{
  "format": "unaveraged-default-model-benchmarks/v1",
  "updated": "2026-10-27",
  "aggregate_only": true,
  "environment": {
    "digit_accuracy": {
      "python": "3.10.12",
      "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
      "cpu": "AMD Ryzen 5 5600U with Radeon Graphics"
    },
    "digit_latency": {
      "native": {
        "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
        "logical_cpus": 4,
        "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
        "python": "3.10.12",
        "numpy": "2.2.6",
        "host_observations": [
          {
            "phase": "before",
            "timestamp": "2026-10-27T06:02:06.596454+00:00",
            "loadavg": [
              0.1982421875,
              0.09814453125,
              0.03369140625
            ]
          },
          {
            "phase": "after",
            "timestamp": "2026-10-27T06:02:27.853666+00:00",
            "loadavg": [
              0.4267578125,
              0.15771484375,
              0.05517578125
            ]
          }
        ]
      },
      "chrome": {
        "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
        "logical_cpus": 4,
        "platform": "Linux 6.8.0-138-generic x64",
        "node": "v24.19.0",
        "host_observations": [
          {
            "phase": "before",
            "timestamp": "2026-10-27T06:02:53.106Z",
            "loadavg": [
              0.28,
              0.14,
              0.05
            ]
          },
          {
            "phase": "after",
            "timestamp": "2026-10-27T06:04:07.170Z",
            "loadavg": [
              0.8,
              0.34,
              0.13
            ]
          }
        ],
        "browser": "Chrome/150.0.7871.24"
      }
    },
    "scope": "Digit accuracy uses Ryzen 5 5600U; Digit latency uses i3-1115G4; VQA latency uses Ryzen 5 5600U. Compare speed within each measured host and protocol.",
    "vqa_latency": {
      "cpu": "AMD Ryzen 5 5600U with Radeon Graphics",
      "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
      "python": "3.10.12",
      "chrome": "150.0.7871.24",
      "wasm_cpu_threads": 1,
      "batch_size": 1,
      "host_idle_claimed": false,
      "node": "v24.19.0",
      "firefox": "firefox/158.0.1"
    }
  },
  "runtime": {
    "native": {
      "package": "volvoxai",
      "version": "0.9.0",
      "source": "pypi",
      "sha256": {
        "libvolvoxai.so": "3dcd7236d71bff5236e60a3e2c8e665e95ad2ee59471641cdea37da72543695d",
        "libsynurang_module_host.so": "a41b36a8ec25847d0dbd23dc9cd8da72d1c1245b7780277897484c35264b3cfb"
      },
      "wheel": "volvoxai-0.9.0-cp310-abi3-manylinux_2_28_x86_64.whl",
      "wheel_sha256": "47b1e1a720d4938bafc152f0d87af907057afc7ae8ba9909606ead3b24a30a5d",
      "url": "https://files.pythonhosted.org/packages/4d/c6/e398c121a68478bb32944383c04bfec10494a25bafc701a6a0fd167553fa/volvoxai-0.9.0-cp310-abi3-manylinux_2_28_x86_64.whl"
    },
    "wasm": {
      "package": "volvoxai-lite",
      "version": "0.9.0",
      "source": "published-npm",
      "sha256": {
        "volvoxai.lite.min.js": "b228f0014189e85f8ec4070ad98ea64eae3c0e9905101ddfe55d4aa3632f5ac2",
        "volvoxai.lite.wasm": "b4b290aa55feb16c2f45a2158438dc342b6a8e42ce5abc9c3648d4b56b3c50f1"
      },
      "tarball_sha256": "4571e36a5be180341ac3d7a1739826c8c35514572a66ab9eaee442f52d574de0",
      "integrity": "sha512-zxQNkaYB69mZ6KvMA5iKmBD+6UgF6pzX21qIzoZ7zCyxsKqRVe/GiYs6gj+pO+v1zgQ27hHxsps4VHMLxyLB4g==",
      "url": "https://registry.npmjs.org/volvoxai-lite/-/volvoxai-lite-0.9.0.tgz"
    },
    "default_dependency": {
      "version": "0.9.0",
      "source": "published"
    }
  },
  "models": {
    "digit": {
      "name": "Digit conv4-linear2 QAT",
      "identity": {
        "profile": "conv4-linear2",
        "training": "joint-CUDA weight-only QAT, selected epoch 2; final CPU A8 calibration",
        "sha256": {
          "graph.json": "2ebaf4611059b396a7fddc3ddea4c8be5dc0993a6065ef65de593e7896ce1d24",
          "model.safetensors": "50d5e4e07abe5b87204f20fe5f3f7b5e9f40c15a10abcd23a185ea1890b9748c"
        },
        "graph_weight_bytes": 1726081,
        "graph_weight_mib": 1.6461191177368164,
        "packed_arithmetic": "dynamic-i8",
        "conv_weight_bits": 4,
        "linear_weight_bits": 2,
        "linear_group_size": 32,
        "activation_policy": "Static INT8 convolution activations; dynamic INT8 packed linear activations; other operators remain FP32.",
        "selected_epoch": 2
      },
      "accuracy": {
        "date": "2026-10-28",
        "method": "complete_phone_street_record",
        "correct": 1913,
        "samples": 2000,
        "percent": 95.65,
        "backend": "native-cpu",
        "cpu_threads": 1,
        "runtime_version": "0.9.0",
        "runtime_source": "pypi",
        "candidate_runtime_reevaluated": true,
        "independent": false,
        "scope": "All 2000 supplied heldout records, previously used for development/evaluation; annotations not fully human-audited.",
        "question_answer_correct": 1968,
        "question_answer_percent": 98.4,
        "chrome": {
          "record_correct": 1913,
          "phone_correct": 1964,
          "street_correct": 1937,
          "answer_correct": 1968,
          "samples": 2000,
          "record_percent": 95.65,
          "answer_percent": 98.4
        }
      },
      "comparators": [
        {
          "profile": "f32",
          "name": "FP32",
          "graph_weight_bytes": 9949321,
          "sha256": {
            "graph.json": "09268ab13dba6b615a04e9891b186497ea09acbe89a4f263aed3ebcd85231071",
            "model.safetensors": "1aebb58a94fe40abfb8ecde5ff2128254438e1c987b4ac7b60a3522cf6ed1b27"
          },
          "accuracy": {
            "date": "2026-10-27",
            "method": "complete_phone_street_record",
            "correct": 1938,
            "samples": 2000,
            "percent": 96.9,
            "backend": "native-cpu",
            "cpu_threads": 1,
            "runtime_version": "0.9.0",
            "runtime_source": "pypi",
            "candidate_runtime_reevaluated": true,
            "independent": false,
            "scope": "All 2000 supplied heldout records, previously used for development/evaluation; annotations not fully human-audited.",
            "question_answer_correct": 1975,
            "question_answer_percent": 98.75,
            "chrome": {
              "record_correct": 1938,
              "phone_correct": 1988,
              "street_correct": 1950,
              "answer_correct": 1975,
              "samples": 2000,
              "record_percent": 96.9,
              "answer_percent": 98.75
            }
          },
          "latency": {
            "date": "2026-10-27",
            "runtime_version": "0.9.0",
            "environment": {
              "native": {
                "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
                "logical_cpus": 4,
                "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
                "python": "3.10.12",
                "numpy": "2.2.6",
                "host_observations": [
                  {
                    "phase": "before",
                    "timestamp": "2026-10-27T06:02:06.596454+00:00",
                    "loadavg": [
                      0.1982421875,
                      0.09814453125,
                      0.03369140625
                    ]
                  },
                  {
                    "phase": "after",
                    "timestamp": "2026-10-27T06:02:27.853666+00:00",
                    "loadavg": [
                      0.4267578125,
                      0.15771484375,
                      0.05517578125
                    ]
                  }
                ]
              },
              "chrome": {
                "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
                "logical_cpus": 4,
                "platform": "Linux 6.8.0-138-generic x64",
                "node": "v24.19.0",
                "host_observations": [
                  {
                    "phase": "before",
                    "timestamp": "2026-10-27T06:02:53.106Z",
                    "loadavg": [
                      0.28,
                      0.14,
                      0.05
                    ]
                  },
                  {
                    "phase": "after",
                    "timestamp": "2026-10-27T06:04:07.170Z",
                    "loadavg": [
                      0.8,
                      0.34,
                      0.13
                    ]
                  }
                ],
                "browser": "Chrome/150.0.7871.24"
              }
            },
            "warm": {
              "native_cpu_1t": {
                "cpu_threads": 1,
                "samples": 45,
                "median_ms": 33.794236,
                "mean_ms": 33.84408542222222,
                "p95_ms": 34.1868754
              },
              "chrome_wasm": {
                "cpu_threads": 1,
                "execution_context": "Dedicated module worker",
                "samples": 45,
                "median_ms": 166.09999990463257,
                "mean_ms": 166.44444443384808,
                "p95_ms": 168.86000003814698
              }
            },
            "protocol": {
              "native": {
                "cycles": 15,
                "warmups_per_input": 4,
                "initial_reference_executions_per_input": 1,
                "input_count": 3,
                "samples_per_model": 45,
                "batch_size": 1,
                "cpu_threads": 1,
                "order": "Alternate forward/reverse model order each cycle, for each fixed input",
                "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
                "host_idle_claimed": false,
                "timer_includes": [
                  "InferenceSession.run, input/output transport and owned output arrays"
                ],
                "timer_excludes": [
                  "input reads",
                  "hashing",
                  "preprocessing",
                  "loading/compilation",
                  "warmups",
                  "numerical validation"
                ],
                "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
                "source_input_indices": [
                  0,
                  999,
                  1999
                ],
                "input_sha256": [
                  "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
                  "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
                  "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
                ],
                "status": "passed",
                "tolerance": {
                  "rtol": 0.0005,
                  "atol": 0.001
                }
              },
              "chrome": {
                "cycles": 15,
                "warmups_per_input": 4,
                "initial_reference_executions_per_input": 1,
                "input_count": 3,
                "samples_per_model": 45,
                "batch_size": 1,
                "cpu_threads": 1,
                "order": "Alternate forward/reverse model order each cycle, for each fixed input",
                "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
                "host_idle_claimed": false,
                "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
                "source_input_indices": [
                  0,
                  999,
                  1999
                ],
                "input_sha256": [
                  "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
                  "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
                  "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
                ],
                "status": "passed",
                "tolerance": {
                  "rtol": 0.0005,
                  "atol": 0.001
                },
                "browser": "Chrome/150.0.7871.24",
                "execution_context": "Chrome dedicated module worker",
                "timer_includes": [
                  "execute dispatch, input/output transport and copied inline output"
                ],
                "timer_excludes": [
                  "preprocessing",
                  "model verification",
                  "load",
                  "compile",
                  "context creation",
                  "numerical validation",
                  "UI",
                  "extension queue"
                ]
              }
            },
            "scope": "Interleaved fixed-input comparison on one Intel i3-1115G4 host; 3 heldout inputs, 15 cycles, 45 executions per model. Full 2000-record accuracy-pass timing is auxiliary in digit-comparison.json."
          }
        },
        {
          "profile": "w8a8",
          "name": "W8A8",
          "graph_weight_bytes": 4132316,
          "sha256": {
            "graph.json": "8e89b05bae252847ba66e4a654f3b0bec4f25f0c5d444e86dedefa752e026c2a",
            "model.safetensors": "4e3f0fc5777d536312573f885ece77865574aba751f82a574b14c14c50218b7b"
          },
          "accuracy": {
            "date": "2026-10-27",
            "method": "complete_phone_street_record",
            "correct": 1936,
            "samples": 2000,
            "percent": 96.8,
            "backend": "native-cpu",
            "cpu_threads": 1,
            "runtime_version": "0.9.0",
            "runtime_source": "pypi",
            "candidate_runtime_reevaluated": true,
            "independent": false,
            "scope": "All 2000 supplied heldout records, previously used for development/evaluation; annotations not fully human-audited.",
            "question_answer_correct": 1976,
            "question_answer_percent": 98.8,
            "chrome": {
              "record_correct": 1936,
              "phone_correct": 1986,
              "street_correct": 1950,
              "answer_correct": 1976,
              "samples": 2000,
              "record_percent": 96.8,
              "answer_percent": 98.8
            }
          },
          "latency": {
            "date": "2026-10-27",
            "runtime_version": "0.9.0",
            "environment": {
              "native": {
                "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
                "logical_cpus": 4,
                "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
                "python": "3.10.12",
                "numpy": "2.2.6",
                "host_observations": [
                  {
                    "phase": "before",
                    "timestamp": "2026-10-27T06:02:06.596454+00:00",
                    "loadavg": [
                      0.1982421875,
                      0.09814453125,
                      0.03369140625
                    ]
                  },
                  {
                    "phase": "after",
                    "timestamp": "2026-10-27T06:02:27.853666+00:00",
                    "loadavg": [
                      0.4267578125,
                      0.15771484375,
                      0.05517578125
                    ]
                  }
                ]
              },
              "chrome": {
                "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
                "logical_cpus": 4,
                "platform": "Linux 6.8.0-138-generic x64",
                "node": "v24.19.0",
                "host_observations": [
                  {
                    "phase": "before",
                    "timestamp": "2026-10-27T06:02:53.106Z",
                    "loadavg": [
                      0.28,
                      0.14,
                      0.05
                    ]
                  },
                  {
                    "phase": "after",
                    "timestamp": "2026-10-27T06:04:07.170Z",
                    "loadavg": [
                      0.8,
                      0.34,
                      0.13
                    ]
                  }
                ],
                "browser": "Chrome/150.0.7871.24"
              }
            },
            "warm": {
              "native_cpu_1t": {
                "cpu_threads": 1,
                "samples": 45,
                "median_ms": 21.849448,
                "mean_ms": 21.832480333333333,
                "p95_ms": 22.0759288
              },
              "chrome_wasm": {
                "cpu_threads": 1,
                "execution_context": "Dedicated module worker",
                "samples": 45,
                "median_ms": 80.80000019073486,
                "mean_ms": 81.25333338843451,
                "p95_ms": 82.78000020980835
              }
            },
            "protocol": {
              "native": {
                "cycles": 15,
                "warmups_per_input": 4,
                "initial_reference_executions_per_input": 1,
                "input_count": 3,
                "samples_per_model": 45,
                "batch_size": 1,
                "cpu_threads": 1,
                "order": "Alternate forward/reverse model order each cycle, for each fixed input",
                "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
                "host_idle_claimed": false,
                "timer_includes": [
                  "InferenceSession.run, input/output transport and owned output arrays"
                ],
                "timer_excludes": [
                  "input reads",
                  "hashing",
                  "preprocessing",
                  "loading/compilation",
                  "warmups",
                  "numerical validation"
                ],
                "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
                "source_input_indices": [
                  0,
                  999,
                  1999
                ],
                "input_sha256": [
                  "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
                  "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
                  "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
                ],
                "status": "passed",
                "tolerance": {
                  "rtol": 0.0005,
                  "atol": 0.001
                }
              },
              "chrome": {
                "cycles": 15,
                "warmups_per_input": 4,
                "initial_reference_executions_per_input": 1,
                "input_count": 3,
                "samples_per_model": 45,
                "batch_size": 1,
                "cpu_threads": 1,
                "order": "Alternate forward/reverse model order each cycle, for each fixed input",
                "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
                "host_idle_claimed": false,
                "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
                "source_input_indices": [
                  0,
                  999,
                  1999
                ],
                "input_sha256": [
                  "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
                  "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
                  "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
                ],
                "status": "passed",
                "tolerance": {
                  "rtol": 0.0005,
                  "atol": 0.001
                },
                "browser": "Chrome/150.0.7871.24",
                "execution_context": "Chrome dedicated module worker",
                "timer_includes": [
                  "execute dispatch, input/output transport and copied inline output"
                ],
                "timer_excludes": [
                  "preprocessing",
                  "model verification",
                  "load",
                  "compile",
                  "context creation",
                  "numerical validation",
                  "UI",
                  "extension queue"
                ]
              }
            },
            "scope": "Interleaved fixed-input comparison on one Intel i3-1115G4 host; 3 heldout inputs, 15 cycles, 45 executions per model. Full 2000-record accuracy-pass timing is auxiliary in digit-comparison.json."
          }
        }
      ],
      "selection_reason": "Frozen epoch-2 joint QAT size-priority selection: 1.646 MiB, 58.2% smaller than W8A8, with 29/2000 fewer complete records (1.45 percentage points). Current Chrome medians on the same i3-1115G4 host are 88.20 ms versus 80.80 ms for W8A8. Conv3/Linear3 QAT is smaller at 1.461 MiB, with 93.50% complete-record accuracy. All 13 frozen alternatives are compared without retraining or choosing a new default from heldout results.",
      "latency": {
        "date": "2026-10-27",
        "runtime_version": "0.9.0",
        "environment": {
          "native": {
            "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
            "logical_cpus": 4,
            "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
            "python": "3.10.12",
            "numpy": "2.2.6",
            "host_observations": [
              {
                "phase": "before",
                "timestamp": "2026-10-27T06:02:06.596454+00:00",
                "loadavg": [
                  0.1982421875,
                  0.09814453125,
                  0.03369140625
                ]
              },
              {
                "phase": "after",
                "timestamp": "2026-10-27T06:02:27.853666+00:00",
                "loadavg": [
                  0.4267578125,
                  0.15771484375,
                  0.05517578125
                ]
              }
            ]
          },
          "chrome": {
            "cpu": "11th Gen Intel(R) Core(TM) i3-1115G4 @ 3.00GHz",
            "logical_cpus": 4,
            "platform": "Linux 6.8.0-138-generic x64",
            "node": "v24.19.0",
            "host_observations": [
              {
                "phase": "before",
                "timestamp": "2026-10-27T06:02:53.106Z",
                "loadavg": [
                  0.28,
                  0.14,
                  0.05
                ]
              },
              {
                "phase": "after",
                "timestamp": "2026-10-27T06:04:07.170Z",
                "loadavg": [
                  0.8,
                  0.34,
                  0.13
                ]
              }
            ],
            "browser": "Chrome/150.0.7871.24"
          }
        },
        "warm": {
          "native_cpu_1t": {
            "cpu_threads": 1,
            "samples": 45,
            "median_ms": 27.103447,
            "mean_ms": 27.1741368,
            "p95_ms": 27.4313216
          },
          "chrome_wasm": {
            "cpu_threads": 1,
            "execution_context": "Dedicated module worker",
            "samples": 45,
            "median_ms": 88.2000002861023,
            "mean_ms": 88.49777782228259,
            "p95_ms": 90.68000020980834
          }
        },
        "protocol": {
          "native": {
            "cycles": 15,
            "warmups_per_input": 4,
            "initial_reference_executions_per_input": 1,
            "input_count": 3,
            "samples_per_model": 45,
            "batch_size": 1,
            "cpu_threads": 1,
            "order": "Alternate forward/reverse model order each cycle, for each fixed input",
            "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
            "host_idle_claimed": false,
            "timer_includes": [
              "InferenceSession.run, input/output transport and owned output arrays"
            ],
            "timer_excludes": [
              "input reads",
              "hashing",
              "preprocessing",
              "loading/compilation",
              "warmups",
              "numerical validation"
            ],
            "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
            "source_input_indices": [
              0,
              999,
              1999
            ],
            "input_sha256": [
              "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
              "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
              "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
            ],
            "status": "passed",
            "tolerance": {
              "rtol": 0.0005,
              "atol": 0.001
            }
          },
          "chrome": {
            "cycles": 15,
            "warmups_per_input": 4,
            "initial_reference_executions_per_input": 1,
            "input_count": 3,
            "samples_per_model": 45,
            "batch_size": 1,
            "cpu_threads": 1,
            "order": "Alternate forward/reverse model order each cycle, for each fixed input",
            "input_selection": "Fixed ordered positions 0,999,1999; no label/accuracy-based selection",
            "host_idle_claimed": false,
            "timing_scope": "Interleaved repeated model timings; separate from full-corpus accuracy and sequential evaluation-pass latency",
            "source_input_indices": [
              0,
              999,
              1999
            ],
            "input_sha256": [
              "68437169b7391da9b2f010464a9d4b94dd7852a3266e7f576344d8198f5eabf0",
              "00e632349cf86f4ac25e0374d4abb9a49865458a31f261540ffb1d0c7db40c32",
              "a31d50c275adb730d6d91a908b468e74af51471c528766dd1d9489074ecc7b55"
            ],
            "status": "passed",
            "tolerance": {
              "rtol": 0.0005,
              "atol": 0.001
            },
            "browser": "Chrome/150.0.7871.24",
            "execution_context": "Chrome dedicated module worker",
            "timer_includes": [
              "execute dispatch, input/output transport and copied inline output"
            ],
            "timer_excludes": [
              "preprocessing",
              "model verification",
              "load",
              "compile",
              "context creation",
              "numerical validation",
              "UI",
              "extension queue"
            ]
          }
        },
        "scope": "Interleaved fixed-input comparison on one Intel i3-1115G4 host; 3 heldout inputs, 15 cycles, 45 executions per model. Full 2000-record accuracy-pass timing is auxiliary in digit-comparison.json."
      },
      "validation": {
        "full_corpus": {
          "slot_mismatches": 0,
          "numerical_mismatches": 118,
          "byte_mismatches": 2000,
          "maximum_absolute_logit_difference": 0.13301706314086914
        },
        "tolerance": {
          "atol": 0.001,
          "rtol": 0.0005
        },
        "input_count": 2000,
        "latency_validation": {
          "native": {
            "native_reference_bytes_equal": true,
            "native_reference_slots_equal": true,
            "native_reference_logits_within_tolerance": true,
            "maximum_absolute_logit_difference": 0.0
          },
          "chrome": {
            "native_reference_slots_equal": true,
            "native_reference_logits_within_tolerance": true,
            "maximum_absolute_logit_difference": 6.67572021484375e-06
          }
        },
        "status": {
          "native": "passed",
          "chrome": "completed-with-numerical-differences"
        },
        "scope": "Full native/Chrome accuracy and slot/logit comparison on the accuracy host; separate 3-input validation on the latency host."
      },
      "comparison_file": "digit-comparison.json"
    },
    "vqa": {
      "name": "VQA W4A8 PTQ",
      "identity": {
        "profile": "w4a8",
        "method": "post-training quantization",
        "sha256": {
          "config.json": "e18e266da200b9b915628c1b8f633cb523520287e2d8f75c8093e3f4d6575f7e",
          "vocab.json": "ae19bf56556649d75e72450ff6c7d43895190404c15dffeb6412dce5cc4d2c80",
          "encoder/graph.json": "c9f1a40e5d0a4d057b03fcb0ee3c65b062768a1991e8b86d674a192bcd20a1c1",
          "encoder/model.safetensors": "3b91f0f688d23397d14cd506a9faafc1b3e85a68ac12aaac2bcddd49bdfa7757",
          "decoder/graph.json": "ae02afd8c84c2927068dbe4a8ef0a77bcec646b3274d6f83b6df6f02e19f2825",
          "decoder/model.safetensors": "5cfd2a08db56605f8ff1aa23ef7786f83e1415e28d57a2ec542e1a0f6de1312c",
          "shared/token-embedding.safetensors": "2ddc31c77c47498d65ee5604f5ee202704aa32a4878b10cf1c171618306a4a60"
        },
        "source_manifest_sha256": "ad4f2576b6401144bf467693e82d12d3ce9db7dc7cbe3d7a7c1e4e57c1f0e3db",
        "package_bytes": 21617105,
        "package_size_scope": "Deployable graphs, weights, shared embedding counted once, vocabulary, config and benchmark manifest; excludes runtime and ZIP. Release wrapper metadata size may differ.",
        "linear_layers": 67,
        "linear_weight_bits": 4,
        "linear_group_size": 32,
        "convolution_layers": 13,
        "conv_weight_bits": 8,
        "packed_arithmetic": "dynamic-i8",
        "activation_policy": "Dynamic INT8 dense activations; static INT8 convolution activations; embeddings, normalization and attention retain FP32."
      },
      "reference_w8a8": {
        "profile": "w8a8",
        "package_bytes": 27337570,
        "comparison_scope": "Existing W8A8 has a different static activation recipe and calibration. This comparison does not isolate weight bit width or prove a kernel-level cause."
      },
      "comparators": [
        {
          "profile": "w8a8",
          "name": "W8A8",
          "package_bytes": 27337570,
          "sha256": {
            "configuration": "e18e266da200b9b915628c1b8f633cb523520287e2d8f75c8093e3f4d6575f7e",
            "decoder_graph": "2b1e0a54b16dbb57ae9315f66b7212e3e986435c4a080922807e5cb0e4d375c7",
            "decoder_weights": "a19a2ef115053174b7e927f76d709cb0a66e2eb8ed0ce3340624d14d12ba1726",
            "encoder_graph": "28fb62de5eee6865fdb30ae362ae3fd466be5e42d0af749b72e5896c5af70727",
            "encoder_weights": "150a513e1f1f30689d5d9bf932c4d4675bca32d4c7f5531bc2f23fd4485117c1",
            "manifest": "1e3931519b47d6a36ea31018561d4e2a8eee8da0ec360bbaf09cdb51c661f5f3",
            "shared_token_embedding": "8b9c2ffa44618dae2bc8169ceb6409cf99bee4145f336999e70a5913b4600e6d",
            "vocabulary": "ae19bf56556649d75e72450ff6c7d43895190404c15dffeb6412dce5cc4d2c80"
          }
        }
      ],
      "selection_reason": "20.9% smaller than W8A8. VolvoxAI 0.9.0 Chrome answer medians are 823.9 ms for W4A8 and 894.5 ms for W8A8 on the same three public questions.",
      "latency": {
        "date": "2026-10-27",
        "w4a8": {
          "complete_answer_median_ms": 823.8999999761581,
          "encoder_median_ms": 618.8000000119209,
          "greedy_decoder_median_ms": 203.59999990463257,
          "fixed_prefix_decoder_median_ms_per_token": 12.474999997764826,
          "samples": 15
        },
        "reference_w8a8": {
          "complete_answer_median_ms": 894.5,
          "encoder_median_ms": 542.6999999880791,
          "greedy_decoder_median_ms": 348.89999997615814,
          "fixed_prefix_decoder_median_ms_per_token": 21.493750002235174,
          "samples": 15
        },
        "protocol": {
          "batch_size": 1,
          "input_count": 3,
          "input_source": "Three public questions (phone, address, store) on one English receipt; preprocessed 320x672 grayscale",
          "input_sha256": [
            "88ecddb4769b6bf6a3b3ec575eebdf1ee2207d0d36915afcab398ff037af5d95",
            "88ecddb4769b6bf6a3b3ec575eebdf1ee2207d0d36915afcab398ff037af5d95",
            "88ecddb4769b6bf6a3b3ec575eebdf1ee2207d0d36915afcab398ff037af5d95"
          ],
          "warmups_per_input": 1,
          "cycles": 5,
          "samples_per_model": 15,
          "fixed_prefix_decoder_steps": 16,
          "ordering": "Alternate forward/reverse model order each measured cycle",
          "execution_context": "Real Chrome dedicated module worker",
          "host_idle_claimed": false
        },
        "timer_includes": [
          "execute dispatch",
          "input/output transport",
          "copied selected outputs",
          "tokenization/argmax/parsing in total generation time"
        ],
        "timer_excludes": [
          "image preprocessing",
          "model verification/load/compile",
          "warmups",
          "UI",
          "extension queue"
        ],
        "interpretation": "Stage medians are pooled independently and need not sum to the complete-answer median. Greedy output length varies; fixed-prefix timing holds 16 tokens and cache lengths constant.",
        "environment": {
          "cpu": "AMD Ryzen 5 5600U with Radeon Graphics",
          "platform": "Linux-6.8.0-138-generic-x86_64-with-glibc2.35",
          "python": "3.10.12",
          "chrome": "150.0.7871.24",
          "wasm_cpu_threads": 1,
          "batch_size": 1,
          "host_idle_claimed": false,
          "node": "v24.19.0",
          "firefox": "firefox/158.0.1"
        },
        "runtime_version": "0.9.0",
        "runtime_sha256": {
          "volvoxai.lite.min.js": "75a5611d873b25c4997523d8fb527f33dde4da882efd64f11bcdd2a0eb38be6b",
          "volvoxai.lite.wasm": "5fc82164fba4fa368b7c6ded2c02f1f0fa030ac15c8b6fc36ee1798e0e8529d6"
        }
      },
      "validation": {
        "timed_w4a8_public_questions_valid": 3,
        "timed_w4a8_public_questions_total": 3,
        "repeated_encoder_outputs_bit_exact": true
      },
      "accuracy_status": "Full-heldout accuracy evaluated on 0.9.0 (97.15%, 1943/2000)."
    }
  },
  "interpretation": [
    "Digit reports exact complete phone/street records and the answer requested by each question, both evaluated on 2000 inputs with VolvoxAI 0.9.0.",
    "All 13 Digit graph/weight artifacts are unchanged. Current package runtime version and original model-exporter provenance are recorded separately.",
    "Native and Chrome decoded outputs match for all Digit inputs. Numerical logit tolerance differences remain recorded per model.",
    "Primary Digit latency uses three fixed heldout inputs and interleaved model order on one i3-1115G4. Full-corpus accuracy-pass timing is auxiliary.",
    "VQA reports 0.9.0 latency over three public questions; full-heldout 0.9.0 accuracy is 97.15% (1,943/2,000).",
    "Warm timings exclude image preprocessing, model loading and extension UI/queue overhead.",
    "The heldout set has 1986 distinct image contents across 2000 records and prior development/evaluation use."
  ],
  "evidence": {
    "consolidation": "Only frozen final model choices, relevant controls, current runtime measurements and selection rationale retained; historical optimization logs removed. Private records excluded.",
    "current_reports_sha256": {
      "vqa_chrome": "915afeae3f022fb271b578cc68ac62bc2c87478913c7d537cfd092fc6e7414d2"
    },
    "vqa_sources_sha256": {
      "benchmarks/vqa_wasm.mjs": "fd3f0c24490f4b82953f082deeb310385e031b2bda4feddf0666391870665b30",
      "benchmarks/vqa-wasm-worker.js": "3e9d569a6842c9539572245e9217ee3ebab12bbdfbd4ea40ebd976e473ecbb3d",
      "vqa/adapter.js": "ca421c4efcf8e22b876dd1c5bae8130d0fcce0a0c719cac55858ce6d421e3133",
      "vqa/tokenizer.js": "c2c72e977d786674fda4aa4008719fd10dcccd7779ad56af0324f409c88b9ce5"
    },
    "vqa_fixture_sha256": "0f10332fe8adf7cae0303ee5cdc229e77b14fdc6cd2d4b423d344b6e08af57c4",
    "digit_comparison_sha256": "b4ca90e83f29fae05d233a1f480f861fadb2443fe8bc636b36cfd77e91bc059e",
    "digit_comparison_file": "digit-comparison.json",
    "vqa_sources_scope": "Hashes identify the code used for the recorded 0.9.0 measurements."
  },
  "qualification": {
    "status": "completed-with-numerical-differences",
    "scope": "VolvoxAI 0.9.0 Digit native/Chrome evaluation of 13 models on 2000 records plus CPU/Chrome latency. Numerical tolerance differences are preserved.",
    "digit_full_corpus_accuracy_reevaluated": true,
    "vqa_full_corpus_accuracy_reevaluated": true,
    "runtime_parity_passed": false,
    "cuda_tested": true,
    "artifact_exporter_version": "0.8.0",
    "comparison_file": "digit-comparison.json",
    "integration_validation": {
      "runtime_version": "0.9.0",
      "javascript": {
        "passed": 36,
        "total": 36
      },
      "python": {
        "passed": 118,
        "total": 118
      },
      "models": {
        "digit": {
          "browser_questions": {
            "correct": 8,
            "total": 8
          },
          "native_questions": {
            "correct": 8,
            "total": 8
          },
          "lifecycle_checks": 6,
          "browser_report_sha256": "9aee7e73a4d77ec5c3ce99a94a148489d722d324a0be05149feeeca1312e3ba1",
          "native_report_sha256": "20bb9e2d03a60f84b5920af1e45511db85b3c6c835f6eb68843ab15403137466",
          "lifecycle_report_sha256": "2ae462978ff767b65949a59119b47b67b013a8b249d87e2646bd07420ee53529",
          "build_info_sha256": "0864968988d39de9eb08b6dd5d6b79ff6df8f88a90ee7e5d74c250b3bd1720a0"
        },
        "vqa": {
          "browser_questions": {
            "correct": 16,
            "total": 16
          },
          "native_questions": {
            "correct": 16,
            "total": 16
          },
          "lifecycle_checks": 6,
          "browser_report_sha256": "6d744b87cffff828db803c7cceb4b88f337a5c14a99ba593f1f9cdd294a7a7a0",
          "native_report_sha256": "66ecdde0d1fa1e2421be5e44053b86b41368d4e089f131ccb2304762ac692e4f",
          "lifecycle_report_sha256": "a0d0fc2d97ac9a819c0b28f9128eead5112ad4702a21b7465352c8c3cc7b1270",
          "build_info_sha256": "f3f4085b2fc25d5cd15b166f0fc324a8aaa73e5e14e02632030d9c086758cdac"
        }
      }
    }
  }
}