{
  "schema_version": 2,
  "evaluated_at": "2026-07-24",
  "models": {
    "teacher_reference": "gemma-4-12B-it BF16",
    "candidate": "gemma-4-12B-it AutoRound INT4"
  },
  "shared_controls": {
    "tokenizer": "same tokenizer for both runs",
    "chat_template": "same explicit Jinja template for both runs",
    "datasets": [
      "WikiText-2 test split",
      "HumanEval canonical solutions"
    ],
    "checkpoint_tokenizer_configs_differed": true,
    "template_note": "The checkpoint tokenizer configurations differed, so both runs used the same explicit chat template."
  },
  "runtime": {
    "max_model_len": 131072,
    "max_num_seqs": 4,
    "max_num_batched_tokens": 8192,
    "gpu_memory_utilization": 0.98,
    "kv_cache": "native BF16",
    "prefix_caching": true,
    "language_model_only": true,
    "graph_mode": "FULL_DECODE_ONLY",
    "graph_capture_sizes": [1, 2, 4],
    "server_template_kwargs": {
      "enable_thinking": true,
      "preserve_thinking": true
    },
    "generation_controls": {
      "temperature": 1.0,
      "top_p": 0.95,
      "top_k": 64,
      "min_p": 0.0,
      "presence_penalty": 0.0,
      "repetition_penalty": 1.0
    }
  },
  "evaluation": {
    "fixed_target": {
      "metric": "conditional assistant-target perplexity",
      "wikitext_2_cases": 16,
      "humaneval_canonical_code_cases": 64,
      "executes_code": false
    },
    "monte_carlo_kl": {
      "prompt_classes": 6,
      "samples_per_prompt": 4,
      "sequences_per_direction": 24,
      "scored_tokens_per_sequence_when_capped": 192,
      "thinking_enabled": false,
      "confidence_interval": "95% bootstrap",
      "rendered_text_retokenized": true,
      "stop_token_probability_included": false
    },
    "needle_matrix": {
      "deterministic": true,
      "contexts": [4096, 16384, 32768, 65536, 98304, 126000],
      "depth_fractions": [0.1, 0.5, 0.9],
      "cases_per_model": 18
    }
  },
  "limitations": [
    "The evaluator source revision was not recorded.",
    "The Monte Carlo KL values come from finite rendered-and-retokenized sequence samples rather than aligned full-vocabulary logits.",
    "The 18 deterministic needle cases do not set an upper bound on retrieval failures."
  ]
}