Publish curated RAP benchmark dataset release

ARXIV_METADATA_NOTICE.md

@@ -0,0 +1,20 @@
+# arXiv metadata notice
+
+The Search shards in this release contain descriptive arXiv metadata only:
+arXiv identifier, title, abstract, submission date, and arXiv categories. They
+do not contain PDFs, TeX/source files, figures, or full-text article content.
+
+arXiv's API terms identify title, abstract, authors, identifiers, and
+classification terms as descriptive metadata that may be used, stored,
+transformed, and shared under CC0. See:
+
+<https://info.arxiv.org/help/api/tou.html>
+
+This release preserves the arXiv identifier for each record. The corresponding
+abstract page is `https://arxiv.org/abs/{id}`. RAP's field selection, episode
+boundaries, assignments, and target labels are derived RAP artifacts and are
+covered separately by the dataset license selected by the authors.
+
+This release is not endorsed by arXiv. It is an independent research artifact.
+The project acknowledges: “Thank you to arXiv for use of its open access
+interoperability.”

DATA_CARD.md

@@ -0,0 +1,76 @@
+# RAP Benchmark Dataset Card
+
+## Task
+
+RAP evaluates whether an agent can estimate the distribution of newly posted
+papers across eight operational research directions and forecast the
+distribution in the following six months. Each episode uses `forecast_date=T`
+as its boundary. The history windows describe the available past periods and
+the future window is `[T, T+6 months)`.
+
+The State task estimates the distribution observed in the most recent
+six-month window. The Forecast task predicts the distribution in the next
+six-month window.
+
+## Package contents
+
+- `inputs/` contains the model-facing Dev/Test episode inputs.
+- `targets/` contains evaluator-only recent and future counts, shares, and
+  diagnostics.
+- `evaluator/` contains combined episode records for scoring and offline
+  baselines.
+- `data/` contains the frozen evaluation scope and Search-manifest indexes.
+- `search_shards/` contains the metadata-only Search snapshot.
+
+Each row in `inputs/` contains the field, cutoff date, time-window metadata,
+and eight direction definitions. It does not contain history counts, recent
+shares, future counts, or future shares. Those evaluator-only values are stored
+in `targets/` and `evaluator/`.
+
+## Splits and evaluation
+
+Dev and Test are frozen field-level splits rather than random episode-level
+splits. Dev contains 335 episodes from 67 fields; Test contains 1,055 episodes
+from 211 fields. The evaluation-code repository provides the prompt protocol,
+scorer, contract checks, and offline statistical baselines.
+
+The canonical count assigns each `(field, paper_id)` to one primary direction.
+Secondary assignments are retained in the evaluator records for the specified
+sensitivity analyses. The direction definitions and episode boundaries are
+fixed by the accompanying manifests.
+
+## Search snapshot
+
+`search_shards/fixed/` enforces the window `[T-6m, T)`, while
+`search_shards/expanding/` enforces `<T`. Each shard contains exactly
+`id`, `date`, `title`, `abstract`, and `categories`. The shard manifests record
+SHA-256 digests, row counts, and date-bound checks.
+
+The Search snapshot contains descriptive arXiv metadata only. See
+`ARXIV_METADATA_NOTICE.md` for provenance and source terms.
+
+## Integrity
+
+From the dataset root, run:
+
+```bash
+sha256sum -c SHA256SUMS
+```
+
+## License
+
+This repository is marked `other` because it combines original RAP-derived
+artifacts with arXiv descriptive metadata. Read `LICENSE` and
+`ARXIV_METADATA_NOTICE.md` before redistribution or adaptation.
+
+## Citation
+
+```bibtex
+@misc{wu2026rap,
+  title  = {RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition Biases},
+  author = {Wu, Yingqian and Liang, Jingcong and Wang, Siyuan and Yin, Zhenfei
+            and Torr, Philip and Yu, Junchi and Wei, Zhongyu},
+  year   = {2026},
+  url    = {https://arxiv.org/abs/2609.10092}
+}
+```

LICENSE

@@ -0,0 +1,53 @@
+RAP Benchmark Dataset — Composite License Notice
+
+Copyright (c) 2026 Yingqian Wu, Jingcong Liang, Siyuan Wang, Zhenfei Yin,
+Philip Torr, Junchi Yu, and Zhongyu Wei.
+
+This dataset repository is marked `other` because it contains both original
+RAP-derived material and descriptive metadata obtained from arXiv. No single
+license label should be applied to every file without reading the notices
+below.
+
+## 1. Original RAP-derived material
+
+The original RAP-derived portions of this release—including the episode
+definitions, field and direction assignments, codebooks, split manifests,
+target tables, aggregate labels, and the authors' selection and arrangement of
+these materials—are licensed under the Creative Commons Attribution 4.0
+International license (CC BY 4.0):
+
+<https://creativecommons.org/licenses/by/4.0/>
+
+You may share and adapt those portions, including commercially, provided that
+you give appropriate credit, link to the license, and indicate whether changes
+were made. Please cite the RAP paper and link to this repository when using the
+release:
+
+<https://arxiv.org/abs/2609.10092>
+
+## 2. arXiv metadata in Search shards
+
+The files under `search_shards/` contain descriptive arXiv metadata only. They
+do not contain PDFs, TeX/source files, figures, or full-text article content.
+The arXiv API terms describe the treatment of metadata such as identifiers,
+titles, abstracts, dates, authors, and classification terms. The applicable
+source terms and attribution boundary are documented in:
+
+<https://info.arxiv.org/help/api/tou.html>
+
+<ARXIV_METADATA_NOTICE.md>
+
+This notice does not change or relicense arXiv's source terms. Each Search row
+preserves an arXiv identifier so that users can identify the source record.
+This independent research artifact is not endorsed by arXiv.
+
+## 3. Excluded materials
+
+This release does not redistribute PDFs, TeX sources, figures, full-text
+papers, provider responses, model checkpoints, or private human-audit records.
+
+## 4. No warranty
+
+The authors provide the original RAP-derived material and this compilation
+without warranty. For the complete terms of CC BY 4.0 and the applicable arXiv
+source terms, consult the linked licenses and notices above.

README.md

@@ -0,0 +1,91 @@
+---
+license: other
+license_name: rap-benchmark-composite-license
+license_link: LICENSE
+pretty_name: RAP Benchmark Dataset
+---
+
+# RAP Benchmark Dataset
+
+This is the companion data release for **Research Attention Prediction (RAP)**.
+RAP evaluates whether an agent can estimate the distribution of newly posted
+papers across eight fixed research directions, then forecast the distribution
+in the following six months.
+
+The package contains the frozen benchmark interface and the metadata-only
+Search snapshot. It does not contain model outputs, training data, provider
+responses, checkpoints, PDFs, TeX sources, or full-text papers.
+
+## Snapshot
+
+| Item | Count |
+| --- | ---: |
+| Fields | 278 |
+| Dev episodes | 335 |
+| Test episodes | 1,055 |
+| Fixed Search shards | 1,390 |
+| Expanding Search shards | 1,390 |
+| Directions per field | 8 |
+
+## Layout
+
+| Path | Use |
+| --- | --- |
+| `inputs/` | Safe model-facing Dev/Test episode inputs |
+| `targets/` | Evaluator-only recent/future counts, shares, and diagnostics |
+| `evaluator/` | Combined episodes for scoring and no-API baselines; do not send to agents |
+| `data/` | Public runner scope and Search-manifest indexes |
+| `search_shards/` | Fixed and Expanding metadata-only Search corpora |
+| `DATA_CARD.md` | Task semantics, leakage boundary, and aggregation rules |
+| `ARXIV_METADATA_NOTICE.md` | Provenance and redistribution boundary for arXiv metadata |
+
+Each row in `inputs/` contains the field, cutoff date, history-window metadata,
+future-window metadata, and the eight direction definitions. It intentionally
+does not contain history counts, recent shares, future counts, or future shares.
+Those evaluator-only values live in `targets/` and `evaluator/`.
+
+## Search snapshot
+
+- `search_shards/fixed/` enforces `[T-6m, T)`.
+- `search_shards/expanding/` enforces `<T`.
+- Each shard contains exactly `id`, `date`, `title`, `abstract`, and
+  `categories`.
+- Each manifest records the shard SHA-256, row count, and date-bound checks.
+
+The shards contain descriptive arXiv metadata only—not PDFs, TeX, figures, or
+full text. See `ARXIV_METADATA_NOTICE.md` for the source terms and attribution
+boundary. The Search snapshot is large; Hugging Face or Zenodo is preferable to
+ordinary GitHub blobs.
+
+## Integrity
+
+Verify the complete snapshot from its root:
+
+```bash
+sha256sum -c SHA256SUMS
+```
+
+The evaluator contract and scoring commands are documented in the companion
+RAP evaluation-code repository.
+
+## License
+
+This dataset repository is marked `other` because it combines two provenance
+classes. Original RAP-derived artifacts are released under CC BY 4.0. The
+metadata-only Search shards retain their arXiv provenance and source terms.
+Read `LICENSE` and `ARXIV_METADATA_NOTICE.md` before redistributing or adapting
+the snapshot.
+
+## Citation
+
+Please cite the RAP paper:
+
+```bibtex
+@misc{wu2026rap,
+  title  = {RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition Biases},
+  author = {Wu, Yingqian and Liang, Jingcong and Wang, Siyuan and Yin, Zhenfei
+            and Torr, Philip and Yu, Junchi and Wei, Zhongyu},
+  year   = {2026},
+  url    = {https://arxiv.org/abs/2609.10092}
+}
+```

data/MODEL_EVALUATION_EPISODES_V1_MANIFEST.json

@@ -0,0 +1,19 @@
+{
+  "protocol": "model-evaluation-episodes-v1",
+  "source": "[retained source path]",
+  "source_sha256": "e993c4f5a55d69fc8f89987762bfcd41963aa705e04c5a8266eaaea684fc9af0",
+  "scope": "[retained source path]",
+  "scope_sha256": "accac591086458edf9fec00518287f90a4847d6ebd73945f82e18d32c7537793",
+  "dev": {
+    "path": "[retained source path]",
+    "sha256": "2930d89885fa0667a9a8df12fd4e50a30480f2fcab0a2587c2b4a6219557df74",
+    "fields": 67,
+    "episodes": 335
+  },
+  "test": {
+    "path": "[retained source path]",
+    "sha256": "a3accfcae1152e3cb2440a977ec21a3d80bd9ad10b9e875302d90c530a394197",
+    "fields": 211,
+    "episodes": 1055
+  }
+}

data/MODEL_EVALUATION_SCOPE_FREEZE_V1.json

@@ -0,0 +1,2633 @@
+{
+  "protocol": "model-evaluation-scope-strong-block-v1",
+  "forecast_results_used": false,
+  "membership_cutoff": "2024-01-01",
+  "assignments": "[retained source path]",
+  "assignments_sha256": "2c3f97ccb9f79bba64d98bed7f7b619d8cd880e75a8602d9a3491c0da3964ae3",
+  "parallel_seed_source": "[retained source path]",
+  "parallel_seed_source_sha256": "2020eba215f1abc5b5d96cc0a38df64eb4b38eaa43ff6e9bbbe5292f7653d8e1",
+  "strong_block_rule": {
+    "overlap_coefficient_minimum": 0.6,
+    "minimum_intersection": 10,
+    "connected_components": true
+  },
+  "threshold_scan": {
+    "0.40": {
+      "edges": 192,
+      "blocks": 121,
+      "non_singleton_blocks": 47,
+      "fields_in_non_singletons": 204,
+      "largest_block": 40,
+      "top_block_sizes": [
+        40,
+        34,
+        10,
+        7,
+        6,
+        5,
+        4,
+        4,
+        4,
+        4
+      ]
+    },
+    "0.50": {
+      "edges": 156,
+      "blocks": 141,
+      "non_singleton_blocks": 53,
+      "fields_in_non_singletons": 190,
+      "largest_block": 21,
+      "top_block_sizes": [
+        21,
+        20,
+        9,
+        8,
+        7,
+        7,
+        6,
+        5,
+        4,
+        4
+      ]
+    },
+    "0.60": {
+      "edges": 113,
+      "blocks": 175,
+      "non_singleton_blocks": 53,
+      "fields_in_non_singletons": 156,
+      "largest_block": 13,
+      "top_block_sizes": [
+        13,
+        10,
+        9,
+        6,
+        5,
+        5,
+        4,
+        4,
+        3,
+        3
+      ]
+    },
+    "0.70": {
+      "edges": 85,
+      "blocks": 199,
+      "non_singleton_blocks": 46,
+      "fields_in_non_singletons": 125,
+      "largest_block": 9,
+      "top_block_sizes": [
+        9,
+        9,
+        8,
+        6,
+        4,
+        3,
+        3,
+        3,
+        3,
+        3
+      ]
+    },
+    "0.80": {
+      "edges": 51,
+      "blocks": 228,
+      "non_singleton_blocks": 33,
+      "fields_in_non_singletons": 83,
+      "largest_block": 8,
+      "top_block_sizes": [
+        8,
+        8,
+        4,
+        3,
+        3,
+        3,
+        2,
+        2,
+        2,
+        2
+      ]
+    }
+  },
+  "fields": 278,
+  "pre2024_unique_papers": 73587,
+  "pre2024_memberships": 138946,
+  "blocks": [
+    {
+      "block_id": 0,
+      "fields": [
+        "language model alignment",
+        "language model reasoning",
+        "large language model alignment",
+        "large language model efficiency",
+        "large language model evaluation",
+        "large language model fine-tuning",
+        "large language model inference",
+        "large language model inference efficiency",
+        "large language model reasoning",
+        "large language model safety",
+        "large language model systems",
+        "large language model training",
+        "large language models"
+      ],
+      "size": 13,
+      "split": "test"
+    },
+    {
+      "block_id": 1,
+      "fields": [
+        "image segmentation",
+        "medical image analysis",
+        "medical image classification",
+        "medical image object detection",
+        "medical image registration",
+        "medical image report generation",
+        "medical image segmentation",
+        "medical image synthesis",
+        "retinal image analysis",
+        "semantic segmentation"
+      ],
+      "size": 10,
+      "split": "test"
+    },
+    {
+      "block_id": 2,
+      "fields": [
+        "constrained reinforcement learning",
+        "model-based reinforcement learning",
+        "multi-agent reinforcement learning",
+        "multi-agent systems",
+        "offline reinforcement learning",
+        "reinforcement learning",
+        "reinforcement learning from human feedback",
+        "reinforcement learning theory",
+        "safe reinforcement learning"
+      ],
+      "size": 9,
+      "split": "dev"
+    },
+    {
+      "block_id": 3,
+      "fields": [
+        "explainable artificial intelligence",
+        "language model interpretability",
+        "large language model interpretability",
+        "mechanistic interpretability",
+        "model interpretability",
+        "neural network interpretability"
+      ],
+      "size": 6,
+      "split": "test"
+    },
+    {
+      "block_id": 4,
+      "fields": [
+        "affective computing",
+        "multimodal emotion recognition",
+        "multimodal large language models",
+        "multimodal learning",
+        "multimodal representation learning"
+      ],
+      "size": 5,
+      "split": "dev"
+    },
+    {
+      "block_id": 5,
+      "fields": [
+        "deep learning theory",
+        "neural network optimization theory",
+        "neural network theory",
+        "optimization theory",
+        "stochastic optimization"
+      ],
+      "size": 5,
+      "split": "dev"
+    },
+    {
+      "block_id": 6,
+      "fields": [
+        "anomaly detection",
+        "industrial anomaly detection",
+        "time series anomaly detection",
+        "video anomaly detection"
+      ],
+      "size": 4,
+      "split": "dev"
+    },
+    {
+      "block_id": 7,
+      "fields": [
+        "medical visual question answering",
+        "question answering",
+        "video question answering",
+        "visual question answering"
+      ],
+      "size": 4,
+      "split": "test"
+    },
+    {
+      "block_id": 8,
+      "fields": [
+        "3d object detection",
+        "object detection",
+        "remote sensing object detection"
+      ],
+      "size": 3,
+      "split": "dev"
+    },
+    {
+      "block_id": 9,
+      "fields": [
+        "adversarial machine learning",
+        "adversarial robustness",
+        "backdoor attacks"
+      ],
+      "size": 3,
+      "split": "dev"
+    },
+    {
+      "block_id": 10,
+      "fields": [
+        "causal discovery",
+        "causal inference",
+        "causal representation learning"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 11,
+      "fields": [
+        "clinical outcome prediction",
+        "clinical predictive modeling",
+        "clinical risk prediction"
+      ],
+      "size": 3,
+      "split": "dev"
+    },
+    {
+      "block_id": 12,
+      "fields": [
+        "dialogue systems",
+        "spoken dialogue systems",
+        "task-oriented dialogue systems"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 13,
+      "fields": [
+        "domain adaptation",
+        "test-time adaptation",
+        "unsupervised domain adaptation"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 14,
+      "fields": [
+        "model compression",
+        "neural network compression",
+        "neural network model compression"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 15,
+      "fields": [
+        "multi-armed bandits",
+        "online learning",
+        "online learning theory"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 16,
+      "fields": [
+        "physics-informed machine learning",
+        "physics-informed neural networks",
+        "scientific machine learning"
+      ],
+      "size": 3,
+      "split": "test"
+    },
+    {
+      "block_id": 17,
+      "fields": [
+        "robot learning",
+        "robot learning and control",
+        "robot manipulation"
+      ],
+      "size": 3,
+      "split": "dev"
+    },
+    {
+      "block_id": 18,
+      "fields": [
+        "3d human pose estimation",
+        "human pose estimation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 19,
+      "fields": [
+        "3d point cloud processing",
+        "lidar point cloud processing"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 20,
+      "fields": [
+        "3d reconstruction",
+        "3d scene reconstruction"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 21,
+      "fields": [
+        "3d shape modeling",
+        "text-to-3d generation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 22,
+      "fields": [
+        "action recognition",
+        "video action recognition"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 23,
+      "fields": [
+        "autonomous driving",
+        "autonomous driving perception"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 24,
+      "fields": [
+        "biomedical natural language processing",
+        "clinical natural language processing"
+      ],
+      "size": 2,
+      "split": "dev"
+    },
+    {
+      "block_id": 25,
+      "fields": [
+        "class-incremental learning",
+        "continual learning"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 26,
+      "fields": [
+        "code generation",
+        "large language model code generation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 27,
+      "fields": [
+        "combinatorial optimization",
+        "neural combinatorial optimization"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 28,
+      "fields": [
+        "computational learning theory",
+        "statistical learning theory"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 29,
+      "fields": [
+        "deep learning optimization",
+        "neural network optimization"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 30,
+      "fields": [
+        "depth estimation",
+        "monocular depth estimation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 31,
+      "fields": [
+        "differential privacy",
+        "privacy-preserving machine learning"
+      ],
+      "size": 2,
+      "split": "dev"
+    },
+    {
+      "block_id": 32,
+      "fields": [
+        "diffusion model image generation",
+        "diffusion models"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 33,
+      "fields": [
+        "fake news detection",
+        "misinformation detection"
+      ],
+      "size": 2,
+      "split": "dev"
+    },
+    {
+      "block_id": 34,
+      "fields": [
+        "financial time series forecasting",
+        "time series forecasting"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 35,
+      "fields": [
+        "graph neural networks",
+        "graph representation learning"
+      ],
+      "size": 2,
+      "split": "dev"
+    },
+    {
+      "block_id": 36,
+      "fields": [
+        "human-ai interaction",
+        "human-computer interaction"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 37,
+      "fields": [
+        "hyperspectral image analysis",
+        "hyperspectral image classification"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 38,
+      "fields": [
+        "image generation",
+        "text-to-image generation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 39,
+      "fields": [
+        "implicit neural representations",
+        "neural implicit representations"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 40,
+      "fields": [
+        "information extraction",
+        "relation extraction"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 41,
+      "fields": [
+        "knowledge graph completion",
+        "knowledge graph embedding"
+      ],
+      "size": 2,
+      "split": "dev"
+    },
+    {
+      "block_id": 42,
+      "fields": [
+        "language model agents",
+        "large language model agents"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 43,
+      "fields": [
+        "language model pretraining",
+        "large language model pretraining"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 44,
+      "fields": [
+        "large language model hallucination",
+        "large language model reliability"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 45,
+      "fields": [
+        "machine translation",
+        "neural machine translation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 46,
+      "fields": [
+        "medical image reconstruction",
+        "mri reconstruction"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 47,
+      "fields": [
+        "neural architecture search",
+        "neural network architecture design"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 48,
+      "fields": [
+        "out-of-distribution detection",
+        "out-of-distribution generalization"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 49,
+      "fields": [
+        "recommender systems",
+        "sequential recommendation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 50,
+      "fields": [
+        "speech synthesis",
+        "text-to-speech synthesis"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 51,
+      "fields": [
+        "text-to-video generation",
+        "video generation"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 52,
+      "fields": [
+        "video object segmentation",
+        "video understanding"
+      ],
+      "size": 2,
+      "split": "test"
+    },
+    {
+      "block_id": 53,
+      "fields": [
+        "3d scene generation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 54,
+      "fields": [
+        "3d scene understanding"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 55,
+      "fields": [
+        "active learning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 56,
+      "fields": [
+        "ai governance and policy"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 57,
+      "fields": [
+        "ai in education"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 58,
+      "fields": [
+        "ai-generated content detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 59,
+      "fields": [
+        "algorithmic fairness"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 60,
+      "fields": [
+        "audio-visual learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 61,
+      "fields": [
+        "automated machine learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 62,
+      "fields": [
+        "automatic speech recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 63,
+      "fields": [
+        "bayesian deep learning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 64,
+      "fields": [
+        "bayesian optimization"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 65,
+      "fields": [
+        "brain-computer interfaces"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 66,
+      "fields": [
+        "clinical decision support systems"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 67,
+      "fields": [
+        "clustering"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 68,
+      "fields": [
+        "computational linguistics"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 69,
+      "fields": [
+        "computational pathology"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 70,
+      "fields": [
+        "computational social science"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 71,
+      "fields": [
+        "conformal prediction"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 72,
+      "fields": [
+        "deepfake detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 73,
+      "fields": [
+        "distributed machine learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 74,
+      "fields": [
+        "document understanding"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 75,
+      "fields": [
+        "domain generalization"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 76,
+      "fields": [
+        "dynamic scene reconstruction"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 77,
+      "fields": [
+        "educational data mining"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 78,
+      "fields": [
+        "educational natural language processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 79,
+      "fields": [
+        "eeg signal processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 80,
+      "fields": [
+        "efficient deep learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 81,
+      "fields": [
+        "efficient transformers"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 82,
+      "fields": [
+        "equivariant neural networks"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 83,
+      "fields": [
+        "event-based vision"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 84,
+      "fields": [
+        "face recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 85,
+      "fields": [
+        "facial expression recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 86,
+      "fields": [
+        "federated learning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 87,
+      "fields": [
+        "few-shot learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 88,
+      "fields": [
+        "gaussian processes"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 89,
+      "fields": [
+        "generative adversarial networks"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 90,
+      "fields": [
+        "generative modeling"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 91,
+      "fields": [
+        "hate speech detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 92,
+      "fields": [
+        "human activity recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 93,
+      "fields": [
+        "human motion generation"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 94,
+      "fields": [
+        "human-robot interaction"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 95,
+      "fields": [
+        "image captioning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 96,
+      "fields": [
+        "image classification"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 97,
+      "fields": [
+        "image editing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 98,
+      "fields": [
+        "image quality assessment"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 99,
+      "fields": [
+        "image restoration"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 100,
+      "fields": [
+        "image retrieval"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 101,
+      "fields": [
+        "image super-resolution"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 102,
+      "fields": [
+        "imitation learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 103,
+      "fields": [
+        "in-context learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 104,
+      "fields": [
+        "industrial defect detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 105,
+      "fields": [
+        "information retrieval"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 106,
+      "fields": [
+        "knowledge distillation"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 107,
+      "fields": [
+        "knowledge graph construction"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 108,
+      "fields": [
+        "knowledge graph reasoning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 109,
+      "fields": [
+        "language model evaluation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 110,
+      "fields": [
+        "large language model adaptation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 111,
+      "fields": [
+        "large language model fairness"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 112,
+      "fields": [
+        "large language model prompting"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 113,
+      "fields": [
+        "learned image compression"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 114,
+      "fields": [
+        "legal natural language processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 115,
+      "fields": [
+        "low-resource natural language processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 116,
+      "fields": [
+        "machine unlearning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 117,
+      "fields": [
+        "malware detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 118,
+      "fields": [
+        "medical question answering"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 119,
+      "fields": [
+        "mental health nlp"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 120,
+      "fields": [
+        "meta-learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 121,
+      "fields": [
+        "mixture of experts"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 122,
+      "fields": [
+        "molecular generation"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 123,
+      "fields": [
+        "molecular representation learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 124,
+      "fields": [
+        "multi-object tracking"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 125,
+      "fields": [
+        "multi-task learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 126,
+      "fields": [
+        "multilingual language models"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 127,
+      "fields": [
+        "multilingual natural language processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 128,
+      "fields": [
+        "multimodal reasoning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 129,
+      "fields": [
+        "music generation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 130,
+      "fields": [
+        "music information retrieval"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 131,
+      "fields": [
+        "named entity recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 132,
+      "fields": [
+        "natural language generation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 133,
+      "fields": [
+        "network intrusion detection"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 134,
+      "fields": [
+        "neural network hardware acceleration"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 135,
+      "fields": [
+        "neural network verification"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 136,
+      "fields": [
+        "neural operator learning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 137,
+      "fields": [
+        "neural radiance fields"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 138,
+      "fields": [
+        "neuroimaging analysis"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 139,
+      "fields": [
+        "neuromorphic computing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 140,
+      "fields": [
+        "novel view synthesis"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 141,
+      "fields": [
+        "object pose estimation"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 142,
+      "fields": [
+        "parameter-efficient fine-tuning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 143,
+      "fields": [
+        "person re-identification"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 144,
+      "fields": [
+        "probabilistic graphical models"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 145,
+      "fields": [
+        "quantum machine learning"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 146,
+      "fields": [
+        "remote sensing image analysis"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 147,
+      "fields": [
+        "representation learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 148,
+      "fields": [
+        "retrieval-augmented generation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 149,
+      "fields": [
+        "robot motion planning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 150,
+      "fields": [
+        "self-supervised learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 151,
+      "fields": [
+        "semantic communications"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 152,
+      "fields": [
+        "semi-supervised learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 153,
+      "fields": [
+        "sentiment analysis"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 154,
+      "fields": [
+        "sign language processing"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 155,
+      "fields": [
+        "spatiotemporal forecasting"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 156,
+      "fields": [
+        "spiking neural networks"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 157,
+      "fields": [
+        "surgical scene understanding"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 158,
+      "fields": [
+        "tabular data learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 159,
+      "fields": [
+        "talking head generation"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 160,
+      "fields": [
+        "text classification"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 161,
+      "fields": [
+        "text generation"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 162,
+      "fields": [
+        "text summarization"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 163,
+      "fields": [
+        "text-to-sql"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 164,
+      "fields": [
+        "time series analysis"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 165,
+      "fields": [
+        "trajectory prediction"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 166,
+      "fields": [
+        "uncertainty quantification"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 167,
+      "fields": [
+        "variational inference"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 168,
+      "fields": [
+        "vision transformer"
+      ],
+      "size": 1,
+      "split": "dev"
+    },
+    {
+      "block_id": 169,
+      "fields": [
+        "vision-language models"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 170,
+      "fields": [
+        "visual object tracking"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 171,
+      "fields": [
+        "visual place recognition"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 172,
+      "fields": [
+        "weakly supervised learning"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 173,
+      "fields": [
+        "weather forecasting"
+      ],
+      "size": 1,
+      "split": "test"
+    },
+    {
+      "block_id": 174,
+      "fields": [
+        "zero-shot learning"
+      ],
+      "size": 1,
+      "split": "test"
+    }
+  ],
+  "field_to_block": {
+    "language model alignment": 0,
+    "language model reasoning": 0,
+    "large language model alignment": 0,
+    "large language model efficiency": 0,
+    "large language model evaluation": 0,
+    "large language model fine-tuning": 0,
+    "large language model inference": 0,
+    "large language model inference efficiency": 0,
+    "large language model reasoning": 0,
+    "large language model safety": 0,
+    "large language model systems": 0,
+    "large language model training": 0,
+    "large language models": 0,
+    "image segmentation": 1,
+    "medical image analysis": 1,
+    "medical image classification": 1,
+    "medical image object detection": 1,
+    "medical image registration": 1,
+    "medical image report generation": 1,
+    "medical image segmentation": 1,
+    "medical image synthesis": 1,
+    "retinal image analysis": 1,
+    "semantic segmentation": 1,
+    "constrained reinforcement learning": 2,
+    "model-based reinforcement learning": 2,
+    "multi-agent reinforcement learning": 2,
+    "multi-agent systems": 2,
+    "offline reinforcement learning": 2,
+    "reinforcement learning": 2,
+    "reinforcement learning from human feedback": 2,
+    "reinforcement learning theory": 2,
+    "safe reinforcement learning": 2,
+    "explainable artificial intelligence": 3,
+    "language model interpretability": 3,
+    "large language model interpretability": 3,
+    "mechanistic interpretability": 3,
+    "model interpretability": 3,
+    "neural network interpretability": 3,
+    "affective computing": 4,
+    "multimodal emotion recognition": 4,
+    "multimodal large language models": 4,
+    "multimodal learning": 4,
+    "multimodal representation learning": 4,
+    "deep learning theory": 5,
+    "neural network optimization theory": 5,
+    "neural network theory": 5,
+    "optimization theory": 5,
+    "stochastic optimization": 5,
+    "anomaly detection": 6,
+    "industrial anomaly detection": 6,
+    "time series anomaly detection": 6,
+    "video anomaly detection": 6,
+    "medical visual question answering": 7,
+    "question answering": 7,
+    "video question answering": 7,
+    "visual question answering": 7,
+    "3d object detection": 8,
+    "object detection": 8,
+    "remote sensing object detection": 8,
+    "adversarial machine learning": 9,
+    "adversarial robustness": 9,
+    "backdoor attacks": 9,
+    "causal discovery": 10,
+    "causal inference": 10,
+    "causal representation learning": 10,
+    "clinical outcome prediction": 11,
+    "clinical predictive modeling": 11,
+    "clinical risk prediction": 11,
+    "dialogue systems": 12,
+    "spoken dialogue systems": 12,
+    "task-oriented dialogue systems": 12,
+    "domain adaptation": 13,
+    "test-time adaptation": 13,
+    "unsupervised domain adaptation": 13,
+    "model compression": 14,
+    "neural network compression": 14,
+    "neural network model compression": 14,
+    "multi-armed bandits": 15,
+    "online learning": 15,
+    "online learning theory": 15,
+    "physics-informed machine learning": 16,
+    "physics-informed neural networks": 16,
+    "scientific machine learning": 16,
+    "robot learning": 17,
+    "robot learning and control": 17,
+    "robot manipulation": 17,
+    "3d human pose estimation": 18,
+    "human pose estimation": 18,
+    "3d point cloud processing": 19,
+    "lidar point cloud processing": 19,
+    "3d reconstruction": 20,
+    "3d scene reconstruction": 20,
+    "3d shape modeling": 21,
+    "text-to-3d generation": 21,
+    "action recognition": 22,
+    "video action recognition": 22,
+    "autonomous driving": 23,
+    "autonomous driving perception": 23,
+    "biomedical natural language processing": 24,
+    "clinical natural language processing": 24,
+    "class-incremental learning": 25,
+    "continual learning": 25,
+    "code generation": 26,
+    "large language model code generation": 26,
+    "combinatorial optimization": 27,
+    "neural combinatorial optimization": 27,
+    "computational learning theory": 28,
+    "statistical learning theory": 28,
+    "deep learning optimization": 29,
+    "neural network optimization": 29,
+    "depth estimation": 30,
+    "monocular depth estimation": 30,
+    "differential privacy": 31,
+    "privacy-preserving machine learning": 31,
+    "diffusion model image generation": 32,
+    "diffusion models": 32,
+    "fake news detection": 33,
+    "misinformation detection": 33,
+    "financial time series forecasting": 34,
+    "time series forecasting": 34,
+    "graph neural networks": 35,
+    "graph representation learning": 35,
+    "human-ai interaction": 36,
+    "human-computer interaction": 36,
+    "hyperspectral image analysis": 37,
+    "hyperspectral image classification": 37,
+    "image generation": 38,
+    "text-to-image generation": 38,
+    "implicit neural representations": 39,
+    "neural implicit representations": 39,
+    "information extraction": 40,
+    "relation extraction": 40,
+    "knowledge graph completion": 41,
+    "knowledge graph embedding": 41,
+    "language model agents": 42,
+    "large language model agents": 42,
+    "language model pretraining": 43,
+    "large language model pretraining": 43,
+    "large language model hallucination": 44,
+    "large language model reliability": 44,
+    "machine translation": 45,
+    "neural machine translation": 45,
+    "medical image reconstruction": 46,
+    "mri reconstruction": 46,
+    "neural architecture search": 47,
+    "neural network architecture design": 47,
+    "out-of-distribution detection": 48,
+    "out-of-distribution generalization": 48,
+    "recommender systems": 49,
+    "sequential recommendation": 49,
+    "speech synthesis": 50,
+    "text-to-speech synthesis": 50,
+    "text-to-video generation": 51,
+    "video generation": 51,
+    "video object segmentation": 52,
+    "video understanding": 52,
+    "3d scene generation": 53,
+    "3d scene understanding": 54,
+    "active learning": 55,
+    "ai governance and policy": 56,
+    "ai in education": 57,
+    "ai-generated content detection": 58,
+    "algorithmic fairness": 59,
+    "audio-visual learning": 60,
+    "automated machine learning": 61,
+    "automatic speech recognition": 62,
+    "bayesian deep learning": 63,
+    "bayesian optimization": 64,
+    "brain-computer interfaces": 65,
+    "clinical decision support systems": 66,
+    "clustering": 67,
+    "computational linguistics": 68,
+    "computational pathology": 69,
+    "computational social science": 70,
+    "conformal prediction": 71,
+    "deepfake detection": 72,
+    "distributed machine learning": 73,
+    "document understanding": 74,
+    "domain generalization": 75,
+    "dynamic scene reconstruction": 76,
+    "educational data mining": 77,
+    "educational natural language processing": 78,
+    "eeg signal processing": 79,
+    "efficient deep learning": 80,
+    "efficient transformers": 81,
+    "equivariant neural networks": 82,
+    "event-based vision": 83,
+    "face recognition": 84,
+    "facial expression recognition": 85,
+    "federated learning": 86,
+    "few-shot learning": 87,
+    "gaussian processes": 88,
+    "generative adversarial networks": 89,
+    "generative modeling": 90,
+    "hate speech detection": 91,
+    "human activity recognition": 92,
+    "human motion generation": 93,
+    "human-robot interaction": 94,
+    "image captioning": 95,
+    "image classification": 96,
+    "image editing": 97,
+    "image quality assessment": 98,
+    "image restoration": 99,
+    "image retrieval": 100,
+    "image super-resolution": 101,
+    "imitation learning": 102,
+    "in-context learning": 103,
+    "industrial defect detection": 104,
+    "information retrieval": 105,
+    "knowledge distillation": 106,
+    "knowledge graph construction": 107,
+    "knowledge graph reasoning": 108,
+    "language model evaluation": 109,
+    "large language model adaptation": 110,
+    "large language model fairness": 111,
+    "large language model prompting": 112,
+    "learned image compression": 113,
+    "legal natural language processing": 114,
+    "low-resource natural language processing": 115,
+    "machine unlearning": 116,
+    "malware detection": 117,
+    "medical question answering": 118,
+    "mental health nlp": 119,
+    "meta-learning": 120,
+    "mixture of experts": 121,
+    "molecular generation": 122,
+    "molecular representation learning": 123,
+    "multi-object tracking": 124,
+    "multi-task learning": 125,
+    "multilingual language models": 126,
+    "multilingual natural language processing": 127,
+    "multimodal reasoning": 128,
+    "music generation": 129,
+    "music information retrieval": 130,
+    "named entity recognition": 131,
+    "natural language generation": 132,
+    "network intrusion detection": 133,
+    "neural network hardware acceleration": 134,
+    "neural network verification": 135,
+    "neural operator learning": 136,
+    "neural radiance fields": 137,
+    "neuroimaging analysis": 138,
+    "neuromorphic computing": 139,
+    "novel view synthesis": 140,
+    "object pose estimation": 141,
+    "parameter-efficient fine-tuning": 142,
+    "person re-identification": 143,
+    "probabilistic graphical models": 144,
+    "quantum machine learning": 145,
+    "remote sensing image analysis": 146,
+    "representation learning": 147,
+    "retrieval-augmented generation": 148,
+    "robot motion planning": 149,
+    "self-supervised learning": 150,
+    "semantic communications": 151,
+    "semi-supervised learning": 152,
+    "sentiment analysis": 153,
+    "sign language processing": 154,
+    "spatiotemporal forecasting": 155,
+    "spiking neural networks": 156,
+    "surgical scene understanding": 157,
+    "tabular data learning": 158,
+    "talking head generation": 159,
+    "text classification": 160,
+    "text generation": 161,
+    "text summarization": 162,
+    "text-to-sql": 163,
+    "time series analysis": 164,
+    "trajectory prediction": 165,
+    "uncertainty quantification": 166,
+    "variational inference": 167,
+    "vision transformer": 168,
+    "vision-language models": 169,
+    "visual object tracking": 170,
+    "visual place recognition": 171,
+    "weakly supervised learning": 172,
+    "weather forecasting": 173,
+    "zero-shot learning": 174
+  },
+  "dev": {
+    "seed_fields": [
+      "3d object detection",
+      "3d scene understanding",
+      "active learning",
+      "adversarial robustness",
+      "bayesian deep learning",
+      "biomedical natural language processing",
+      "clinical predictive modeling",
+      "fake news detection",
+      "federated learning",
+      "graph neural networks",
+      "human motion generation",
+      "human-robot interaction",
+      "image restoration",
+      "image super-resolution",
+      "industrial anomaly detection",
+      "knowledge distillation",
+      "knowledge graph construction",
+      "knowledge graph embedding",
+      "large language model fairness",
+      "mental health nlp",
+      "molecular generation",
+      "multi-agent systems",
+      "multimodal large language models",
+      "multimodal reasoning",
+      "neural network optimization theory",
+      "neural operator learning",
+      "neuroimaging analysis",
+      "object detection",
+      "object pose estimation",
+      "privacy-preserving machine learning",
+      "quantum machine learning",
+      "robot learning",
+      "safe reinforcement learning",
+      "spatiotemporal forecasting",
+      "stochastic optimization",
+      "talking head generation",
+      "uncertainty quantification",
+      "vision transformer"
+    ],
+    "seed_count": 38,
+    "block_ids": [
+      2,
+      4,
+      5,
+      6,
+      8,
+      9,
+      11,
+      17,
+      24,
+      31,
+      33,
+      35,
+      41,
+      54,
+      55,
+      63,
+      86,
+      93,
+      94,
+      99,
+      101,
+      106,
+      107,
+      111,
+      119,
+      122,
+      128,
+      136,
+      138,
+      141,
+      145,
+      155,
+      159,
+      166,
+      168
+    ],
+    "fields": [
+      "3d object detection",
+      "3d scene understanding",
+      "active learning",
+      "adversarial machine learning",
+      "adversarial robustness",
+      "affective computing",
+      "anomaly detection",
+      "backdoor attacks",
+      "bayesian deep learning",
+      "biomedical natural language processing",
+      "clinical natural language processing",
+      "clinical outcome prediction",
+      "clinical predictive modeling",
+      "clinical risk prediction",
+      "constrained reinforcement learning",
+      "deep learning theory",
+      "differential privacy",
+      "fake news detection",
+      "federated learning",
+      "graph neural networks",
+      "graph representation learning",
+      "human motion generation",
+      "human-robot interaction",
+      "image restoration",
+      "image super-resolution",
+      "industrial anomaly detection",
+      "knowledge distillation",
+      "knowledge graph completion",
+      "knowledge graph construction",
+      "knowledge graph embedding",
+      "large language model fairness",
+      "mental health nlp",
+      "misinformation detection",
+      "model-based reinforcement learning",
+      "molecular generation",
+      "multi-agent reinforcement learning",
+      "multi-agent systems",
+      "multimodal emotion recognition",
+      "multimodal large language models",
+      "multimodal learning",
+      "multimodal reasoning",
+      "multimodal representation learning",
+      "neural network optimization theory",
+      "neural network theory",
+      "neural operator learning",
+      "neuroimaging analysis",
+      "object detection",
+      "object pose estimation",
+      "offline reinforcement learning",
+      "optimization theory",
+      "privacy-preserving machine learning",
+      "quantum machine learning",
+      "reinforcement learning",
+      "reinforcement learning from human feedback",
+      "reinforcement learning theory",
+      "remote sensing object detection",
+      "robot learning",
+      "robot learning and control",
+      "robot manipulation",
+      "safe reinforcement learning",
+      "spatiotemporal forecasting",
+      "stochastic optimization",
+      "talking head generation",
+      "time series anomaly detection",
+      "uncertainty quantification",
+      "video anomaly detection",
+      "vision transformer"
+    ],
+    "field_count": 67,
+    "episodes": 335,
+    "purpose": "prompt/analysis development and parallel-slate construction robustness"
+  },
+  "test": {
+    "fields": [
+      "3d human pose estimation",
+      "3d point cloud processing",
+      "3d reconstruction",
+      "3d scene generation",
+      "3d scene reconstruction",
+      "3d shape modeling",
+      "action recognition",
+      "ai governance and policy",
+      "ai in education",
+      "ai-generated content detection",
+      "algorithmic fairness",
+      "audio-visual learning",
+      "automated machine learning",
+      "automatic speech recognition",
+      "autonomous driving",
+      "autonomous driving perception",
+      "bayesian optimization",
+      "brain-computer interfaces",
+      "causal discovery",
+      "causal inference",
+      "causal representation learning",
+      "class-incremental learning",
+      "clinical decision support systems",
+      "clustering",
+      "code generation",
+      "combinatorial optimization",
+      "computational learning theory",
+      "computational linguistics",
+      "computational pathology",
+      "computational social science",
+      "conformal prediction",
+      "continual learning",
+      "deep learning optimization",
+      "deepfake detection",
+      "depth estimation",
+      "dialogue systems",
+      "diffusion model image generation",
+      "diffusion models",
+      "distributed machine learning",
+      "document understanding",
+      "domain adaptation",
+      "domain generalization",
+      "dynamic scene reconstruction",
+      "educational data mining",
+      "educational natural language processing",
+      "eeg signal processing",
+      "efficient deep learning",
+      "efficient transformers",
+      "equivariant neural networks",
+      "event-based vision",
+      "explainable artificial intelligence",
+      "face recognition",
+      "facial expression recognition",
+      "few-shot learning",
+      "financial time series forecasting",
+      "gaussian processes",
+      "generative adversarial networks",
+      "generative modeling",
+      "hate speech detection",
+      "human activity recognition",
+      "human pose estimation",
+      "human-ai interaction",
+      "human-computer interaction",
+      "hyperspectral image analysis",
+      "hyperspectral image classification",
+      "image captioning",
+      "image classification",
+      "image editing",
+      "image generation",
+      "image quality assessment",
+      "image retrieval",
+      "image segmentation",
+      "imitation learning",
+      "implicit neural representations",
+      "in-context learning",
+      "industrial defect detection",
+      "information extraction",
+      "information retrieval",
+      "knowledge graph reasoning",
+      "language model agents",
+      "language model alignment",
+      "language model evaluation",
+      "language model interpretability",
+      "language model pretraining",
+      "language model reasoning",
+      "large language model adaptation",
+      "large language model agents",
+      "large language model alignment",
+      "large language model code generation",
+      "large language model efficiency",
+      "large language model evaluation",
+      "large language model fine-tuning",
+      "large language model hallucination",
+      "large language model inference",
+      "large language model inference efficiency",
+      "large language model interpretability",
+      "large language model pretraining",
+      "large language model prompting",
+      "large language model reasoning",
+      "large language model reliability",
+      "large language model safety",
+      "large language model systems",
+      "large language model training",
+      "large language models",
+      "learned image compression",
+      "legal natural language processing",
+      "lidar point cloud processing",
+      "low-resource natural language processing",
+      "machine translation",
+      "machine unlearning",
+      "malware detection",
+      "mechanistic interpretability",
+      "medical image analysis",
+      "medical image classification",
+      "medical image object detection",
+      "medical image reconstruction",
+      "medical image registration",
+      "medical image report generation",
+      "medical image segmentation",
+      "medical image synthesis",
+      "medical question answering",
+      "medical visual question answering",
+      "meta-learning",
+      "mixture of experts",
+      "model compression",
+      "model interpretability",
+      "molecular representation learning",
+      "monocular depth estimation",
+      "mri reconstruction",
+      "multi-armed bandits",
+      "multi-object tracking",
+      "multi-task learning",
+      "multilingual language models",
+      "multilingual natural language processing",
+      "music generation",
+      "music information retrieval",
+      "named entity recognition",
+      "natural language generation",
+      "network intrusion detection",
+      "neural architecture search",
+      "neural combinatorial optimization",
+      "neural implicit representations",
+      "neural machine translation",
+      "neural network architecture design",
+      "neural network compression",
+      "neural network hardware acceleration",
+      "neural network interpretability",
+      "neural network model compression",
+      "neural network optimization",
+      "neural network verification",
+      "neural radiance fields",
+      "neuromorphic computing",
+      "novel view synthesis",
+      "online learning",
+      "online learning theory",
+      "out-of-distribution detection",
+      "out-of-distribution generalization",
+      "parameter-efficient fine-tuning",
+      "person re-identification",
+      "physics-informed machine learning",
+      "physics-informed neural networks",
+      "probabilistic graphical models",
+      "question answering",
+      "recommender systems",
+      "relation extraction",
+      "remote sensing image analysis",
+      "representation learning",
+      "retinal image analysis",
+      "retrieval-augmented generation",
+      "robot motion planning",
+      "scientific machine learning",
+      "self-supervised learning",
+      "semantic communications",
+      "semantic segmentation",
+      "semi-supervised learning",
+      "sentiment analysis",
+      "sequential recommendation",
+      "sign language processing",
+      "speech synthesis",
+      "spiking neural networks",
+      "spoken dialogue systems",
+      "statistical learning theory",
+      "surgical scene understanding",
+      "tabular data learning",
+      "task-oriented dialogue systems",
+      "test-time adaptation",
+      "text classification",
+      "text generation",
+      "text summarization",
+      "text-to-3d generation",
+      "text-to-image generation",
+      "text-to-speech synthesis",
+      "text-to-sql",
+      "text-to-video generation",
+      "time series analysis",
+      "time series forecasting",
+      "trajectory prediction",
+      "unsupervised domain adaptation",
+      "variational inference",
+      "video action recognition",
+      "video generation",
+      "video object segmentation",
+      "video question answering",
+      "video understanding",
+      "vision-language models",
+      "visual object tracking",
+      "visual place recognition",
+      "visual question answering",
+      "weakly supervised learning",
+      "weather forecasting",
+      "zero-shot learning"
+    ],
+    "field_count": 211,
+    "episodes": 1055,
+    "purpose": "confirmatory model evaluation; no model outputs viewed before freeze"
+  },
+  "leakage_audit": {
+    "strong_cross_edges": 0,
+    "all_nonzero_cross_edges": 2418,
+    "total_pair_intersection_mass": 96027,
+    "cross_pair_intersection_mass": 20436,
+    "cross_intersection_mass_fraction": 0.21281514574026056,
+    "top_cross_edges": [
+      {
+        "field_a": "industrial anomaly detection",
+        "field_b": "industrial defect detection",
+        "intersection": 111,
+        "overlap_coefficient": 0.5935828877005348,
+        "jaccard": 0.4220532319391635
+      },
+      {
+        "field_a": "medical image analysis",
+        "field_b": "neuroimaging analysis",
+        "intersection": 217,
+        "overlap_coefficient": 0.5755968169761273,
+        "jaccard": 0.0460330929147221
+      },
+      {
+        "field_a": "time series analysis",
+        "field_b": "time series anomaly detection",
+        "intersection": 101,
+        "overlap_coefficient": 0.5489130434782609,
+        "jaccard": 0.09730250481695568
+      },
+      {
+        "field_a": "clinical natural language processing",
+        "field_b": "medical question answering",
+        "intersection": 57,
+        "overlap_coefficient": 0.5480769230769231,
+        "jaccard": 0.08755760368663594
+      },
+      {
+        "field_a": "spatiotemporal forecasting",
+        "field_b": "weather forecasting",
+        "intersection": 71,
+        "overlap_coefficient": 0.5419847328244275,
+        "jaccard": 0.12456140350877193
+      },
+      {
+        "field_a": "neural network optimization",
+        "field_b": "neural network optimization theory",
+        "intersection": 210,
+        "overlap_coefficient": 0.5357142857142857,
+        "jaccard": 0.2900552486187845
+      },
+      {
+        "field_a": "knowledge graph completion",
+        "field_b": "knowledge graph reasoning",
+        "intersection": 145,
+        "overlap_coefficient": 0.5178571428571429,
+        "jaccard": 0.3251121076233184
+      },
+      {
+        "field_a": "imitation learning",
+        "field_b": "robot learning",
+        "intersection": 145,
+        "overlap_coefficient": 0.5123674911660777,
+        "jaccard": 0.10591672753834916
+      },
+      {
+        "field_a": "efficient transformers",
+        "field_b": "vision transformer",
+        "intersection": 147,
+        "overlap_coefficient": 0.5086505190311419,
+        "jaccard": 0.29458917835671344
+      },
+      {
+        "field_a": "algorithmic fairness",
+        "field_b": "large language model fairness",
+        "intersection": 91,
+        "overlap_coefficient": 0.5055555555555555,
+        "jaccard": 0.07564422277639235
+      },
+      {
+        "field_a": "remote sensing image analysis",
+        "field_b": "remote sensing object detection",
+        "intersection": 84,
+        "overlap_coefficient": 0.5029940119760479,
+        "jaccard": 0.08870116156282999
+      },
+      {
+        "field_a": "neural operator learning",
+        "field_b": "scientific machine learning",
+        "intersection": 77,
+        "overlap_coefficient": 0.46107784431137727,
+        "jaccard": 0.05873379099923722
+      },
+      {
+        "field_a": "medical image object detection",
+        "field_b": "object detection",
+        "intersection": 65,
+        "overlap_coefficient": 0.46099290780141844,
+        "jaccard": 0.036931818181818184
+      },
+      {
+        "field_a": "large language models",
+        "field_b": "multimodal large language models",
+        "intersection": 163,
+        "overlap_coefficient": 0.4604519774011299,
+        "jaccard": 0.05077881619937695
+      },
+      {
+        "field_a": "knowledge graph embedding",
+        "field_b": "knowledge graph reasoning",
+        "intersection": 124,
+        "overlap_coefficient": 0.45925925925925926,
+        "jaccard": 0.2713347921225383
+      },
+      {
+        "field_a": "large language model alignment",
+        "field_b": "reinforcement learning from human feedback",
+        "intersection": 61,
+        "overlap_coefficient": 0.4485294117647059,
+        "jaccard": 0.1564102564102564
+      },
+      {
+        "field_a": "affective computing",
+        "field_b": "facial expression recognition",
+        "intersection": 75,
+        "overlap_coefficient": 0.43859649122807015,
+        "jaccard": 0.10822510822510822
+      },
+      {
+        "field_a": "clinical decision support systems",
+        "field_b": "clinical predictive modeling",
+        "intersection": 79,
+        "overlap_coefficient": 0.42934782608695654,
+        "jaccard": 0.13787085514834205
+      },
+      {
+        "field_a": "deep learning optimization",
+        "field_b": "optimization theory",
+        "intersection": 188,
+        "overlap_coefficient": 0.41228070175438597,
+        "jaccard": 0.15524360033030554
+      },
+      {
+        "field_a": "neural network optimization",
+        "field_b": "optimization theory",
+        "intersection": 223,
+        "overlap_coefficient": 0.4114391143911439,
+        "jaccard": 0.17670364500792393
+      },
+      {
+        "field_a": "language model alignment",
+        "field_b": "reinforcement learning from human feedback",
+        "intersection": 54,
+        "overlap_coefficient": 0.39705882352941174,
+        "jaccard": 0.20224719101123595
+      },
+      {
+        "field_a": "knowledge graph construction",
+        "field_b": "knowledge graph reasoning",
+        "intersection": 94,
+        "overlap_coefficient": 0.3900414937759336,
+        "jaccard": 0.2052401746724891
+      },
+      {
+        "field_a": "spatiotemporal forecasting",
+        "field_b": "time series forecasting",
+        "intersection": 196,
+        "overlap_coefficient": 0.3843137254901961,
+        "jaccard": 0.17314487632508835
+      },
+      {
+        "field_a": "multimodal large language models",
+        "field_b": "vision-language models",
+        "intersection": 132,
+        "overlap_coefficient": 0.3728813559322034,
+        "jaccard": 0.09992429977289932
+      },
+      {
+        "field_a": "deep learning optimization",
+        "field_b": "neural network optimization theory",
+        "intersection": 145,
+        "overlap_coefficient": 0.36989795918367346,
+        "jaccard": 0.20625889046941678
+      },
+      {
+        "field_a": "human-computer interaction",
+        "field_b": "human-robot interaction",
+        "intersection": 90,
+        "overlap_coefficient": 0.36885245901639346,
+        "jaccard": 0.12448132780082988
+      },
+      {
+        "field_a": "distributed machine learning",
+        "field_b": "federated learning",
+        "intersection": 132,
+        "overlap_coefficient": 0.36464088397790057,
+        "jaccard": 0.056338028169014086
+      },
+      {
+        "field_a": "anomaly detection",
+        "field_b": "industrial defect detection",
+        "intersection": 67,
+        "overlap_coefficient": 0.3582887700534759,
+        "jaccard": 0.05861767279090114
+      },
+      {
+        "field_a": "robot learning",
+        "field_b": "robot motion planning",
+        "intersection": 129,
+        "overlap_coefficient": 0.3543956043956044,
+        "jaccard": 0.08799454297407913
+      },
+      {
+        "field_a": "computational social science",
+        "field_b": "misinformation detection",
+        "intersection": 84,
+        "overlap_coefficient": 0.3514644351464435,
+        "jaccard": 0.10579345088161209
+      },
+      {
+        "field_a": "clinical decision support systems",
+        "field_b": "clinical risk prediction",
+        "intersection": 64,
+        "overlap_coefficient": 0.34782608695652173,
+        "jaccard": 0.1285140562248996
+      },
+      {
+        "field_a": "clinical decision support systems",
+        "field_b": "clinical outcome prediction",
+        "intersection": 64,
+        "overlap_coefficient": 0.34782608695652173,
+        "jaccard": 0.11657559198542805
+      },
+      {
+        "field_a": "multimodal representation learning",
+        "field_b": "representation learning",
+        "intersection": 151,
+        "overlap_coefficient": 0.3463302752293578,
+        "jaccard": 0.06336550566512798
+      },
+      {
+        "field_a": "computational social science",
+        "field_b": "fake news detection",
+        "intersection": 58,
+        "overlap_coefficient": 0.34523809523809523,
+        "jaccard": 0.0774365821094793
+      },
+      {
+        "field_a": "deep learning theory",
+        "field_b": "neural network optimization",
+        "intersection": 184,
+        "overlap_coefficient": 0.33948339483394835,
+        "jaccard": 0.14556962025316456
+      },
+      {
+        "field_a": "deep learning optimization",
+        "field_b": "stochastic optimization",
+        "intersection": 99,
+        "overlap_coefficient": 0.3289036544850498,
+        "jaccard": 0.15045592705167174
+      },
+      {
+        "field_a": "3d object detection",
+        "field_b": "lidar point cloud processing",
+        "intersection": 163,
+        "overlap_coefficient": 0.3247011952191235,
+        "jaccard": 0.19019836639439908
+      },
+      {
+        "field_a": "conformal prediction",
+        "field_b": "uncertainty quantification",
+        "intersection": 38,
+        "overlap_coefficient": 0.3140495867768595,
+        "jaccard": 0.06365159128978225
+      },
+      {
+        "field_a": "large language model fairness",
+        "field_b": "large language models",
+        "intersection": 55,
+        "overlap_coefficient": 0.3055555555555556,
+        "jaccard": 0.017493638676844784
+      },
+      {
+        "field_a": "biomedical natural language processing",
+        "field_b": "medical question answering",
+        "intersection": 31,
+        "overlap_coefficient": 0.2980769230769231,
+        "jaccard": 0.05410122164048865
+      },
+      {
+        "field_a": "imitation learning",
+        "field_b": "robot learning and control",
+        "intersection": 84,
+        "overlap_coefficient": 0.2968197879858657,
+        "jaccard": 0.08341608738828203
+      },
+      {
+        "field_a": "graph representation learning",
+        "field_b": "representation learning",
+        "intersection": 213,
+        "overlap_coefficient": 0.29542302357836336,
+        "jaccard": 0.0817344589409056
+      },
+      {
+        "field_a": "human-ai interaction",
+        "field_b": "human-robot interaction",
+        "intersection": 71,
+        "overlap_coefficient": 0.29098360655737704,
+        "jaccard": 0.11563517915309446
+      },
+      {
+        "field_a": "large language model evaluation",
+        "field_b": "large language model fairness",
+        "intersection": 52,
+        "overlap_coefficient": 0.28888888888888886,
+        "jaccard": 0.04529616724738676
+      },
+      {
+        "field_a": "clinical natural language processing",
+        "field_b": "medical image report generation",
+        "intersection": 23,
+        "overlap_coefficient": 0.2875,
+        "jaccard": 0.03479576399394856
+      },
+      {
+        "field_a": "spatiotemporal forecasting",
+        "field_b": "trajectory prediction",
+        "intersection": 93,
+        "overlap_coefficient": 0.28615384615384615,
+        "jaccard": 0.12533692722371967
+      },
+      {
+        "field_a": "lidar point cloud processing",
+        "field_b": "object detection",
+        "intersection": 147,
+        "overlap_coefficient": 0.28378378378378377,
+        "jaccard": 0.07153284671532846
+      },
+      {
+        "field_a": "anomaly detection",
+        "field_b": "out-of-distribution detection",
+        "intersection": 111,
+        "overlap_coefficient": 0.2803030303030303,
+        "jaccard": 0.08486238532110092
+      },
+      {
+        "field_a": "video anomaly detection",
+        "field_b": "video understanding",
+        "intersection": 31,
+        "overlap_coefficient": 0.27927927927927926,
+        "jaccard": 0.029779058597502402
+      },
+      {
+        "field_a": "large language model fairness",
+        "field_b": "large language model safety",
+        "intersection": 50,
+        "overlap_coefficient": 0.2777777777777778,
+        "jaccard": 0.0880281690140845
+      }
+    ],
+    "interpretation": "Strong dependencies are fully grouped. Weaker legitimate hierarchy/partial overlap remains and is reported rather than deleted."
+  },
+  "inference": {
+    "primary_ci": "field-clustered bootstrap",
+    "robustness_ci": "resample frozen strong-dependency blocks"
+  }
+}

data/MODEL_EVALUATION_SCOPE_FREEZE_V1_MANIFEST.json

@@ -0,0 +1,15 @@
+{
+  "protocol": "model-evaluation-scope-strong-block-v1-manifest",
+  "frozen_at_utc": "2026-07-19T05:47:00Z",
+  "frozen_before_rebuilt_slate_model_test": true,
+  "selection_uses_model_outputs": false,
+  "builder": "freeze_model_evaluation_scope_v1.py",
+  "builder_sha256": "12e82066dd1784458aaa6649223565f0caed72f911b7a0e02fb249d5d211aae1",
+  "scope": "MODEL_EVALUATION_SCOPE_FREEZE_V1.json",
+  "scope_sha256": "accac591086458edf9fec00518287f90a4847d6ebd73945f82e18d32c7537793",
+  "strong_block_rule": "pre-2024 overlap coefficient>=0.60 and intersection>=10",
+  "dev": "38 outcome-blind parallel-instrument seeds plus complete strong-block closure; 67 fields/335 episodes",
+  "test": "all remaining complete blocks; 211 fields/1055 episodes",
+  "training_split": "none",
+  "amendment_policy": "Any change after model Test begins requires a new version and both scopes must be reported."
+}

evaluator/MODEL_TEST211_EPISODES_V1_WITH_TARGETS.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3accfcae1152e3cb2440a977ec21a3d80bd9ad10b9e875302d90c530a394197
+size 11984720

evaluator/README.md

@@ -0,0 +1,11 @@
+# Evaluator-only combined episodes
+
+这两个 JSONL 将每条 episode 的可见输入和未来 targets 放在同一行，方便代码
+仓库中的统计基线和评测脚本读取：
+
+- `MODEL_DEV67_EPISODES_V1_WITH_TARGETS.jsonl`
+- `MODEL_TEST211_EPISODES_V1_WITH_TARGETS.jsonl`
+
+它们**不是**模型输入文件。给 agent 的数据请使用上一级 `inputs/`；只有评测器
+在模型运行结束后才读取 `targets/`（或这里的合并副本）。这样可以避免把
+`future_counts`、`future_share`、`rank_change` 等未来标签泄漏进 prompt。

inputs/MODEL_TEST_EPISODES_INPUTS_V1.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00a8950062e22a8238b3b6996cd5983d95ba765f9e49652efcdc06eb6f2eab74
+size 10723710

search_shards/README.md

@@ -0,0 +1,15 @@
+# Search shard snapshot
+
+This directory is a metadata-only snapshot for the RAP Search tool. Every JSONL
+record has exactly `id`, `date`, `title`, `abstract`, and `categories`; no PDF,
+TeX source, figure, or full-text content is present.
+
+- `fixed/`: 1,390 shards, one per episode, with the enforced window `[T-6m, T)`.
+- `expanding/`: 1,390 shards, one per episode, with the enforced window `<T`.
+- `fixed/manifest.json` and `expanding/manifest.json`: public manifests with
+  per-shard SHA-256, record counts, and date-bound checks.
+
+The two `data/SEARCH_MANIFEST_*_INDEX_V1.json` files are convenient indexes for
+the runner in the sibling evaluation repository. The source and licensing
+boundary for the arXiv descriptive metadata is documented in
+`../ARXIV_METADATA_NOTICE.md`.

search_shards/expanding/F002__2026-01-01.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fc74e2f65bf2e6423b35aef483b21d20cee9251f4a37403dd78627092038d4f
+size 11179455