@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.076_aanet.yaml

@@ -0,0 +1,169 @@
+id: aisb.t3.076_aanet
+name: 'AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation'
+version: 0.1.0
+one_line: Two-stage alignment-and-aggregation framework for uncertainty-aware pocket-based
+  virtual screening on holo, apo, and AlphaFold2-predicted protein structures.
+task_description: 'This packaged benchmark covers structure-based virtual screening
+  under structural uncertainty. AANet addresses the challenge of scoring or ranking
+  candidate binders when reliable pocket annotations are unavailable, as occurs with
+  apo or predicted protein structures. The method uses a two-phase training strategy:
+  (1) alignment via tri-modal contrastive learning that aligns ligand, holo pocket,
+  and Fpocket-detected cavity representations with hard negative mining; and (2) aggregation
+  via cross-attention adapter for dynamic weighting of candidate binding sites, enabling
+  training on pocket-agnostic activity data from ChEMBL. Evaluation benchmarks include
+  DUD-E (38 targets) with holo, AlphaFold2-predicted, and experimental apo structures,
+  plus LIT-PCBA (12 targets).
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1-2d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- bedroc
+- virtual_screening_scores
+- screening_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged DUD-E or LIT-PCBA evaluation route using the released
+    checkpoint. Downloads the pretrained model checkpoint from Google Drive, places
+    it in the designated savedir subdirectory, then evaluates against the appropriate
+    benchmark. Suitable for reproducing published screening scores without retraining.
+
+    '
+- id: two_phase_train
+  label: Two-Phase Train
+  description: 'Execute the full AANet training workflow: Phase 1 (alignment) via
+    fpocket_neg_10A_siglip.sh for tri-modal contrastive representation learning, followed
+    by Phase 2 (aggregation) via finetune_chembl_fpocket_neg_10A_siglip_icrossatt_mollinear_wocollision_attn_kl.sh
+    for cross-attention-based pocket aggregation training on ChEMBL activity data.
+    Requires external pretrained Uni-Mol checkpoints and dataset preparation.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: benchmark_archive
+    url: https://deepscientist.cc/AISB/076_aanet
+    type: zip_archive
+    description: Main packaged benchmark containing code, configs, and evaluation
+      scripts
+  - name: model_checkpoint
+    url: https://drive.google.com/file/d/11Hixd7vVKg6RZcZ81LEXKWoc68p61csV
+    type: google_drive
+    description: Trained AANet model checkpoint, must be placed in savedir/drugclip_fpocket_neg_10A_siglip/[timestamp]/checkpoint_best.pt
+  - name: dude_apo_dataset
+    description: 38-target DUD-E subset with holo, AlphaFold2, and experimental apo
+      structures, bundled in dataset/dude_apo/
+  - name: lit_pcba_dataset
+    description: 12-target LIT-PCBA pocket data subset, bundled in dataset/lit_pcba/
+  - name: molecule_data
+    url: https://drive.google.com/drive/folders/1zW1MGpgunynFxTKXC2Q4RgWxZmg6CInV
+    type: google_drive
+    description: Molecule embeddings from DrugCLIP, decompress into dataset/
+  - name: unimol_pretrained
+    description: Uni-Mol pretrained weights (mol_pre_no_h_220816.pt, pocket_pre_220816.pt),
+      required for training
+  - name: chembl_training_data
+    description: ChEMBL35 assay-filtered ligand similarity data (chembl35_assay2_filtered_ligsim_ge5_dedup.lmdb)
+  notes:
+  - Model checkpoint requires external Google Drive access; no auth needed
+  - DUD-E and LIT-PCBA are included in benchmark archive
+  - Molecule data and Uni-Mol pretrained weights must be obtained externally
+  - Training requires Phase 1 alignment checkpoint before Phase 2 aggregation
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All datasets are publicly available
+  - Google Drive download requires no authentication for public sharing links
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 4
+    gpu_vram_gb: 48
+  notes:
+  - Multi-GPU (n_gpu=4) recommended for training phases; single-GPU evaluation supported
+  - Training batch_size=32, validation batch_size=128; adjust based on available VRAM
+  - Disk space accounts for datasets, pretrained models, and checkpoint saves during
+    training
+environment:
+  python: '3.9'
+  cuda: '11.3'
+  pytorch: 1.11.0
+  flash_attn: null
+  key_packages:
+  - rdkit-pypi==2022.9.3
+  - zstandard
+  - unimol
+  - unicore
+  notes:
+  - Core dependencies mirror Uni-Mol requirements; install via UniCore Docker (dptechnology/unicore:0.0.1-pytorch1.11.0-cuda11.3)
+  - Additional pip install zstandard rdkit-pypi==2022.9.3 after Docker activation
+  - Compatible conda environment: unimol
+risk_flags:
+- multi_phase_training_dependency
+- external_data_dependency
+- gpu_memory_sensitivity
+risk_notes:
+- Phase 1 alignment checkpoint (fpocket_neg_10A_siglip) must complete before Phase
+  2 aggregation (finetune_chembl_*) can start; misalignment will cause training failure
+- External dependencies on Google Drive for model checkpoint and DrugCLIP molecules
+  introduce availability and versioning risk
+- Batch sizes (32 train / 128 valid) may require adjustment for GPUs with less than
+  24GB VRAM; OOM errors common on smaller cards
+- Training with recycling=3 and dist_threshold=8.0 increases memory footprint; monitor
+  NCCL communication if using multi-GPU
+recommended_when: 'Use this benchmark when developing or evaluating structure-based
+  virtual screening methods that must handle structural uncertainty, including cases
+  where only apo or AlphaFold2-predicted protein structures are available without
+  ligand-bound pocket annotations. Ideal for testing robustness to pocket mislocalization,
+  evaluating early enrichment (EF 1%, BEDROC), or benchmarking against DUD-E/LIT-PCBA
+  in blind-apo settings.
+
+  '
+not_recommended_when: 'Do not use this benchmark if your workflow cannot handle molecular
+  3D structure preprocessing, protein pocket detection (Fpocket), or large-scale ligand
+  datasets. Not suitable for environments without GPU access or where external data
+  downloads are restricted.
+
+  '
+paper:
+  title: 'AANet: Virtual Screening under Structural Uncertainty via Alignment and
+    Aggregation'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=TUh4GDposM
+  github: https://github.com/Wiley-Z/AANet
+  license: Apache 2.0
+  model_weights_license: CC BY-NC 4.0
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.076_aanet.zip
+  archive_type: zip
+  local_dir_name: paper-76-AANet
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.076_aanet.zip
+  sha256: 821275c509de0ed669874cf74aaa284ee5623cbb15fff595b693eaeb25393e5b
+  size_bytes: 45967
+display:
+  palette_seed: emerald-bone-pocket
+  art_style: molecular-discovery
+  accent_priority: high
+image_path: ../image/076_aisb.t3.076_aanet.jpg

package/AISB/catalog/aisb.t3.076_aanet.zh.yaml

@@ -0,0 +1,129 @@
+id: aisb.t3.076_aanet
+name: 'AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation'
+version: 0.1.0
+one_line: 面向结构不确定性的两阶段“对齐-聚合”虚拟筛选框架，适用于 holo、apo 与 AlphaFold2 预测蛋白结构。
+task_description: 该基准覆盖在结构不确定性下的结构生物学虚拟筛选。AANet 解决在缺少可靠口袋标注（如 apo 或预测结构）时对候选配体打分/排序的问题。方法采用两阶段训练：(1) 对齐阶段通过三模态对比学习对齐配体、holo 口袋与 Fpocket 腔体表示，并进行难负样本挖掘；(2) 聚合阶段通过交叉注意力适配器动态加权候选结合位点，从而在 ChEMBL 的口袋无关活性数据上训练。评测包括 DUD-E（38 个靶点，含 holo、AlphaFold2、实验 apo）与 LIT-PCBA（12 个靶点）。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1-2d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- bedroc
+- virtual_screening_scores
+- screening_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged DUD-E or LIT-PCBA evaluation route using the released checkpoint. Downloads the pretrained model checkpoint from Google Drive, places it in the designated savedir subdirectory, then evaluates against the appropriate benchmark. Suitable for reproducing published screening scores without retraining.
+
+    '
+- id: two_phase_train
+  label: Two-Phase Train
+  description: 'Execute the full AANet training workflow: Phase 1 (alignment) via fpocket_neg_10A_siglip.sh for tri-modal contrastive representation learning, followed by Phase 2 (aggregation) via finetune_chembl_fpocket_neg_10A_siglip_icrossatt_mollinear_wocollision_attn_kl.sh for cross-attention-based pocket aggregation training on ChEMBL activity data. Requires external pretrained Uni-Mol checkpoints and dataset preparation.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: benchmark_archive
+    url: https://deepscientist.cc/AISB/076_aanet
+    type: zip_archive
+    description: Main packaged benchmark containing code, configs, and evaluation scripts
+  - name: model_checkpoint
+    url: https://drive.google.com/file/d/11Hixd7vVKg6RZcZ81LEXKWoc68p61csV
+    type: google_drive
+    description: Trained AANet model checkpoint, must be placed in savedir/drugclip_fpocket_neg_10A_siglip/[timestamp]/checkpoint_best.pt
+  - name: dude_apo_dataset
+    description: 38-target DUD-E subset with holo, AlphaFold2, and experimental apo structures, bundled in dataset/dude_apo/
+  - name: lit_pcba_dataset
+    description: 12-target LIT-PCBA pocket data subset, bundled in dataset/lit_pcba/
+  - name: molecule_data
+    url: https://drive.google.com/drive/folders/1zW1MGpgunynFxTKXC2Q4RgWxZmg6CInV
+    type: google_drive
+    description: Molecule embeddings from DrugCLIP, decompress into dataset/
+  - name: unimol_pretrained
+    description: Uni-Mol pretrained weights (mol_pre_no_h_220816.pt, pocket_pre_220816.pt), required for training
+  - name: chembl_training_data
+    description: ChEMBL35 assay-filtered ligand similarity data (chembl35_assay2_filtered_ligsim_ge5_dedup.lmdb)
+  notes:
+  - Model checkpoint requires external Google Drive access; no auth needed
+  - DUD-E and LIT-PCBA are included in benchmark archive
+  - Molecule data and Uni-Mol pretrained weights must be obtained externally
+  - Training requires Phase 1 alignment checkpoint before Phase 2 aggregation
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All datasets are publicly available
+  - Google Drive download requires no authentication for public sharing links
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 4
+    gpu_vram_gb: 48
+  notes:
+  - Multi-GPU (n_gpu=4) recommended for training phases; single-GPU evaluation supported
+  - Training batch_size=32, validation batch_size=128; adjust based on available VRAM
+  - Disk space accounts for datasets, pretrained models, and checkpoint saves during training
+environment:
+  python: '3.9'
+  cuda: '11.3'
+  pytorch: 1.11.0
+  flash_attn: null
+  key_packages:
+  - rdkit-pypi==2022.9.3
+  - zstandard
+  - unimol
+  - unicore
+  notes:
+  - Core dependencies mirror Uni-Mol requirements; install via UniCore Docker (dptechnology/unicore:0.0.1-pytorch1.11.0-cuda11.3)
+  - Additional pip install zstandard rdkit-pypi==2022.9.3 after Docker activation
+  - Compatible conda environment: unimol
+risk_flags:
+- multi_phase_training_dependency
+- external_data_dependency
+- gpu_memory_sensitivity
+risk_notes:
+- Phase 1 alignment checkpoint (fpocket_neg_10A_siglip) must complete before Phase 2 aggregation (finetune_chembl_*) can start; misalignment will cause training failure
+- External dependencies on Google Drive for model checkpoint and DrugCLIP molecules introduce availability and versioning risk
+- Batch sizes (32 train / 128 valid) may require adjustment for GPUs with less than 24GB VRAM; OOM errors common on smaller cards
+- Training with recycling=3 and dist_threshold=8.0 increases memory footprint; monitor NCCL communication if using multi-GPU
+recommended_when: 当你要开发或评估能处理结构不确定性的结构基础虚拟筛选方法时使用本基准，尤其适用于只有 apo/AlphaFold2 结构且缺少配体结合口袋标注的场景。可用于测试口袋定位偏差鲁棒性、早期富集指标（EF 1%、BEDROC）与 DUD-E/LIT-PCBA 盲 apo 设定对比。
+not_recommended_when: 若你的流程无法处理分子 3D 预处理、Fpocket 口袋检测或大规模配体数据集，不建议使用。无 GPU 或受限外部下载环境也不适合。
+paper:
+  title: 'AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=TUh4GDposM
+  github: https://github.com/Wiley-Z/AANet
+  license: Apache 2.0
+  model_weights_license: CC BY-NC 4.0
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.076_aanet.zip
+  archive_type: zip
+  local_dir_name: paper-76-AANet
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.076_aanet.zip
+  sha256: 821275c509de0ed669874cf74aaa284ee5623cbb15fff595b693eaeb25393e5b
+  size_bytes: 45967
+display:
+  palette_seed: emerald-bone-pocket
+  art_style: molecular-discovery
+  accent_priority: high
+image_path: ../image/076_aisb.t3.076_aanet.jpg

package/AISB/catalog/aisb.t3.077_cmnn.yaml

@@ -0,0 +1,199 @@
+id: aisb.t3.077_cmnn
+name: 'Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation
+  Beyond Bounded Activations'
+version: 0.1.0
+one_line: Constrained monotonic neural network theory and tabular modeling evaluation
+  requiring upstream source restoration before execution.
+task_description: 'This benchmark entry covers constrained monotonic neural networks
+  for tabular modeling under monotonicity constraints. The theoretical contribution
+  demonstrates that MLPs with non-negative weight constraints and alternating saturating
+  activations achieve universal approximation for monotonic functions, extending beyond
+  traditional bounded-activation approaches. Additionally, the work proves that MLPs
+  with non-positive constrained weights and ReLU activation also qualify as universal
+  approximators with at least 4 layers. The proposed parametrization eliminates weight
+  reparameterization requirements, improving training stability. The current local
+  snapshot preserves only metadata, audit files, and documentation rather than executable
+  training or evaluation code; upstream source restoration is required before this
+  benchmark becomes runnable.
+
+  '
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- id: test_rmse
+  description: Root mean squared error on held-out test set
+  status: blocked_no_code
+  note: No executable code anchor in current snapshot; requires restoration
+  code_anchor: null
+- id: metadata_review
+  description: Preserved benchmark framing and monotonic model documentation
+  status: available
+capability_tags:
+- research_code_optimization
+- monotonic_networks
+- tabular_ml
+- constrained_learning
+- trustworthy_ml
+- universal_approximation
+- hard_monotonicity
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Review the preserved metadata, paper documentation, and monotonic-model
+    benchmark framing without execution.
+  estimated_time: <1h
+  resource_intensive: false
+- id: restore_first
+  label: Restore Then Execute
+  description: Restore the missing constrained monotonic model implementation and
+    dataset setup before attempting benchmark execution.
+  estimated_time: 1-2h
+  resource_intensive: false
+  prerequisites:
+  - source_code_restoration
+  - dataset_acquisition
+  - dependency_verification
+dataset_download:
+  primary_method: pending_restoration
+  sources: []
+  notes:
+  - The current local package does not preserve executable benchmark routes; dataset
+    setup is part of upstream restoration
+  - Treat dataset requirements as undetermined pending source code recovery
+  - No data files are bundled in the current snapshot
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No authenticated data sources identified in current snapshot
+  - Public paper archive access sufficient for documentation review
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution viable for analysis route; no GPU required for metadata
+      review
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU acceleration would be needed for restored training execution; 8GB VRAM
+      sufficient for tabular MLP training
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - torch
+  - numpy
+  - scikit-learn
+  notes:
+  - Environment specifications are provisional pending upstream source restoration
+  - Current snapshot contains no executable setup files or runnable source code
+  - Bundled README and requirements.txt expected but not yet verified in snapshot
+  - Actual dependency versions should be confirmed from restored source before execution
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+- metric_not_anchored
+risk_notes:
+- The current snapshot preserves metadata only; no executable training or evaluation
+  code is available
+- Primary metric test_rmse cannot be evaluated without restored code anchor
+- No benchmark execution was performed in this packaging pass
+- Treat this entry as blocked until source restoration or richer local snapshot is
+  recovered
+- Top-level README absent in current snapshot; documentation may be incomplete
+recommended_when:
+- Conducting paper-backed planning or feasibility assessment for monotonic tabular
+  modeling
+- Performing code-audit review of theoretical contributions without execution
+- Preparing upstream restoration roadmap for constrained monotonic network benchmarks
+- Reviewing theoretical framework for hard monotonicity in neural networks
+not_recommended_when:
+- Expecting a ready-to-run benchmark with executable training and evaluation paths
+- Requiring immediate metric evaluation or model checkpoint generation
+- Needing bundled datasets and data loading pipelines
+- Treating this as a production-ready benchmark entry
+paper:
+  title: 'Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation
+    Beyond Bounded Activations'
+  authors:
+  - Davide Sartor
+  - Alberto Sinigaglia
+  - Gian Antonio Susto
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.02537
+  abstract_summary: 'Generalizes theoretical results for monotonic MLPs showing that
+    networks with non-negative weight constraints and alternating saturating activations
+    are universal approximators. Proves equivalence between activation saturation
+    side and weight constraint sign. Demonstrates that ReLU with non-positive constrained
+    weights also achieves universal approximation with greater-than-or-equal-to 4
+    layers. Proposes alternative parametrization eliminating weight reparameterization
+    for improved training stability.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.077_cmnn.zip
+  archive_type: zip
+  local_dir_name: paper-77-CMNN
+  notes:
+  - Archive contains metadata, documentation, and LaTeX source only
+  - No executable Python code bundled in current archive version
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.077_cmnn.zip
+  sha256: 4e7704808fbec948ed14f73c416eeb37bc3e1e1bd1735df0b5c66617a267c15f
+  size_bytes: 34676
+display:
+  palette_seed: sand-olive-monotone
+  art_style: trustworthy-minimal
+  accent_priority: medium
+  summary_cards:
+  - label: Theory Focus
+    value: Universal Approximation for Monotonic MLPs
+  - label: Architecture
+    value: Constrained Weight Networks
+  - label: Activation
+    value: Saturating Functions (ReLU family)
+  - label: Restoration Status
+    value: Blocked - No Code Anchor
+image_path: ../image/077_aisb.t3.077_cmnn.jpg
+audit_metadata:
+  static_code_audit_status: blocked_no_code
+  executable_anchors: none_detected
+  code_bearing_roots: none_detected
+  route_classification: prepackaged_import_candidate
+  owner_stage: baseline_prepackage
+recovery_rule: 'Prefer existing code files over README-only claims when choosing origin_path,
+  source_ref, and evaluation_protocol.code_paths. Keep metrics_summary empty until
+  trusted execution outputs exist. If the current snapshot has no executable anchor
+  for a metric, mark that metric as provisional or blocked rather than inventing support.
+
+  '
+execution_status_summary: 'The current snapshot is blocked or only partially anchorable
+  from code. No benchmark execution was performed in this packaging pass. Recovery
+  requires upstream source restoration before evaluation can proceed.
+
+  '

package/AISB/catalog/aisb.t3.077_cmnn.zh.yaml

@@ -0,0 +1,165 @@
+id: aisb.t3.077_cmnn
+name: 'Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation Beyond Bounded Activations'
+version: 0.1.0
+one_line: 受约束单调神经网络理论与表格建模评测，当前执行前需先恢复上游源码。
+task_description: 该条目覆盖用于表格建模的受约束单调神经网络。其理论贡献表明：采用非负权重约束与交替饱和激活的 MLP 可实现单调函数的通用逼近，突破传统有界激活设定；并证明在至少 4 层条件下，非正权重约束 + ReLU 的 MLP 也具备通用逼近性。所提参数化减少了权重重参数化需求，提升训练稳定性。当前本地快照仅保留元数据、审计文件与文档，不含可执行训练/评估代码；需上游恢复后方可运行。
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- id: test_rmse
+  description: Root mean squared error on held-out test set
+  status: blocked_no_code
+  note: No executable code anchor in current snapshot; requires restoration
+  code_anchor: null
+- id: metadata_review
+  description: Preserved benchmark framing and monotonic model documentation
+  status: available
+capability_tags:
+- research_code_optimization
+- monotonic_networks
+- tabular_ml
+- constrained_learning
+- trustworthy_ml
+- universal_approximation
+- hard_monotonicity
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Review the preserved metadata, paper documentation, and monotonic-model benchmark framing without execution.
+  estimated_time: <1h
+  resource_intensive: false
+- id: restore_first
+  label: Restore Then Execute
+  description: Restore the missing constrained monotonic model implementation and dataset setup before attempting benchmark execution.
+  estimated_time: 1-2h
+  resource_intensive: false
+  prerequisites:
+  - source_code_restoration
+  - dataset_acquisition
+  - dependency_verification
+dataset_download:
+  primary_method: pending_restoration
+  sources: []
+  notes:
+  - The current local package does not preserve executable benchmark routes; dataset setup is part of upstream restoration
+  - Treat dataset requirements as undetermined pending source code recovery
+  - No data files are bundled in the current snapshot
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No authenticated data sources identified in current snapshot
+  - Public paper archive access sufficient for documentation review
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution viable for analysis route; no GPU required for metadata review
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU acceleration would be needed for restored training execution; 8GB VRAM sufficient for tabular MLP training
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - torch
+  - numpy
+  - scikit-learn
+  notes:
+  - Environment specifications are provisional pending upstream source restoration
+  - Current snapshot contains no executable setup files or runnable source code
+  - Bundled README and requirements.txt expected but not yet verified in snapshot
+  - Actual dependency versions should be confirmed from restored source before execution
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+- metric_not_anchored
+risk_notes:
+- The current snapshot preserves metadata only; no executable training or evaluation code is available
+- Primary metric test_rmse cannot be evaluated without restored code anchor
+- No benchmark execution was performed in this packaging pass
+- Treat this entry as blocked until source restoration or richer local snapshot is recovered
+- Top-level README absent in current snapshot; documentation may be incomplete
+recommended_when:
+- 用于基于论文的单调表格建模规划或可行性评估
+- 在不执行代码时进行理论贡献审阅
+- 为受约束单调网络基准制定上游恢复路线图
+- 审阅硬单调性神经网络理论框架
+not_recommended_when:
+- 期望立即可运行并可训练评估的基准
+- 需要即刻产出指标或检查点
+- 需要随包提供的数据集与数据管道
+- 将其视为生产可用基准条目
+paper:
+  title: 'Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation Beyond Bounded Activations'
+  authors:
+  - Davide Sartor
+  - Alberto Sinigaglia
+  - Gian Antonio Susto
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.02537
+  abstract_summary: 'Generalizes theoretical results for monotonic MLPs showing that networks with non-negative weight constraints and alternating saturating activations are universal approximators. Proves equivalence between activation saturation side and weight constraint sign. Demonstrates that ReLU with non-positive constrained weights also achieves universal approximation with greater-than-or-equal-to 4 layers. Proposes alternative parametrization eliminating weight reparameterization for improved training stability.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.077_cmnn.zip
+  archive_type: zip
+  local_dir_name: paper-77-CMNN
+  notes:
+  - Archive contains metadata, documentation, and LaTeX source only
+  - No executable Python code bundled in current archive version
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.077_cmnn.zip
+  sha256: 4e7704808fbec948ed14f73c416eeb37bc3e1e1bd1735df0b5c66617a267c15f
+  size_bytes: 34676
+display:
+  palette_seed: sand-olive-monotone
+  art_style: trustworthy-minimal
+  accent_priority: medium
+  summary_cards:
+  - label: Theory Focus
+    value: Universal Approximation for Monotonic MLPs
+  - label: Architecture
+    value: Constrained Weight Networks
+  - label: Activation
+    value: Saturating Functions (ReLU family)
+  - label: Restoration Status
+    value: Blocked - No Code Anchor
+image_path: ../image/077_aisb.t3.077_cmnn.jpg
+audit_metadata:
+  static_code_audit_status: blocked_no_code
+  executable_anchors: none_detected
+  code_bearing_roots: none_detected
+  route_classification: prepackaged_import_candidate
+  owner_stage: baseline_prepackage
+recovery_rule: 'Prefer existing code files over README-only claims when choosing origin_path, source_ref, and evaluation_protocol.code_paths. Keep metrics_summary empty until trusted execution outputs exist. If the current snapshot has no executable anchor for a metric, mark that metric as provisional or blocked rather than inventing support.
+
+  '
+execution_status_summary: 'The current snapshot is blocked or only partially anchorable from code. No benchmark execution was performed in this packaging pass. Recovery requires upstream source restoration before evaluation can proceed.
+
+  '