{
  "paper": {
    "title": "Learning What to Remember: Test-Time Training via Context Distillation"
  },
  "dataset": {
    "name": "deterministic long-context public mini-suite",
    "samples": 64,
    "sequence_length": 96
  },
  "setup": {
    "adapter": "ttcd",
    "seed": 42,
    "same_examples": true
  },
  "baseline": {
    "name": "recent-window attention",
    "accuracy": 0.078125,
    "memory_fraction": 0.25,
    "relative_cost": 1.0
  },
  "method": {
    "name": "ttcd",
    "accuracy": 0.859375,
    "memory_fraction": 0.25,
    "relative_cost": 0.42,
    "context_distillation_updates": 64.0
  },
  "relative": {
    "accuracy_percent": 1000.0,
    "accuracy_points": 78.125,
    "cost_percent": -58.00000000000001
  },
  "paper_results": {
    "reported": "IP-TTCD 在长上下文语言建模中持续优于 DeltaNet、Gated DeltaNet、滑窗注意力和 TTT。"
  },
  "scope": "固定 numpy 长上下文/多模态/评测 mini-suite；不冒充论文规模预训练、GPU kernel 或完整公开 benchmark。",
  "schema_version": 2,
  "manifest_ref": "reproduction:ttcd",
  "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/2608.01672-ttcd/metrics/public-seed42.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}