{
  "date": "2026-10-02",
  "scope": "Documentation only. Private SI code, weights, user memory and raw datasets excluded.",
  "reports": {
    "math_256": {
      "reader": {
        "fresh": {
          "before": {
            "n": 140,
            "correct": 140,
            "wrong_accepted": 0,
            "lost_correct": 0
          },
          "after": {
            "n": 140,
            "correct": 140,
            "wrong_accepted": 0,
            "lost_correct": 0
          }
        },
        "regression": {
          "before": {
            "n": 282,
            "correct": 282,
            "wrong_accepted": 0,
            "lost_correct": 0
          },
          "after": {
            "n": 282,
            "correct": 282,
            "wrong_accepted": 0,
            "lost_correct": 0
          }
        },
        "parameters": 526086,
        "checkpoint_bytes": 2106643,
        "gate_passed": true
      },
      "planner": {
        "fresh": {
          "before": {
            "n": 96,
            "correct": 96,
            "wrong_accepted": 0,
            "lost_correct": 0
          },
          "after": {
            "n": 96,
            "correct": 96,
            "wrong_accepted": 0,
            "lost_correct": 0
          }
        },
        "regression": {
          "before": {
            "n": 192,
            "correct": 192,
            "wrong_accepted": 0,
            "lost_correct": 0
          },
          "after": {
            "n": 192,
            "correct": 192,
            "wrong_accepted": 0,
            "lost_correct": 0
          }
        },
        "parameters": 525829,
        "checkpoint_bytes": 2105629,
        "gate_passed": true
      },
      "source_unchanged": true,
      "fresh_unchanged": true,
      "activated": false
    },
    "language_comparison": {
      "bielik": {
        "fresh": {
          "n": 8,
          "correct": 3,
          "supported_correct": 0,
          "wrong_accepted": 0
        }
      },
      "qwen": {
        "fresh": {
          "n": 8,
          "correct": 8,
          "supported_correct": 5,
          "wrong_accepted": 0
        },
        "regression": {
          "n": 8,
          "correct": 8,
          "supported_correct": 5,
          "wrong_accepted": 0
        }
      },
      "gate_passed": true,
      "dataset_unchanged": true,
      "source_unchanged": true
    },
    "qwen_modes": {
      "n": 8,
      "passed": 8,
      "gate_passed": true,
      "dataset_unchanged": true,
      "source_unchanged": true,
      "weights_changed": false
    }
  },
  "sizes_bytes": {
    "qwen": 2497280448,
    "inactive_bielik": 1699568288,
    "math_reader_256": 2106643,
    "math_planner_256": 2105629
  },
  "limitations": [
    "Small development tests in known controlled grammar; not general AI benchmark.",
    "In language comparison, three refusal inputs are blocked before language inference.",
    "Arithmetic values checked independently; this does not validate arbitrary problem assumptions.",
    "256-neuron modules matched 128-neuron modules; no improvement demonstrated by doubling.",
    "Different language quantization: Qwen Q4_K_M, Bielik Q8_0.",
    "Eight Qwen mode checks are functional smoke tests, not proof of general quality or security."
  ]
}