{
  "paper": {
    "arxiv_id": "2607.25915",
    "title": "Penelope: Localized Latent Recurrence for Efficient Structured Reasoning",
    "url": "https://arxiv.org/abs/2607.25915",
    "organization": "Academic author team"
  },
  "dataset": {
    "name": "WikiText-2",
    "train_tokens": 120000,
    "validation_tokens": 8000,
    "test_tokens": 8000,
    "narrative_tokens": 120000,
    "instruction_train": 320,
    "instruction_validation": 64,
    "preference_validation": 64,
    "reasoning_validation": 0,
    "benchmark_suite": "core",
    "vocab_size": 512,
    "tokenizer": "data/llm-evolution/wikitext-narrative-bpe-512.json"
  },
  "setup": {
    "seed": 42,
    "steps_per_variant": 12,
    "dimensions": 64,
    "layers": 2,
    "sequence_length": 64,
    "same_tokens_optimizer_and_budget": true,
    "evolve_architecture": "penelope",
    "evolve_optimizer": "adamw"
  },
  "baseline": {
    "name": "llama_modern",
    "lm_loss": 6.043058067560196,
    "perplexity": 421.17906153137636,
    "instruction_loss": 6.0724897384643555,
    "primary": -6.953931528329849,
    "public_composite": -6.953931528329849,
    "composite_loss": 6.953931528329849,
    "fitness": -6.953931528329849,
    "initial_loss": 6.17989764213562,
    "final_loss": 6.145263242721557,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 139584,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {},
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "method": {
    "name": "penelope",
    "lm_loss": 6.004443973302841,
    "perplexity": 405.2256098254262,
    "instruction_loss": 6.037044286727905,
    "primary": -6.910000616312027,
    "public_composite": -6.910000616312027,
    "composite_loss": 6.910000616312027,
    "fitness": -6.910000616312027,
    "initial_loss": 6.1616936206817625,
    "final_loss": 6.120111656188965,
    "post_training_loss": 0.0,
    "reward_bearing_group_rate": 0.0,
    "parameters": 222018,
    "device": "mps",
    "optimizer": "adamw",
    "architecture_stats": {
      "localized_decoder_interval": [
        0,
        1
      ],
      "latent_recurrence_steps": 2.0,
      "full_decoder_reexecution": 0.0
    },
    "muon_orthogonality_error": 0.0,
    "seeds": [
      42
    ]
  },
  "relative": {
    "lm_loss_percent": -0.6389826777379397,
    "perplexity_percent": -3.7878074109250757
  },
  "paper_results": {
    "claim": "competitive structured-reasoning accuracy with lower measured latency",
    "latent_recurrence_localized": true
  },
  "scope": "真实在 micro decoder 中只对一个中间边界执行两步 GRU latent refinement，下层前缀不重复计算；WikiText-2 同预算训练不等同于论文结构推理 benchmark 和完整 CoT-to-latent curriculum。",
  "schema_version": 2,
  "manifest_ref": "reproduction:penelope",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2607.25915-penelope/metrics/result-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}