_step_ms": 781.765,
    "requested_cache_experts": 4600,
    "replicas": 2,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 34.059007368351566,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 12.595323000103235,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-7S3",
    "rung": "fork-plus-two",
    "order": 7,
    "prefill_tok_s": 43.6,
    "generation_tok_s": 21.02,
    "steady_tok_s": 22.35,
    "first_decode_step_ms": 607.071,
    "requested_cache_experts": 4600,
    "replicas": 2,
    "gpu_active_mhz": 1619.9999999999998,
    "gpu_w": 34.56429371773004,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 11.897517999866977,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-8U",
    "rung": "upstream-internal",
    "order": 8,
    "prefill_tok_s": 18.09,
    "generation_tok_s": 9.82,
    "steady_tok_s": 10.11,
    "first_decode_step_ms": 660.796,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1524.2224120273672,
    "gpu_w": 17.38363887442531,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-9S2",
    "rung": "fork-plus-one",
    "order": 9,
    "prefill_tok_s": 36.87,
    "generation_tok_s": 19.56,
    "steady_tok_s": 21.07,
    "first_decode_step_ms": 777.1,
    "requested_cache_experts": 4600,
    "replicas": 1,
    "gpu_active_mhz": 1620.0000000000005,
    "gpu_w": 33.016413961625,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 14.087217000080273,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-10U",
    "rung": "upstream-internal",
    "order": 10,
    "prefill_tok_s": 17.93,
    "generation_tok_s": 9.88,
    "steady_tok_s": 10.15,
    "first_decode_step_ms": 634.798,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1526.9959356387285,
    "gpu_w": 17.67910107249445,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-11S1",
    "rung": "fork-internal",
    "order": 11,
    "prefill_tok_s": 30.87,
    "generation_tok_s": 17.54,
    "steady_tok_s": 18.74,
    "first_decode_step_ms": 785.526,
    "requested_cache_experts": 4600,
    "replicas": 0,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 30.11901772551422,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": 16.780602999962866,
    "steady_misses_per_token": 12.21608040201005,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "ladder-0927-12U",
    "rung": "upstream-internal",
    "order": 12,
    "prefill_tok_s": 18.04,
    "generation_tok_s": 9.96,
    "steady_tok_s": 10.24,
    "first_decode_step_ms": 639.952,
    "requested_cache_experts": null,
    "replicas": 0,
    "gpu_active_mhz": 1529.8816227520285,
    "gpu_w": 17.77575935411936,
    "swap_growth_mb": 0.0,
    "engine_launch_to_first_token_seconds": null,
    "steady_misses_per_token": null,
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   }
  ],
  "scope": "Steady excludes the first decode step; inclusive includes it; both exclude startup and prefill. Upstream median of six arms, fork rungs median of two: observed ranges, not confidence intervals. Output identity is regression evidence, not broad quality validation. One machine, one model, one prompt."
 },
 "followup_2026_09_28": {
  "date": "2026-09-28",
  "profile": "v41-stack-20260928",
  "adds": {
   "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
   "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
  },
  "control": "published binary built from tag v41-stack-20260927 (engine source 9ad4a61) with profile v41-stack-20260927, cache 4600",
  "output_identical_on_all_arms": true,
  "groups": [
   {
    "id": "final512",
    "role": "qualification",
    "sequence": "BAAB",
    "generated_tokens": 512,
    "A": "published v41-stack-20260927 binary (tag build), stack profile, cache 4600",
    "B": "candidate: stack profile + POST_MOE_FLUSH=1 + VICTIM_PRESCAN=1 (final binary)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.92,
      "generation_tok_s": 22.47,
      "prefill_tok_s": 43.93,
      "first_decode_step_ms": 480.93899999999996
     },
     "B": {
      "steady_tok_s": 23.259999999999998,
      "generation_tok_s": 22.755,
      "prefill_tok_s": 44.41,
      "first_decode_step_ms": 524.0455
     }
    },
    "steady_gain_pct": 1.483420593368212,
    "inclusive_gain_pct": 1.2683578104138782,
    "pair_deltas_steady": [
     0.3,
     0.38
    ]
   },
   {
    "id": "final200",
    "role": "qualification",
    "sequence": "ABBA",
    "generated_tokens": 200,
    "A": "published v41-stack-20260927 binary (tag build), stack profile, cache 4600",
    "B": "candidate: stack profile + POST_MOE_FLUSH=1 + VICTIM_PRESCAN=1 (final binary)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.555,
      "generation_tok_s": 21.485,
      "prefill_tok_s": 43.94,
      "first_decode_step_ms": 481.726
     },
     "B": {
      "steady_tok_s": 22.94,
      "generation_tok_s": 21.810000000000002,
      "prefill_tok_s": 44.129999999999995,
      "first_decode_step_ms": 493.4425
     }
    },
    "steady_gain_pct": 1.7069385945466697,
    "inclusive_gain_pct": 1.5126832673958646,
    "pair_deltas_steady": [
     0.33,
     0.44
    ]
   },
   {
    "id": "flush512",
    "role": "qualification (earlier candidate build, flush only)",
    "sequence": "BAAB",
    "generated_tokens": 512,
    "A": "published v41-stack-20260927 binary (tag build), stack profile, cache 4600",
    "B": "candidate build 2: stack profile + DS4_ARGODRIVE_POST_MOE_FLUSH=1",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.83,
      "generation_tok_s": 22.365000000000002,
      "prefill_tok_s": 44.175,
      "first_decode_step_ms": 502.56949999999995
     },
     "B": {
      "steady_tok_s": 23.12,
      "generation_tok_s": 22.64,
      "prefill_tok_s": 44.34,
      "first_decode_step_ms": 503.991
     }
    },
    "steady_gain_pct": 1.270258431887883,
    "inclusive_gain_pct": 1.2295998211491188,
    "pair_deltas_steady": [
     0.37,
     0.21
    ]
   },
   {
    "id": "flush200",
    "role": "qualification (earlier candidate build, flush only)",
    "sequence": "ABBA",
    "generated_tokens": 200,
    "A": "published v41-stack-20260927 binary (tag build), stack profile, cache 4600",
    "B": "candidate build 2: stack profile + DS4_ARGODRIVE_POST_MOE_FLUSH=1",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.505000000000003,
      "generation_tok_s": 21.405,
      "prefill_tok_s": 44.230000000000004,
      "first_decode_step_ms": 496.028
     },
     "B": {
      "steady_tok_s": 22.634999999999998,
      "generation_tok_s": 21.5,
      "prefill_tok_s": 44.065,
      "first_decode_step_ms": 506.9825
     }
    },
    "steady_gain_pct": 0.5776494112419206,
    "inclusive_gain_pct": 0.44382153702404725,
    "pair_deltas_steady": [
     0.18,
     0.08
    ]
   },
   {
    "id": "post-moe-flush-miss-layers",
    "role": "single lever",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile",
    "B": "+ DS4_ARGODRIVE_POST_MOE_FLUSH=1 (commit after the routed MoE on layers that waited for reads)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.41,
      "generation_tok_s": 21.085,
      "prefill_tok_s": 43.335,
      "first_decode_step_ms": 602.6179999999999
     },
     "B": {
      "steady_tok_s": 22.72,
      "generation_tok_s": 21.365000000000002,
      "prefill_tok_s": 43.5,
      "first_decode_step_ms": 599.4784999999999
     }
    },
    "steady_gain_pct": 1.3833110218652411,
    "inclusive_gain_pct": 1.3279582641688492,
    "pair_deltas_steady": [
     0.32,
     0.3
    ]
   },
   {
    "id": "post-moe-flush-all-layers",
    "role": "single lever",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile",
    "B": "+ DS4_ARGODRIVE_POST_MOE_FLUSH=2 (commit after every layer's routed MoE)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 21.955,
      "generation_tok_s": 20.630000000000003,
      "prefill_tok_s": 43.74,
      "first_decode_step_ms": 635.9395
     },
     "B": {
      "steady_tok_s": 22.7,
      "generation_tok_s": 21.59,
      "prefill_tok_s": 44.480000000000004,
      "first_decode_step_ms": 493.093
     }
    },
    "steady_gain_pct": 3.3933044864495665,
    "inclusive_gain_pct": 4.6534173533688605,
    "pair_deltas_steady": [
     1.26,
     0.23
    ]
   },
   {
    "id": "victim-prescan",
    "role": "single lever",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile + POST_MOE_FLUSH=1",
    "B": "+ DS4_ARGODRIVE_VICTIM_PRESCAN=1 (eviction candidates ranked during the mailbox wait)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.725,
      "generation_tok_s": 21.65,
      "prefill_tok_s": 43.965,
      "first_decode_step_ms": 477.476
     },
     "B": {
      "steady_tok_s": 22.835,
      "generation_tok_s": 21.735,
      "prefill_tok_s": 44.16,
      "first_decode_step_ms": 483.8455
     }
    },
    "steady_gain_pct": 0.48404840484048695,
    "inclusive_gain_pct": 0.39260969976906424,
    "pair_deltas_steady": [
     0.19,
     0.03
    ]
   },
   {
    "id": "lookahead-mailbox-k2",
    "role": "single lever (rejected)",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile",
    "B": "+ DS4_ARGODRIVE_DECODE_AHEAD=1 (next-layer prediction through the mailbox, two most confident experts)",
    "complete": true,
    "medians": {
     "A": {
      "steady_tok_s": 22.28,
      "generation_tok_s": 20.73,
      "prefill_tok_s": 42.895,
      "first_decode_step_ms": 713.2825
     },
     "B": {
      "steady_tok_s": 22.060000000000002,
      "generation_tok_s": 20.69,
      "prefill_tok_s": 42.875,
      "first_decode_step_ms": 642.2570000000001
     }
    },
    "steady_gain_pct": -0.9874326750448748,
    "inclusive_gain_pct": -0.1929570670525793,
    "pair_deltas_steady": [
     -0.21,
     -0.23
    ]
   },
   {
    "id": "lookahead-missing-only-k3",
    "role": "single lever (rejected, stopped after one pair)",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile",
    "B": "+ DECODE_AHEAD=1, MISSING_ONLY=1, K=3 (speculative reads only for predicted non-resident experts)",
    "complete": false,
    "medians": {
     "A": {
      "steady_tok_s": 22.42,
      "generation_tok_s": 21.36,
      "prefill_tok_s": 44.05,
      "first_decode_step_ms": 483.196
     },
     "B": {
      "steady_tok_s": 20.59,
      "generation_tok_s": 19.71,
      "prefill_tok_s": 44.06,
      "first_decode_step_ms": 478.159
     }
    },
    "steady_gain_pct": -8.162355040142732,
    "inclusive_gain_pct": -7.72471910112359,
    "pair_deltas_steady": [
     -1.83
    ]
   },
   {
    "id": "lookahead-missing-only-k2",
    "role": "excluded: swap guard",
    "sequence": "ABAB",
    "generated_tokens": 200,
    "A": "candidate binary, stack profile",
    "B": "+ DECODE_AHEAD=1, MISSING_ONLY=1, K=2 (arm 2B stopped by the swap-growth guard)",
    "complete": false,
    "medians": {
     "A": {
      "steady_tok_s": 22.47,
      "generation_tok_s": 21.25,
      "prefill_tok_s": 43.58,
      "first_decode_step_ms": 552.214
     },
     "B": null
    },
    "steady_gain_pct": null,
    "inclusive_gain_pct": null,
    "pair_deltas_steady": null
   }
  ],
  "arms": [
   {
    "id": "final512-1B",
    "group": "final512",
    "label": "B",
    "generated_tokens": 512,
    "prefill_tok_s": 44.56,
    "generation_tok_s": 22.65,
    "steady_tok_s": 23.18,
    "first_decode_step_ms": 552.833,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999993,
    "gpu_w": 35.56370735963624,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "final512-2A",
    "group": "final512",
    "label": "A",
    "generated_tokens": 512,
    "prefill_tok_s": 43.93,
    "generation_tok_s": 22.44,
    "steady_tok_s": 22.88,
    "first_decode_step_ms": 474.255,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 35.41087118420163,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "final512-3A",
    "group": "final512",
    "label": "A",
    "generated_tokens": 512,
    "prefill_tok_s": 43.93,
    "generation_tok_s": 22.5,
    "steady_tok_s": 22.96,
    "first_decode_step_ms": 487.623,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000011,
    "gpu_w": 35.32181037270135,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "final512-4B",
    "group": "final512",
    "label": "B",
    "generated_tokens": 512,
    "prefill_tok_s": 44.26,
    "generation_tok_s": 22.86,
    "steady_tok_s": 23.34,
    "first_decode_step_ms": 495.258,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000018,
    "gpu_w": 35.67275581593567,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "final200-1A",
    "group": "final200",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.77,
    "generation_tok_s": 21.5,
    "steady_tok_s": 22.59,
    "first_decode_step_ms": 488.399,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 34.78824552946018,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "final200-2B",
    "group": "final200",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.18,
    "generation_tok_s": 21.8,
    "steady_tok_s": 22.92,
    "first_decode_step_ms": 491.036,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.15893776276427,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "final200-3B",
    "group": "final200",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.08,
    "generation_tok_s": 21.82,
    "steady_tok_s": 22.96,
    "first_decode_step_ms": 495.849,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.19304146116703,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "final200-4A",
    "group": "final200",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.11,
    "generation_tok_s": 21.47,
    "steady_tok_s": 22.52,
    "first_decode_step_ms": 475.053,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 33.67455563197465,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "flush512-1B",
    "group": "flush512",
    "label": "B",
    "generated_tokens": 512,
    "prefill_tok_s": 44.25,
    "generation_tok_s": 22.66,
    "steady_tok_s": 23.14,
    "first_decode_step_ms": 507.487,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 35.58335829245858,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "flush512-2A",
    "group": "flush512",
    "label": "A",
    "generated_tokens": 512,
    "prefill_tok_s": 44.09,
    "generation_tok_s": 22.28,
    "steady_tok_s": 22.77,
    "first_decode_step_ms": 528.554,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000011,
    "gpu_w": 35.36637007435779,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "flush512-3A",
    "group": "flush512",
    "label": "A",
    "generated_tokens": 512,
    "prefill_tok_s": 44.26,
    "generation_tok_s": 22.45,
    "steady_tok_s": 22.89,
    "first_decode_step_ms": 476.585,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 35.43045191207663,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "flush512-4B",
    "group": "flush512",
    "label": "B",
    "generated_tokens": 512,
    "prefill_tok_s": 44.43,
    "generation_tok_s": 22.62,
    "steady_tok_s": 23.1,
    "first_decode_step_ms": 500.495,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000014,
    "gpu_w": 35.566178071764675,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.060665362035225,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "f14f5a1dcb74375d8d3e98fba56cc1cdea2d410117fd41decb82d0d273fd1ec4"
   },
   {
    "id": "flush200-1A",
    "group": "flush200",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.15,
    "generation_tok_s": 21.43,
    "steady_tok_s": 22.53,
    "first_decode_step_ms": 493.336,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999998,
    "gpu_w": 34.865078270404275,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "flush200-2B",
    "group": "flush200",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.32,
    "generation_tok_s": 21.61,
    "steady_tok_s": 22.71,
    "first_decode_step_ms": 485.548,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 33.962447441905375,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "flush200-3B",
    "group": "flush200",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 43.81,
    "generation_tok_s": 21.39,
    "steady_tok_s": 22.56,
    "first_decode_step_ms": 528.417,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 33.76219694107462,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "flush200-4A",
    "group": "flush200",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.31,
    "generation_tok_s": 21.38,
    "steady_tok_s": 22.48,
    "first_decode_step_ms": 498.72,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 33.67172727118506,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "published tag",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-miss-layers-1A",
    "group": "post-moe-flush-miss-layers",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.03,
    "generation_tok_s": 20.91,
    "steady_tok_s": 22.35,
    "first_decode_step_ms": 658.406,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 33.799379325065686,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-miss-layers-2B",
    "group": "post-moe-flush-miss-layers",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 43.66,
    "generation_tok_s": 21.25,
    "steady_tok_s": 22.67,
    "first_decode_step_ms": 628.486,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 34.72615978661302,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-miss-layers-3A",
    "group": "post-moe-flush-miss-layers",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.64,
    "generation_tok_s": 21.26,
    "steady_tok_s": 22.47,
    "first_decode_step_ms": 546.83,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999993,
    "gpu_w": 34.72088762162371,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-miss-layers-4B",
    "group": "post-moe-flush-miss-layers",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 43.34,
    "generation_tok_s": 21.48,
    "steady_tok_s": 22.77,
    "first_decode_step_ms": 570.471,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999998,
    "gpu_w": 35.30115064671999,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-all-layers-1A",
    "group": "post-moe-flush-all-layers",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.03,
    "generation_tok_s": 19.82,
    "steady_tok_s": 21.42,
    "first_decode_step_ms": 796.361,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 34.97409567381799,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-all-layers-2B",
    "group": "post-moe-flush-all-layers",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.56,
    "generation_tok_s": 21.6,
    "steady_tok_s": 22.68,
    "first_decode_step_ms": 479.749,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.70197302808477,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "2"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-all-layers-3A",
    "group": "post-moe-flush-all-layers",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.45,
    "generation_tok_s": 21.44,
    "steady_tok_s": 22.49,
    "first_decode_step_ms": 475.518,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.26246519410018,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "post-moe-flush-all-layers-4B",
    "group": "post-moe-flush-all-layers",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.4,
    "generation_tok_s": 21.58,
    "steady_tok_s": 22.72,
    "first_decode_step_ms": 506.437,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999993,
    "gpu_w": 34.554637658361585,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "2"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "victim-prescan-1A",
    "group": "victim-prescan",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.92,
    "generation_tok_s": 21.61,
    "steady_tok_s": 22.67,
    "first_decode_step_ms": 471.679,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999998,
    "gpu_w": 34.11126140335785,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "victim-prescan-2B",
    "group": "victim-prescan",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.49,
    "generation_tok_s": 21.79,
    "steady_tok_s": 22.86,
    "first_decode_step_ms": 471.879,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 34.30506919604722,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "victim-prescan-3A",
    "group": "victim-prescan",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.01,
    "generation_tok_s": 21.69,
    "steady_tok_s": 22.78,
    "first_decode_step_ms": 483.273,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000005,
    "gpu_w": 34.18726697608164,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "victim-prescan-4B",
    "group": "victim-prescan",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 43.83,
    "generation_tok_s": 21.68,
    "steady_tok_s": 22.81,
    "first_decode_step_ms": 495.812,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.41937484395672,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {
     "DS4_ARGODRIVE_POST_MOE_FLUSH": "1",
     "DS4_ARGODRIVE_VICTIM_PRESCAN": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-mailbox-k2-1A",
    "group": "lookahead-mailbox-k2",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 42.77,
    "generation_tok_s": 20.54,
    "steady_tok_s": 22.23,
    "first_decode_step_ms": 780.955,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.000000000001,
    "gpu_w": 33.12296870005367,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-mailbox-k2-2B",
    "group": "lookahead-mailbox-k2",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 43.06,
    "generation_tok_s": 20.69,
    "steady_tok_s": 22.02,
    "first_decode_step_ms": 626.15,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0,
    "gpu_w": 34.009584264111325,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.371859296482413,
    "environment_delta": {
     "DS4_ARGODRIVE_DECODE_AHEAD": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-mailbox-k2-3A",
    "group": "lookahead-mailbox-k2",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.02,
    "generation_tok_s": 20.92,
    "steady_tok_s": 22.33,
    "first_decode_step_ms": 645.61,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.35291255296814,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-mailbox-k2-4B",
    "group": "lookahead-mailbox-k2",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 42.69,
    "generation_tok_s": 20.69,
    "steady_tok_s": 22.1,
    "first_decode_step_ms": 658.364,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1619.9999999999998,
    "gpu_w": 34.35877204769018,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 11.371859296482413,
    "environment_delta": {
     "DS4_ARGODRIVE_DECODE_AHEAD": "1"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-missing-only-k3-1A",
    "group": "lookahead-missing-only-k3",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 44.05,
    "generation_tok_s": 21.36,
    "steady_tok_s": 22.42,
    "first_decode_step_ms": 483.196,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 34.687590112850884,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-missing-only-k3-2B",
    "group": "lookahead-missing-only-k3",
    "label": "B",
    "generated_tokens": 200,
    "prefill_tok_s": 44.06,
    "generation_tok_s": 19.71,
    "steady_tok_s": 20.59,
    "first_decode_step_ms": 478.159,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000007,
    "gpu_w": 32.32701918352478,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 8.371859296482413,
    "environment_delta": {
     "DS4_ARGODRIVE_DECODE_AHEAD": "1",
     "DS4_ARGODRIVE_DECODE_AHEAD_MISSING_ONLY": "1",
     "DS4_ARGODRIVE_DECODE_AHEAD_K": "3"
    },
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   },
   {
    "id": "lookahead-missing-only-k2-1A",
    "group": "lookahead-missing-only-k2",
    "label": "A",
    "generated_tokens": 200,
    "prefill_tok_s": 43.58,
    "generation_tok_s": 21.25,
    "steady_tok_s": 22.47,
    "first_decode_step_ms": 552.214,
    "requested_cache_experts": 4600,
    "gpu_active_mhz": 1620.0000000000002,
    "gpu_w": 35.70494999349978,
    "swap_growth_mb": 0.0,
    "steady_misses_per_token": 12.21608040201005,
    "environment_delta": {},
    "binary_tree": "candidate",
    "output_sha256": "8182ab832dcc07c11d4d36028b9b3f4f5cc672538011495e60fcb4f40b91044f"
   }
  ],
  "scope": "Steady excludes the first decode step; inclusive includes it. Both exclude startup and prefill. Two repeats are an observed range, not a confidence interval. Output identity is regression evidence, not broad quality validation. Single-lever groups are screens on the candidate binary and are not the qualification. Excluded groups are retained in full. No comparison with current upstream is established.",
  "release": "v41-stack-20260928"
 },
 "prefill_20260930": {
  "date": "2026-09-30",
  "profile": "v41-stack-20260930",
  "qualification_cache_experts": 4200,
  "ladder_cache_experts": 3600,
  "qualification": {
   "pp512_tg512": {
    "sequence": "BAAB",
    "generated_tokens": 512,
    "prefill_tok_s": {
     "A": 45.555,
     "B": 68.58500000000001
    },
    "prefill_gain_pct": 50.55427505213479,
    "steady_gain_pct": -0.7117437722420017,
    "pair_deltas_prefill": [
     23.01,
     23.05
    ]
   },
   "pp512_tg200": {
    "sequence": "ABBA",
    "generated_tokens": 200,
    "prefill_tok_s": {
     "A": 45.5,
     "B": 68.08500000000001
    },
    "prefill_gain_pct": 49.63736263736265,
    "steady_gain_pct": -0.5986590038314143,
    "pair_deltas_prefill": [
     21.93,
     23.24
    ]
   }
  },
  "ladder": {
   "upstream-internal": {
    "runs": 6,
    "medians": {
     "steady_tok_s": 10.114999999999998,
     "generation_tok_s": 9.725000000000001,
     "prefill_tok_s": 16.119999999999997,
     "first_decode_step_ms": 893.7465
    },
    "vs_upstream": {
     "steady_tok_s": 1.0,
     "generation_tok_s": 1.0,
     "prefill_tok_s": 1.0
    }
   },
   "fork-internal": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 16.775,
     "generation_tok_s": 16.15,
     "prefill_tok_s": 43.75,
     "first_decode_step_ms": 519.8995
    },
    "vs_upstream": {
     "steady_tok_s": 1.6584280771131983,
     "generation_tok_s": 1.6606683804627245,
     "prefill_tok_s": 2.7140198511166256
    }
   },
   "fork-plus-one": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 18.674999999999997,
     "generation_tok_s": 17.895,
     "prefill_tok_s": 58.129999999999995,
     "first_decode_step_ms": 514.62
    },
    "vs_upstream": {
     "steady_tok_s": 1.8462679189322788,
     "generation_tok_s": 1.8401028277634959,
     "prefill_tok_s": 3.6060794044665014
    }
   },
   "fork-plus-two": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 20.020000000000003,
     "generation_tok_s": 19.22,
     "prefill_tok_s": 68.28999999999999,
     "first_decode_step_ms": 463.1965
    },
    "vs_upstream": {
     "steady_tok_s": 1.97923875432526,
     "generation_tok_s": 1.9763496143958865,
     "prefill_tok_s": 4.2363523573201
    }
   }
  },
  "ladder_sequence": [
   "U",
   "S1",
   "U",
   "S2",
   "U",
   "S3",
   "S3",
   "U",
   "S2",
   "U",
   "S1",
   "U"
  ],
  "ladder_max_swap_growth_mb": 0.0,
  "output_identical_on_all_arms": true,
  "package": "https://github.com/argonautlabsai/ds4-argodrive/blob/v41-stack-20260930/argodrive/candidates/2026-09-30-prefill/README.md",
  "ladder_package": "https://github.com/argonautlabsai/ds4-argodrive/blob/v41-stack-20260930/argodrive/candidates/2026-09-30-ladder/README.md",
  "release": "https://github.com/argonautlabsai/ds4-argodrive/releases/tag/v41-stack-20260930"
 },
 "prefill_20261004": {
  "date": "2026-10-04",
  "profile": "v41-stack-20261004",
  "cache_experts": 3600,
  "qualification": {
   "pp512_tg512": {
    "sequence": "BAAB",
    "generated_tokens": 512,
    "prefill_tok_s": {
     "A": 68.85,
     "B": 73.265
    },
    "prefill_gain_pct": 6.412490922294856,
    "steady_gain_pct": -0.024850894632189036,
    "pair_deltas_prefill": [
     4.3,
     4.53
    ]
   },
   "pp512_tg200": {
    "sequence": "ABBA",
    "generated_tokens": 200,
    "prefill_tok_s": {
     "A": 68.66999999999999,
     "B": 73.155
    },
    "prefill_gain_pct": 6.5312363477501245,
    "steady_gain_pct": -2.065272819989805,
    "pair_deltas_prefill": [
     3.93,
     5.04
    ]
   }
  },
  "ladder": {
   "upstream-internal": {
    "runs": 6,
    "medians": {
     "steady_tok_s": 10.17,
     "generation_tok_s": 9.805,
     "prefill_tok_s": 17.305,
     "first_decode_step_ms": 823.8235
    },
    "vs_upstream": {
     "steady_tok_s": 1.0,
     "generation_tok_s": 1.0,
     "prefill_tok_s": 1.0
    }
   },
   "fork-internal": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 16.935000000000002,
     "generation_tok_s": 16.285,
     "prefill_tok_s": 47.71,
     "first_decode_step_ms": 522.159
    },
    "vs_upstream": {
     "steady_tok_s": 1.6651917404129797,
     "generation_tok_s": 1.6608873023967363,
     "prefill_tok_s": 2.7570066454781856
    }
   },
   "fork-plus-one": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 18.64,
     "generation_tok_s": 17.825000000000003,
     "prefill_tok_s": 62.989999999999995,
     "first_decode_step_ms": 537.539
    },
    "vs_upstream": {
     "steady_tok_s": 1.832841691248771,
     "generation_tok_s": 1.8179500254971956,
     "prefill_tok_s": 3.639988442646634
    }
   },
   "fork-plus-two": {
    "runs": 2,
    "medians": {
     "steady_tok_s": 19.314999999999998,
     "generation_tok_s": 18.560000000000002,
     "prefill_tok_s": 75.065,
     "first_decode_step_ms": 471.746
    },
    "vs_upstream": {
     "steady_tok_s": 1.8992133726646998,
     "generation_tok_s": 1.8929117797042327,
     "prefill_tok_s": 4.337763652123663
    }
   }
  },
  "ladder_sequence": [
   "U",
   "S1",
   "U",
   "S2",
   "U",
   "S3",
   "S3",
   "U",
   "S2",
   "U",
   "S1",
   "U"
  ],
  "ladder_max_swap_growth_mb": 0.0,
  "output_identical_on_all_arms": true,
  "package": "https://github.com/argonautlabsai/ds4-argodrive/blob/v41-stack-20261004/argodrive/candidates/2026-10-04-engram/README.md",
  "ladder_package": "https://github.com/argonautlabsai/ds4-argodrive/blob/v41-stack-20261004/argodrive/candidates/2026-10-04-ladder/README.md",
  "release": "https://github.com/argonautlabsai/ds4-argodrive/releases/tag/v41-stack-20261004"
 }
}
