@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.099_ollalanding.yaml

@@ -0,0 +1,248 @@
+id: aisb.t3.099_ollalanding
+name: OLLA Constrained Sampling
+version: 0.1.0
+one_line: Constrained non-log-concave sampling with projection-free OLLA/OLLA-H samplers
+  and diverse benchmark suite.
+task_description: 'This packaged benchmark implements OLLA (Overdamped Langevin with
+  LAnding), a suite of projection-free stochastic dynamics for sampling from constrained
+  statistical distributions. The framework handles both nonlinear equality and inequality
+  constraints without relying on projections or slack variables. OLLA enforces constraints
+  by projecting the overdamped Langevin drift and diffusion onto the tangent space
+  of the feasible set and uses a restitution ("landing") mechanism to guarantee convergence
+  to the constraint manifold. The benchmark includes five sampler implementations
+  (OLLA, OLLA-H, CLangevin, CHMC, CGHMC) and seven constraint problems spanning 2D
+  toy geometries, high-dimensional synthetic stress tests, polymer chain simulations,
+  and a fairness-aware Bayesian neural network. OLLA-H uses Hutchinson trace estimation
+  for scalable high-dimensional performance. The benchmark evaluates sampling quality
+  via test negative log-likelihood (test_nll), effective sample size (ESS), CPU time,
+  and constraint violation metrics.
+
+  '
+capability_tags:
+- research_code_optimization
+- sampling
+- constrained_optimization
+- probabilistic_modeling
+- monte_carlo
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- test_nll
+- cpu_time
+- ess
+- constraint_violations
+- sampling_report
+- csv_summaries
+- pdf_plots
+snapshot_status: runnable
+support_level: advanced
+recommended_when: 'Use this benchmark when you need to evaluate constrained sampling
+  algorithms on challenging nonconvex constraint sets, compare OLLA/OLLA-H against
+  classical constrained samplers (Langevin, HMC, GHMC), or stress-test sampler scalability
+  across low-dimensional (2D) to high-dimensional (100+ dim) problems.
+
+  '
+not_recommended_when: 'Do not use this benchmark for simple supervised-learning tasks,
+  unconstrained sampling problems, or when only basic log-concave convex-constrained
+  sampling is required.
+
+  '
+dataset_download:
+  primary_method: automatic
+  sources:
+  - url: https://deepscientist.cc/AISB/099_ollalanding
+    type: archive
+  - name: OpenML
+    type: runtime
+    description: German Credit dataset (german_credit experiment) downloaded automatically
+      on first use
+  notes:
+  - Ensure outbound network access is available when running the german_credit constraint
+    experiment
+  - Archive contains full source code, environment.yml, and experiment runner
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No authentication required for dataset downloads
+  - OpenML access requires standard internet connectivity
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution fully supported for all minimum-route experiments
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: GPU recommended only for parallel multichain sampling; CPU preferred for
+      standard experiments
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: 2.3.0
+  flash_attn: null
+  key_packages:
+  - numpy
+  - pandas
+  - scikit-learn
+  - matplotlib
+  - arviz
+  - tqdm
+  - torch
+  notes:
+  - Environment managed via conda/environment.yml
+  - Package installed in editable mode (pip install -e .)
+  - CPU-only execution is preferred across all experiments unless running parallel
+    sampling for multichain
+  - GPU support available via PyTorch CUDA builds (e.g., cu130); not required for
+    primary experiments
+risk_flags:
+- network_on_first_run
+- long_runtime
+risk_notes:
+- German Credit experiment downloads from OpenML on first execution
+- Full sampler suite may require up to 24 hours depending on dimensionality and sampler
+  count
+launch_profiles:
+- id: quick_analysis
+  label: Quick Analysis
+  description: 'Run constrained-sampling analysis on a small experiment subset. Executes
+    experiments/run_analysis.py with a default constraint and sampler selection on
+    CPU. Suitable for validation runs and method familiarization. Typical runtime:
+    1-4 hours.
+
+    '
+  command_template: python experiments/run_analysis.py --experiments circle --samplers
+    OLLA-H --device cpu
+- id: full_sampler_suite
+  label: Full Sampler Suite
+  description: 'Execute the complete OLLA benchmark workflow across all constraint
+    problems and all five samplers (OLLA, OLLA-H, CLangevin, CHMC, CGHMC). Includes
+    2D experiments with PDF visualization, high-dimensional stress tests, polymer
+    chain sampling, and German Credit Bayesian neural network. Produces CSV summaries,
+    ESS metrics, constraint violation diagnostics, and test_nll comparisons. Typical
+    runtime: 6-24 hours.
+
+    '
+  command_template: python experiments/run_analysis.py --experiments circle two_lobe
+    quadratic_poly mix_gaussian highdim_stress highdim_polymer german_credit --samplers
+    OLLA OLLA-H CLangevin CHMC CGHMC --device cpu
+paper:
+  title: Fast Non-Log-Concave Sampling under Nonconvex Equality and Inequality Constraints
+    with Landing
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2510.22044
+  authors:
+  - Jeon, Kijung
+  - Muehlebach, Michael
+  - Tao, Molei
+  institution:
+  - Georgia Institute of Technology
+  - MPI-IS
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.099_ollalanding.zip
+  archive_type: zip
+  local_dir_name: paper-99-OLLALanding
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.099_ollalanding.zip
+  sha256: 65bf98fd2de3dabac25cdb41e57ac2f63756a3336378835d48cb5462ec3e9d4a
+  size_bytes: 845998
+display:
+  palette_seed: charcoal-mint-landing
+  art_style: constrained-geometry
+  accent_priority: high
+  tags:
+  - sampling
+  - constrained-optimization
+  - langevin-dynamics
+  - bayesian-inference
+image_path: ../image/099_aisb.t3.099_ollalanding.jpg
+citation: "@inproceedings{jeon2025olla,\n  title     = {Fast Non-Log-Concave Sampling\
+  \ under Nonconvex Equality and Inequality Constraints with Landing},\n  author \
+  \   = {Jeon, Kijung and Muehlebach, Michael and Tao, Molei},\n  booktitle = {Advances\
+  \ in Neural Information Processing Systems},\n  year      = {2025},\n}\n"
+metric_evidence:
+  test_nll:
+    origin: experiments/run_analysis.py
+    status: code_backed
+  cpu_time:
+    origin: experiments/run_analysis.py
+    status: code_backed
+  ess:
+    origin: src/constraints/highdim_stress.py
+    status: code_backed
+constraint_library:
+- id: circle
+  description: 2D Gaussian restricted to unit circle (x1^2 + x2^2 = 1). Analytic manifold
+    initialization.
+- id: quadratic_poly
+  description: Polynomial equality (x1^4 x2^2 + x1^2 + x2 = 1) with inequality (x1^3
+    - x2^3 <= 1). Rejection-based init.
+- id: mix_gaussian
+  description: Nine-component isotropic Gaussian mixture constrained to seven-lobed
+    star contour with polynomial inequality.
+- id: two_lobe
+  description: Double-moon density via inequality constraint. Highlights curved boundaries
+    without equalities.
+- id: highdim_stress
+  description: Random high-dimensional linear and spherical equalities with spherical
+    obstacle inequalities. Scalability test.
+- id: highdim_polymer
+  description: Configurable 3D polymer chain with bond-length, bond-angle constraints,
+    steric inequalities, torsion/WCA potentials.
+- id: german_credit
+  description: Fairness-aware 2-layer Bayesian neural network with TPR/FPR parity
+    equality and feature monotonicity inequalities.
+sampler_implementations:
+- id: OLLA
+  class: OLLA
+  path: src/samplers/olla.py
+  description: Original OLLA with analytic trace correction. Low-to-moderate dimensions.
+- id: OLLA-H
+  class: OLLA_H
+  path: src/samplers/olla_h.py
+  description: Hutchinson-trace variant. Preferred for high-dimensional problems or
+    many constraints. Set Hutchinson probes N=0 for maximum speed.
+- id: CLangevin
+  class: CLangevin
+  path: src/samplers/clangevin.py
+  description: Constrained Langevin with slack-variable augmentation and SHAKE-style
+    projections for inequality handling.
+- id: CHMC
+  class: CHMC
+  path: src/samplers/chmc.py
+  description: Constrained Hamiltonian Monte Carlo using RATTLE integrator on augmented
+    state with slack variables.
+- id: CGHMC
+  class: CGHMC
+  path: src/samplers/cghmc.py
+  description: Constrained generalized HMC with Metropolis-Hastings correction. Inequality
+    constraints via rejection during MH step.
+execution_notes: 'All samplers inherit from BaseSampler (src/samplers/__init__.py)
+  which standardizes RNG handling, step-size configuration, and return conventions
+  (trajectory, h_vals, g_vals). The experiment runner (run_analysis.py) generates
+  CSV summaries, optional PDF plots for 2D experiments, and prints diagnostics including
+  ESS, constraint violations, and task-specific metrics. Each constraint module exposes
+  EXPERIMENT_SETTINGS (steps, particle counts) and SAMPLER_SETTINGS (per-sampler hyperparameters)
+  consumed by the runner.
+
+  '

package/AISB/catalog/aisb.t3.099_ollalanding.zh.yaml

@@ -0,0 +1,182 @@
+id: aisb.t3.099_ollalanding
+name: OLLA 约束采样
+version: 0.1.0
+one_line: 使用无投影 OLLA/OLLA-H 采样器和多样化基准套件进行约束非对数凹采样。
+task_description: 该打包基准测试实现了 OLLA（Overdamped Langevin with LANDing），这是一套无投影随机动力学，用于从受约束的统计分布中进行采样。该框架无需依赖投影或松弛变量即可处理非线性等式和不等式约束。 OLLA 通过将过阻尼朗之万漂移和扩散投影到可行集的切线空间上来强制约束，并使用恢复（“着陆”）机制来保证收敛到约束流形。该基准测试包括五个采样器实现（OLLA、OLLA-H、CLangevin、CHMC、CGHMC）和涵盖 2D 玩具几何形状、高维合成压力测试、聚合物链模拟和公平感知贝叶斯神经网络的七个约束问题。 OLLA-H 使用 Hutchinson 迹估计来实现可扩展的高维性能。该基准测试通过测试负对数似然 (test_nll)、有效样本大小 (ESS)、CPU 时间和约束违规指标来评估采样质量。
+capability_tags:
+- research_code_optimization
+- sampling
+- constrained_optimization
+- probabilistic_modeling
+- monte_carlo
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- test_nll
+- cpu_time
+- ess
+- constraint_violations
+- sampling_report
+- csv_summaries
+- pdf_plots
+snapshot_status: runnable
+support_level: advanced
+recommended_when: 当您需要在具有挑战性的非凸约束集上评估约束采样算法、将 OLLA/OLLA-H 与经典约束采样器（Langevin、HMC、GHMC）进行比较，或者跨低维 (2D) 到高维（100+ 暗淡）问题对采样器可扩展性进行压力测试时，请使用此基准测试。
+not_recommended_when: 不要将此基准用于简单的监督学习任务、无约束采样问题或仅需要基本对数凹凸约束采样时。
+dataset_download:
+  primary_method: automatic
+  sources:
+  - url: https://deepscientist.cc/AISB/099_ollalanding
+    type: archive
+  - name: OpenML
+    type: runtime
+    description: German Credit dataset (german_credit experiment) downloaded automatically on first use
+  notes:
+  - 确保运行 german_credit 约束实验时出站网络访问可用
+  - 存档包含完整的源代码、environment.yml 和实验运行程序
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 数据集下载无需身份验证
+  - OpenML 访问需要标准互联网连接
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: 所有最小路由实验完全支持仅 CPU 执行
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: GPU 仅推荐用于并行多链采样；标准实验首选 CPU
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: 2.3.0
+  flash_attn: null
+  key_packages:
+  - numpy
+  - pandas
+  - scikit-learn
+  - matplotlib
+  - arviz
+  - tqdm
+  - torch
+  notes:
+  - 通过 conda/environment.yml 管理的环境
+  - 以可编辑模式安装的软件包 (pip install -e .)
+  - 在所有实验中，仅 CPU 执行是首选，除非为多链运行并行采样
+  - 通过 PyTorch CUDA 构建提供 GPU 支持（e.g.、cu130）；初级实验不需要
+risk_flags:
+- network_on_first_run
+- long_runtime
+risk_notes:
+- 德国信用实验首次执行时从 OpenML 下载
+- 完整的采样器套件可能需要长达 24 小时，具体取决于维度和采样器数量
+launch_profiles:
+- id: quick_analysis
+  label: 快速分析
+  description: 对小型实验子集运行约束抽样分析。在 CPU 上使用默认约束和采样器选择执行experiments/run_analysis.py。适合验证运行和方法熟悉。典型运行时间：1-4 小时。
+  command_template: python experiments/run_analysis.py --experiments circle --samplers OLLA-H --device cpu
+- id: full_sampler_suite
+  label: 完整的采样器套件
+  description: 跨所有约束问题和所有五个采样器（OLLA、OLLA-H、CLangevin、CHMC、CGHMC）执行完整的 OLLA 基准工作流程。包括 PDF 可视化、高维压力测试、聚合物链采样和德国信贷贝叶斯神经网络的 2D 实验。生成 CSV 摘要、ESS 指标、约束违规诊断和 test_nll 比较。典型运行时间：6-24 小时。
+  command_template: python experiments/run_analysis.py --experiments circle two_lobe quadratic_poly mix_gaussian highdim_stress highdim_polymer german_credit --samplers OLLA OLLA-H CLangevin CHMC CGHMC --device cpu
+paper:
+  title: Fast Non-Log-Concave Sampling under Nonconvex Equality and Inequality Constraints with Landing
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2510.22044
+  authors:
+  - Jeon, Kijung
+  - Muehlebach, Michael
+  - Tao, Molei
+  institution:
+  - Georgia Institute of Technology
+  - MPI-IS
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.099_ollalanding.zip
+  archive_type: zip
+  local_dir_name: paper-99-OLLALanding
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.099_ollalanding.zip
+  sha256: 65bf98fd2de3dabac25cdb41e57ac2f63756a3336378835d48cb5462ec3e9d4a
+  size_bytes: 845998
+display:
+  palette_seed: charcoal-mint-landing
+  art_style: constrained-geometry
+  accent_priority: high
+  tags:
+  - sampling
+  - constrained-optimization
+  - langevin-dynamics
+  - bayesian-inference
+image_path: ../image/099_aisb.t3.099_ollalanding.jpg
+citation: "@inproceedings{jeon2025olla,\n  title     = {Fast Non-Log-Concave Sampling under Nonconvex Equality and Inequality Constraints with Landing},\n  author    = {Jeon, Kijung and Muehlebach, Michael and Tao, Molei},\n  booktitle = {Advances in Neural Information Processing Systems},\n  year      = {2025},\n}\n"
+metric_evidence:
+  test_nll:
+    origin: experiments/run_analysis.py
+    status: code_backed
+  cpu_time:
+    origin: experiments/run_analysis.py
+    status: code_backed
+  ess:
+    origin: src/constraints/highdim_stress.py
+    status: code_backed
+constraint_library:
+- id: circle
+  description: 二维高斯仅限于单位圆 (x1^2 + x2^2 = 1)。解析流形初始化。
+- id: quadratic_poly
+  description: 多项式等式 (x1^4 x2^2 + x1^2 + x2 = 1) 和不等式 (x1^3 - x2^3 <= 1)。基于拒绝的 init。
+- id: mix_gaussian
+  description: 九分量各向同性高斯混合约束为具有多项式不等式的七瓣星形轮廓。
+- id: two_lobe
+  description: 通过不等式约束的双月密度。突出显示不等式的弯曲边界。
+- id: highdim_stress
+  description: 随机高维线性和球面等式以及球面障碍物不等式。可扩展性测试。
+- id: highdim_polymer
+  description: 可配置的 3D 聚合物链，具有键长、键角约束、空间不等式、扭转/WCA 势。
+- id: german_credit
+  description: 具有 TPR/FPR 奇偶性和特征单调性不等式的公平感知 2 层贝叶斯神经网络。
+sampler_implementations:
+- id: OLLA
+  class: OLLA
+  path: src/samplers/olla.py
+  description: 带有分析迹线校正的原始 OLLA。低到中等尺寸。
+- id: OLLA-H
+  class: OLLA_H
+  path: src/samplers/olla_h.py
+  description: 哈金森踪迹变体。对于高维问题或许多约束的首选。设置 Hutchinson 探针 N=0 以获得最大速度。
+- id: CLangevin
+  class: CLangevin
+  path: src/samplers/clangevin.py
+  description: 使用松弛变量增强和 SHAKE 式预测来约束 Langevin，以处理不平等问题。
+- id: CHMC
+  class: CHMC
+  path: src/samplers/chmc.py
+  description: 在具有松弛变量的增强状态上使用 RATTLE 积分器进行约束哈密顿蒙特卡罗。
+- id: CGHMC
+  class: CGHMC
+  path: src/samplers/cghmc.py
+  description: 具有 Metropolis-Hastings 校正的约束广义 HMC。 MH 步骤期间通过拒绝进行不平等约束。
+execution_notes: 'All samplers inherit from BaseSampler (src/samplers/__init__.py) which standardizes RNG handling, step-size configuration, and return conventions (trajectory, h_vals, g_vals). The experiment runner (run_analysis.py) generates CSV summaries, optional PDF plots for 2D experiments, and prints diagnostics including ESS, constraint violations, and task-specific metrics. Each constraint module exposes EXPERIMENT_SETTINGS (steps, particle counts) and SAMPLER_SETTINGS (per-sampler hyperparameters) consumed by the runner.
+
+  '

package/AISB/catalog/aisb.t3.100_invmissingdata.yaml

@@ -0,0 +1,179 @@
+id: aisb.t3.100_invmissingdata
+name: Kernel Point Imputation (KPI)
+version: 0.1.0
+one_line: Bi-level kernel-adaptive iterative imputation leveraging oracle-feature
+  supervision for heterogeneous missing-data recovery.
+task_description: 'This benchmark implements KPI (Kernel Point Imputation), a bi-level
+  optimization framework for iterative missing-data imputation. The inner loop adapts
+  per-feature imputation models via functional optimization in a Reproducing Kernel
+  Hilbert Space (RKHS), addressing model misspecification in heterogeneous data-generating
+  processes. The outer loop exploits oracle (fully observed) features as privileged
+  supervision signals. Supports CPU-only execution for minimum route; GPU acceleration
+  available for full workflow.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- mae
+- wass
+- rmse
+- imputation_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run the packaged KPI benchmark script on a prepared dataset with default
+    hyperparameters. Suitable for rapid metric extraction from pre-processed UCI data
+    without full training.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval
+  description: 'Execute the complete KPI pipeline: kernel point imputation optimization,
+    bi-level training with oracle-feature supervision, and full benchmark suite across
+    all supported datasets. Includes Wasserstein distance evaluation.
+
+    '
+dataset_download:
+  primary_method: manual
+  sources:
+  - name: UCI Machine Learning Repository
+    url: https://archive.ics.uci.edu
+    notes: 'Pre-processed tabular datasets (e.g., blood_transfusion, concrete_compression)
+      should be placed in the ./dataset folder. Specific dataset references are documented
+      in benchmark_kpi.py argument defaults and aisb_catalog.yaml.
+
+      '
+  - name: Benchmark archive
+    url: https://deepscientist.cc/AISB/100_invmissingdata
+    archive_type: zip
+    local_dir_name: paper-100-InvMissingData
+  notes:
+  - Both UCI source data and the packaged benchmark archive are required.
+  - Dataset preparation steps are detailed in the bundled README.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution is fully supported for the minimum configuration.
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU recommended for full_train_eval to accelerate kernel computations and
+      bi-level optimization.
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.7.1
+  flash_attn: null
+  key_packages:
+  - ot (Optimal Transport)
+  - numpy
+  - pandas
+  - scipy
+  - torch
+  notes:
+  - CPU-only execution is plausible for the minimum route.
+  - See bundled requirements.txt for the full dependency set.
+  - Optional Docker container with conda-based environment named 'kpi' is provided.
+risk_flags: []
+risk_notes:
+- Static code audit confirmed executable anchors exist for all staged metrics.
+- No runtime execution was performed during this packaging pass; metric values require
+  verification via execution.
+recommended_when: 'Use this benchmark when you need a missing-data benchmark that
+  focuses on flexible nonparametric imputation behavior with kernel-adapted geometry
+  and oracle-feature supervision. Appropriate for tabular data with heterogeneous
+  missingness patterns.
+
+  '
+not_recommended_when: Do not use this if you require a heavy multimodal benchmark
+  or deep learning-based imputation without explicit kernel-method foundations.
+paper:
+  title: Inverse Methods for Missing Data Imputation
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/kim25ag.html
+  code_url: null
+  baseline_id: aisb.t3.100_invmissingdata
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.100_invmissingdata.zip
+  archive_type: zip
+  local_dir_name: paper-100-InvMissingData
+  notes:
+  - Archive contains benchmark_kpi.py, model/KPI.py, scripts/, and requirements.txt.
+  - UCI datasets must be obtained separately from archive.ics.uci.edu.
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.100_invmissingdata.zip
+  sha256: 0b54fbad4ac73321054a33a6dc80e069c22a81cf54788c237d938a8f1caf70ef
+  size_bytes: 1388561
+display:
+  palette_seed: bone-blue-impute
+  art_style: imputation-lab
+  accent_priority: medium
+  tags:
+  - research_code_optimization
+  - missing_data
+  - imputation
+  - kernel_methods
+  - tabular_ml
+image_path: ../image/100_aisb.t3.100_invmissingdata.jpg
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+capability_tags:
+- research_code_optimization
+- missing_data
+- imputation
+- kernel_methods
+- tabular_ml
+executable_anchors:
+  benchmark_script: benchmark_kpi.py
+  core_model: model/KPI.py
+  scripts_dir: scripts/
+  evaluation_protocol:
+  - scripts/qb.sh (quick_eval)
+  - scripts/is.sh (full_train_eval)
+metric_evidence:
+  mae:
+    origin_path: model/remasker/model_mae.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  wass:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  rmse:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+  mse:
+    origin_path: benchmark_kpi.py
+    source_ref: code-backed
+    status: executable_anchor_found
+caveat_snapshot: 'Static code audit confirmed executable anchors for every staged
+  metric. Runtime execution has not been performed; metric values require verification
+  before trust.
+
+  '