{
  "checked_at": "2026-09-23T22:40:31.825277+00:00",
  "scope": "CPU tests + synthetic fixtures; no Ornith GPU training or serving",
  "suites": {
    "toy-examples": {
      "passed": true,
      "output": "test_deterministic (test_distill_toy.DistillationTests.test_deterministic) ... ok\ntest_invalid_temperature (test_distill_toy.DistillationTests.test_invalid_temperature) ... ok\ntest_optimization_improves_held_out (test_distill_toy.DistillationTests.test_optimization_improves_held_out) ... ok\ntest_self_kl_zero (test_distill_toy.DistillationTests.test_self_kl_zero) ... ok\ntest_stable_sigmoid (test_distill_toy.DistillationTests.test_stable_sigmoid) ... ok\ntest_empty_router_rejected (test_prune_quantize_toy.PruningQuantizationTests.test_empty_router_rejected) ... ok\ntest_quant_codes_bound_and_zero_safe (test_prune_quantize_toy.PruningQuantizationTests.test_quant_codes_bound_and_zero_safe) ... ok\ntest_removed_expert_and_disjoint_holdout (test_prune_quantize_toy.PruningQuantizationTests.test_removed_expert_and_disjoint_holdout) ... ok\ntest_router_renormalizes (test_prune_quantize_toy.PruningQuantizationTests.test_router_renormalizes) ... ok\n\n----------------------------------------------------------------------\nRan 9 tests in 0.010s\n\nOK\n"
    },
    "training-examples": {
      "passed": true,
      "output": "test_empty_answer_rejected (test_lora_smoke.MaskingTests.test_empty_answer_rejected) ... ok\ntest_masks_prompt (test_lora_smoke.MaskingTests.test_masks_prompt) ... ok\ntest_mismatch_rejected (test_lora_smoke.MaskingTests.test_mismatch_rejected) ... ok\ntest_existing_output_is_preserved (test_prepare_distill_dataset.DistillationDatasetTests.test_existing_output_is_preserved) ... ok\ntest_nested_reasoning_and_think_are_rejected (test_prepare_distill_dataset.DistillationDatasetTests.test_nested_reasoning_and_think_are_rejected) ... ok\ntest_no_partial_output_when_split_missing (test_prepare_distill_dataset.DistillationDatasetTests.test_no_partial_output_when_split_missing) ... ok\ntest_prepare_emits_separate_sft_files_and_manifest (test_prepare_distill_dataset.DistillationDatasetTests.test_prepare_emits_separate_sft_files_and_manifest) ... ok\ntest_private_reasoning_is_rejected (test_prepare_distill_dataset.DistillationDatasetTests.test_private_reasoning_is_rejected) ... ok\ntest_prompt_contamination_is_rejected (test_prepare_distill_dataset.DistillationDatasetTests.test_prompt_contamination_is_rejected) ... ok\ntest_task_key_contamination_is_rejected (test_prepare_distill_dataset.DistillationDatasetTests.test_task_key_contamination_is_rejected) ... ok\n\n----------------------------------------------------------------------\nRan 10 tests in 0.004s\n\nOK\n"
    },
    "moe-examples": {
      "passed": true,
      "output": "test_20_percent_rounds_to_205_like_published_card (test_plan.CandidatePlanTests.test_20_percent_rounds_to_205_like_published_card) ... ok\ntest_30_percent_is_a_candidate_not_a_found_checkpoint (test_plan.CandidatePlanTests.test_30_percent_is_a_candidate_not_a_found_checkpoint) ... ok\ntest_50_percent_is_128_and_keeps_topk (test_plan.CandidatePlanTests.test_50_percent_is_128_and_keeps_topk) ... ok\ntest_invalid_fraction_fails (test_plan.CandidatePlanTests.test_invalid_fraction_fails) ... ok\ntest_never_retains_fewer_than_topk (test_plan.CandidatePlanTests.test_never_retains_fewer_than_topk) ... ok\n\n----------------------------------------------------------------------\nRan 5 tests in 0.000s\n\nOK\n"
    }
  },
  "distillation_toy": {
    "scope": "CPU toy distillation only; not an Ornith quality result",
    "teacher_coefficients": 3,
    "student_parameters": 2,
    "train_points": 17,
    "held_out_points": 16,
    "epochs": 1500,
    "temperature": 2.0,
    "weights": [
      2.6188783342018813,
      -0.20132097946757457
    ],
    "held_out_kl_before": 0.3468258960081315,
    "held_out_kl_after": 0.008208858418184266,
    "overlap": 0
  },
  "pruning_quantization_toy": {
    "scope": "toy only; dense all-expert routing, not Ornith top-k or REAP",
    "experts_before": 3,
    "experts_after": 2,
    "kept_indices": [
      0,
      1
    ],
    "pruned_expert_weights": [
      [
        1.5,
        0.2
      ],
      [
        -0.4,
        0.6
      ]
    ],
    "salience": [
      0.6881908878562988,
      0.35008240073520247,
      0.24684428924199564
    ],
    "quantized_integer_codes": [
      [
        7,
        1
      ],
      [
        -2,
        3
      ],
      [
        4,
        -1
      ]
    ],
    "scale": 0.21428571428571427,
    "holdout_mse_pruning_only": 0.14662443955775456,
    "holdout_mse_quantization_only": 0.003648624157039457,
    "holdout_mse_combined": 0.18334473352483888,
    "calibration_holdout_overlap": 0,
    "packed_4bit_storage": false
  },
  "moe_plan": {
    "source": "moe-examples/fixtures/config-ornith-observed.json",
    "candidates": [
      {
        "status": "PLAN_ONLY",
        "experts_before": 256,
        "experts_after_candidate": 205,
        "experts_removed_candidate": 51,
        "actual_removed_fraction": 0.19921875,
        "active_experts_per_token": 8,
        "layers": 40,
        "mtp_layers_present": 1,
        "constraints": [
          "do not prune shared experts",
          "verify router renormalization after structural pruning",
          "inspect MTP/vision paths before changing config",
          "run reload + forward before quantization",
          "calibrate and evaluate by domain/language; do not label experts by language"
        ]
      },
      {
        "status": "PLAN_ONLY",
        "experts_before": 256,
        "experts_after_candidate": 179,
        "experts_removed_candidate": 77,
        "actual_removed_fraction": 0.30078125,
        "active_experts_per_token": 8,
        "layers": 40,
        "mtp_layers_present": 1,
        "constraints": [
          "do not prune shared experts",
          "verify router renormalization after structural pruning",
          "inspect MTP/vision paths before changing config",
          "run reload + forward before quantization",
          "calibrate and evaluate by domain/language; do not label experts by language"
        ]
      },
      {
        "status": "PLAN_ONLY",
        "experts_before": 256,
        "experts_after_candidate": 128,
        "experts_removed_candidate": 128,
        "actual_removed_fraction": 0.5,
        "active_experts_per_token": 8,
        "layers": 40,
        "mtp_layers_present": 1,
        "constraints": [
          "do not prune shared experts",
          "verify router renormalization after structural pruning",
          "inspect MTP/vision paths before changing config",
          "run reload + forward before quantization",
          "calibrate and evaluate by domain/language; do not label experts by language"
        ]
      }
    ]
  },
  "synthetic_dataset": {
    "counts": {
      "eval": 2,
      "train": 3
    },
    "format": "openai-compatible-sft-messages",
    "ids": [
      "code-001",
      "code-101",
      "tool-001",
      "tool-002",
      "tool-101"
    ],
    "known_reasoning_fields_and_think_markers_rejected": true,
    "output_sha256": {
      "eval": "d1c0384df703d11609d88eff9ec2e40a2e16896eb9adc0ab575247ce3ff86210",
      "train": "2fc2ad7e72433011eb1910dca677d072f59647c66f112533c63dfc855fd410f2"
    },
    "quality_pass_is_producer_assertion": true,
    "record_count": 5,
    "semantic_content_audited": false,
    "source": "training-examples/teacher_candidates.jsonl",
    "tool_execution": false
  },
  "lora_fixture_plan": {
    "status": "PLAN_ONLY",
    "architecture": "Qwen3_5ForConditionalGeneration",
    "rows_available": 3,
    "rows_used": 1,
    "optimizer_steps": 1,
    "dtype": "bfloat16",
    "rank": 8,
    "max_tokens": 256,
    "scope": "language q_proj/v_proj only; no vision/router/expert adapters"
  },
  "gpu_executed": false
}
