@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.b11.model_efficiency.yaml

@@ -0,0 +1,231 @@
+schema_version: 1
+id: aisb.b11.model_efficiency
+name: AISB B11 Model Efficiency
+version: 0.2.0
+catalog_version: 0.1.0
+one_line: LongBench-v2 mini request set for quality-latency-cost metric validation.
+task_description: AISB B11 Model Efficiency is an AISB BenchStore catalog pilot entry. The package is
+  intended for public-dev workspace initialization, public evaluator checks, and paper/benchmark-track
+  submission scaffolding. Official leaderboard scores require organizer-side hidden replay.
+homepage: https://aisb.deepscientist.cc
+localized:
+  zh:
+    name: AISB B11 模型效率
+    one_line: 用于质量、延迟和成本指标验证的 LongBench-v2 mini 请求集。
+    task_description: AISB B11 模型效率 是 AISB BenchStore catalog pilot 条目。公开包只用于 public-dev 工作区初始化、公开 evaluator
+      检查和论文/benchmark track 提交脚手架；正式 leaderboard 分数必须由组织方 hidden replay 产生。
+capability_tags:
+- model_efficiency
+- long_context
+- latency
+aisb_direction: B11
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: benchmark
+benchmark_mode: quality_latency_cost_logged_inference
+requires_execution: true
+requires_paper: true
+integrity_level: hidden_reference
+snapshot_status: partial
+support_level: preview
+discovery:
+  collection: AISB
+  collection_priority: 100
+  recommendation_weight: 500
+  featured: false
+  featured_reason: null
+display:
+  placement: grid
+  card_size: m
+  badge: AISB
+  accent_priority: normal
+official_links:
+  homepage: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  github: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  benchmark_overview: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B11_ModelEfficiency
+release_gate:
+  release_ready: false
+  status: registry_pending
+  full_real_replay: passed
+  reason: Public-dev and scorer-private replay paths exist locally, but final registry promotion, release
+    assets, and license review remain pending.
+  next_required_step: Validate the published BenchStore release asset on a clean host, then complete
+    registry and license review.
+primary_outputs:
+- results.json
+- paper/paper.pdf
+- paper/claims.json
+- logs/experiment_log.jsonl
+- logs/iterations.jsonl
+cost_band: medium
+time_band: 1-4h
+difficulty: hard
+data_access: public
+resources:
+  minimum:
+    cpu_cores: 2
+    ram_gb: 4
+    disk_gb: 4
+    gpu_count: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+source:
+  benchmark_dir: benchmarks/aisb/B11_ModelEfficiency
+  bench_yaml: benchmarks/aisb/B11_ModelEfficiency/bench.yaml
+  source_repo: ResearAI/NLPCC-2026-Task9-AISB
+  source_commit: b6b7527a226ca7328697156608571f2cb7dcda3a
+download:
+  provider: github_release
+  repo: ResearAI/NLPCC-2026-Task9-AISB
+  tag: aisb-v0.0.1
+  asset_name: aisb.b11.model_efficiency-v0.2.0.zip
+  url: https://github.com/giao-123-sun/DeepScientist/releases/download/aisb-v0.0.1/aisb.b11.model_efficiency-v0.2.0.zip
+  source_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B11_ModelEfficiency
+  archive_type: zip
+  local_dir_name: aisb.b11.model_efficiency
+  sha256: 890c432a9beed8912271a6fa72dde4c08fd58b89d74df5674e741e4a04c904db
+  size_bytes: 974447
+  published_at: '2026-04-30T00:00:00Z'
+  asset_status: published
+dataset_download:
+  primary_method: packaged_public_dev
+  sources:
+  - kind: public_dev
+    access: public
+    url: null
+    note: Public-dev files are included in the published BenchStore package.
+  - kind: hidden_reference
+    access: private
+    url: null
+    note: Organizer-only hidden references are excluded from catalog packages.
+credential_requirements:
+  mode: none
+  notes:
+  - No credential is required for packaged public-dev evaluation.
+  - External model APIs are optional participant choices and not bundled.
+environment:
+  python: '>=3.11'
+  docker: false
+  key_packages:
+  - pytest
+  - PyYAML
+  notes:
+  - Use docker/Dockerfile when available for official replay parity.
+  - Do not copy organizer-private hidden references into participant workspaces.
+launch_profiles:
+- id: b11_public_dev_smoke
+  label: B11 public-dev smoke test
+  description: Initialize an AISB public-dev workspace, validate submission artifacts, and run the public
+    evaluator.
+  startup_contract:
+  - Use only files inside the installed benchmark workspace and the generated submission directory.
+  - Treat hidden references as organizer-only and unavailable to the agent.
+  - Produce benchmark outputs, paper artifacts, claims, and logs before packaging.
+  commands:
+  - python scripts/agent_tools.py workspace init B11 --dest .work/benchstore/B11
+  - python scripts/agent_tools.py evaluate B11 --bench-dir .work/benchstore/B11 --submission .work/benchstore/B11/submission
+  - python scripts/agent_tools.py submission validate .work/benchstore/B11/submission
+  - python scripts/agent_tools.py submission package .work/benchstore/B11/submission --output .work/benchstore/B11/submission.zip
+  - python scripts/agent_tools.py submission replay .work/benchstore/B11/submission --track B11
+  one_click_prompt: Start AISB B11 from the installed BenchStore package. Initialize the public-dev workspace,
+    inspect AGENT.md and README.md, run the baseline or your method, evaluate locally, validate/package
+    the submission, and record any release-gate risks in logs/experiment_log.jsonl.
+paper:
+  venue: AISB
+  year: 2026
+  track_fit:
+  - paper_track
+  - benchmark_track
+  requires_claims_json: true
+  url: null
+benchmark_track:
+  metric: quality_latency_cost_score
+  benchmark_source: LongBench-v2 mini request set
+  packaged_data_status: real LongBench-v2 public-dev request set plus scorer-private hidden rows
+  official_score_requires_hidden_replay: true
+benchmarks:
+  primary:
+    name: LongBench-v2 mini
+    source: THUDM/LongBench-v2 (HF, Apache-2.0)
+    tasks: 3
+    hidden: 3
+    metric: quality_latency_cost_score (quality 0.55, latency 0.20, memory 0.10, tokens 0.10, cost 0.05)
+    sota: 63.3% (Gemini-2.5-Pro, w/ CoT)
+    our_baseline: 0.3529
+  alternatives:
+  - name: RULER
+    source: github.com/hsiehjackson/RULER
+    note: long-context efficiency; alternative metric
+  - name: MLPerf Inference
+    source: mlcommons.org
+    note: hardware-normalized; future candidate
+papers:
+  highlights:
+  - title: 'LongBench v2: Towards Deeper Understanding on Realistic Long-Context Multitasks'
+    source: THUDM/LongBench-v2 (HF, Apache-2.0)
+    why: Primary benchmark; SOTA 63.3% (Gemini-2.5-Pro)
+  - title: 'Youtu-LLM: Unleashing Native Agent Potential for Lightweight LLMs'
+    arxiv: '2512.24618'
+    url: https://arxiv.org/abs/2512.24618
+    venue: Dec 2025
+    why: 1.96B params; sub-2B SOTA; quality-latency-cost balance
+  - title: 'MARS: Efficient Adaptive Co-Scheduling for Heterogeneous Agent Systems'
+    arxiv: '2604.26963'
+    url: https://arxiv.org/abs/2604.26963
+    venue: April 2026
+    why: 5.94x end-to-end latency reduction; GPU-CPU co-scheduling
+  full_references_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B11_*/references/papers.json
+nlpcc_tracks:
+- track: T1
+  name: AI/CS Reasoning & Engineering
+  url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/nlpcc/T1
+private_hidden_reference_strategy:
+  policy: organizer_only
+  public_package_includes_hidden_reference: false
+  public_catalog_discloses_private_paths: false
+  score_release: aggregate_only
+  notes:
+  - Hidden answers, labels, canaries, and private evaluator internals are never packaged.
+  - Official scores must record hidden_reference_version without disclosing hidden content.
+score_card:
+  required_fields:
+  - team_id
+  - system_name
+  - direction
+  - track
+  - benchmark_package_version
+  - evaluator_version
+  - hidden_reference_version
+  - docker_image_digest
+  - data_split_version
+  - benchmark_score
+  - paper_score
+  - cas_score
+  - final_score
+  - security_verdict
+  - license_verdict
+  - used_reference_scope
+  - official_score
+  - created_at
+  track_scoring:
+    paper_track: Final_A = 0.0 * S_benchmark + 1.0 * S_paper
+    benchmark_track: Final_B = 0.7 * S_benchmark + 0.3 * S_paper
+    cas: integrity gate, not bonus
+risk_flags:
+- hidden_reference_not_public
+- not_release_ready
+- registry_promotion_pending
+- hardware_normalization_pending
+risk_notes:
+- This catalog entry is not release_ready.
+- Download URL, checksum, and size point to the published GitHub release asset.
+- Official scores require organizer-side replay against private hidden references.
+recommended_when: Use this entry for AISB public-dev integration, local agent workflow checks, and catalog
+  UI validation.
+not_recommended_when: Do not treat this entry as an official release or public leaderboard claim until
+  release gates pass.
+image_path: ../image/aisb.b11.model_efficiency.svg

package/AISB/catalog/aisb.b12.embodied_ai.yaml

@@ -0,0 +1,238 @@
+schema_version: 1
+id: aisb.b12.embodied_ai
+name: AISB B12 Embodied AI
+version: 0.4.0
+catalog_version: 0.1.0
+one_line: CALVIN task-spec public-dev mini package for headless embodied rollout validation.
+task_description: AISB B12 Embodied AI is an AISB BenchStore catalog pilot entry. The package is intended
+  for public-dev workspace initialization, public evaluator checks, and paper/benchmark-track submission
+  scaffolding. Official leaderboard scores require organizer-side hidden replay.
+homepage: https://aisb.deepscientist.cc
+localized:
+  zh:
+    name: AISB B12 具身智能
+    one_line: 用于无头具身 rollout 验证的 CALVIN 任务规格 public-dev mini 包。
+    task_description: AISB B12 具身智能 是 AISB BenchStore catalog pilot 条目。公开包只用于 public-dev 工作区初始化、公开 evaluator
+      检查和论文/benchmark track 提交脚手架；正式 leaderboard 分数必须由组织方 hidden replay 产生。
+capability_tags:
+- embodied_ai
+- simulator
+- robotics
+aisb_direction: B12
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: benchmark
+benchmark_mode: embodied_episode_success_headless_rollout
+requires_execution: true
+requires_paper: true
+integrity_level: hidden_reference
+snapshot_status: partial
+support_level: preview
+discovery:
+  collection: AISB
+  collection_priority: 100
+  recommendation_weight: 500
+  featured: false
+  featured_reason: null
+display:
+  placement: grid
+  card_size: m
+  badge: AISB
+  accent_priority: normal
+official_links:
+  homepage: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  github: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  benchmark_overview: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B12_EmbodiedAI
+release_gate:
+  release_ready: false
+  status: registry_pending
+  full_real_replay: passed
+  reason: Public-dev and scorer-private replay paths exist locally, but final registry promotion, release
+    assets, and license review remain pending.
+  next_required_step: Validate the published BenchStore release asset on a clean host, then complete
+    registry and license review.
+primary_outputs:
+- results.json
+- paper/paper.pdf
+- paper/claims.json
+- logs/experiment_log.jsonl
+- logs/iterations.jsonl
+cost_band: high
+time_band: 2-8h
+difficulty: hard
+data_access: public
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 0
+source:
+  benchmark_dir: benchmarks/aisb/B12_EmbodiedAI
+  bench_yaml: benchmarks/aisb/B12_EmbodiedAI/bench.yaml
+  source_repo: ResearAI/NLPCC-2026-Task9-AISB
+  source_commit: b6b7527a226ca7328697156608571f2cb7dcda3a
+download:
+  provider: github_release
+  repo: ResearAI/NLPCC-2026-Task9-AISB
+  tag: aisb-v0.0.1
+  asset_name: aisb.b12.embodied_ai-v0.4.0.zip
+  url: https://github.com/giao-123-sun/DeepScientist/releases/download/aisb-v0.0.1/aisb.b12.embodied_ai-v0.4.0.zip
+  source_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B12_EmbodiedAI
+  archive_type: zip
+  local_dir_name: aisb.b12.embodied_ai
+  sha256: cd221f86fddd6123f7487e0ec3e27dfc881aaf005a07c712a9d1fe6bab44e156
+  size_bytes: 25349
+  published_at: '2026-04-30T00:00:00Z'
+  asset_status: published
+dataset_download:
+  primary_method: packaged_public_dev
+  sources:
+  - kind: public_dev
+    access: public
+    url: null
+    note: Public-dev files are included in the published BenchStore package.
+  - kind: hidden_reference
+    access: private
+    url: null
+    note: Organizer-only hidden references are excluded from catalog packages.
+credential_requirements:
+  mode: none
+  notes:
+  - No credential is required for packaged public-dev evaluation.
+  - External model APIs are optional participant choices and not bundled.
+environment:
+  python: '>=3.11'
+  docker: true
+  key_packages:
+  - pytest
+  - PyYAML
+  notes:
+  - Use docker/Dockerfile when available for official replay parity.
+  - Do not copy organizer-private hidden references into participant workspaces.
+launch_profiles:
+- id: b12_public_dev_smoke
+  label: B12 public-dev smoke test
+  description: Initialize an AISB public-dev workspace, validate submission artifacts, and run the public
+    evaluator.
+  startup_contract:
+  - Use only files inside the installed benchmark workspace and the generated submission directory.
+  - Treat hidden references as organizer-only and unavailable to the agent.
+  - Produce benchmark outputs, paper artifacts, claims, and logs before packaging.
+  commands:
+  - python scripts/agent_tools.py workspace init B12 --dest .work/benchstore/B12
+  - python scripts/agent_tools.py evaluate B12 --bench-dir .work/benchstore/B12 --submission .work/benchstore/B12/submission
+  - python scripts/agent_tools.py submission validate .work/benchstore/B12/submission
+  - python scripts/agent_tools.py submission package .work/benchstore/B12/submission --output .work/benchstore/B12/submission.zip
+  - python scripts/agent_tools.py submission replay .work/benchstore/B12/submission --track B12
+  one_click_prompt: Start AISB B12 from the installed BenchStore package. Initialize the public-dev workspace,
+    inspect AGENT.md and README.md, run the baseline or your method, evaluate locally, validate/package
+    the submission, and record any release-gate risks in logs/experiment_log.jsonl.
+paper:
+  venue: AISB
+  year: 2026
+  track_fit:
+  - paper_track
+  - benchmark_track
+  requires_claims_json: true
+  url: null
+benchmark_track:
+  metric: episode_success_rate
+  benchmark_source: CALVIN headless subset
+  packaged_data_status: real CALVIN task-spec mini split with scorer-private debug-dataset headless rollout
+  official_score_requires_hidden_replay: true
+benchmarks:
+  primary:
+    name: CALVIN debug
+    source: github.com/mees/calvin (MIT)
+    tasks: 2
+    hidden: 2
+    metric: episode_success_rate
+    sota: FLOWER 4.53/5 (ABC, 2025)
+    public_score: 0.0
+  alternatives:
+  - name: SimplerEnv
+    source: github.com/simpler-env/SimplerEnv
+    note: lighter simulator; future candidate
+  - name: Behavior-1K
+    source: behavior.stanford.edu
+    note: 1000 everyday tasks; heavy
+  - name: BuilderBench
+    source: arXiv:2510.06288
+    note: 42 target structures; open exploration
+papers:
+  highlights:
+  - title: 'CALVIN: A Benchmark for Language-Conditioned Policy Learning'
+    arxiv: '2112.03227'
+    url: https://arxiv.org/abs/2112.03227
+    why: Primary benchmark; debug dataset used for public-dev (MIT)
+  - title: 'EmboCoach-Bench: Benchmarking AI Agents in Developing Embodied Robots'
+    arxiv: '2601.21570'
+    url: https://arxiv.org/abs/2601.21570
+    venue: Jan 2026
+    why: 32 RL/IL tasks; agents beat human baseline by 26.5%
+  - title: 'BuilderBench: Building Blocks for Agents'
+    arxiv: '2510.06288'
+    url: https://arxiv.org/abs/2510.06288
+    venue: Oct 2025
+    why: 42 target structures; open exploration; no external supervision
+  - title: FLOWER
+    source: arXiv:2509.04996
+    why: 'Current CALVIN SOTA: 4.53/5 on ABC tasks'
+  full_references_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B12_*/references/papers.json
+nlpcc_tracks:
+- track: T1
+  name: AI/CS Reasoning & Engineering
+  url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/nlpcc/T1
+private_hidden_reference_strategy:
+  policy: organizer_only
+  public_package_includes_hidden_reference: false
+  public_catalog_discloses_private_paths: false
+  score_release: aggregate_only
+  notes:
+  - Hidden answers, labels, canaries, and private evaluator internals are never packaged.
+  - Official scores must record hidden_reference_version without disclosing hidden content.
+score_card:
+  required_fields:
+  - team_id
+  - system_name
+  - direction
+  - track
+  - benchmark_package_version
+  - evaluator_version
+  - hidden_reference_version
+  - docker_image_digest
+  - data_split_version
+  - benchmark_score
+  - paper_score
+  - cas_score
+  - final_score
+  - security_verdict
+  - license_verdict
+  - used_reference_scope
+  - official_score
+  - created_at
+  track_scoring:
+    paper_track: Final_A = 0.0 * S_benchmark + 1.0 * S_paper
+    benchmark_track: Final_B = 0.7 * S_benchmark + 0.3 * S_paper
+    cas: integrity gate, not bonus
+risk_flags:
+- hidden_reference_not_public
+- not_release_ready
+- registry_promotion_pending
+- simulator_dependency_pending
+risk_notes:
+- This catalog entry is not release_ready.
+- Download URL, checksum, and size point to the published GitHub release asset.
+- Official scores require organizer-side replay against private hidden references.
+recommended_when: Use this entry for AISB public-dev integration, local agent workflow checks, and catalog
+  UI validation.
+not_recommended_when: Do not treat this entry as an official release or public leaderboard claim until
+  release gates pass.
+image_path: ../image/aisb.b12.embodied_ai.svg

package/AISB/catalog/aisb.b2.agent_systems.yaml

@@ -0,0 +1,229 @@
+schema_version: 1
+id: aisb.b2.agent_systems
+name: AISB B2 Agent Systems
+version: 0.3.0
+catalog_version: 0.1.0
+one_line: tau2-bench airline public-dev mini package for tool-use task success validation.
+task_description: AISB B2 Agent Systems is an AISB BenchStore catalog pilot entry. The package is intended
+  for public-dev workspace initialization, public evaluator checks, and paper/benchmark-track submission
+  scaffolding. Official leaderboard scores require organizer-side hidden replay.
+homepage: https://aisb.deepscientist.cc
+localized:
+  zh:
+    name: AISB B2 智能体系统
+    one_line: 用于工具调用任务成功率验证的 tau2-bench airline public-dev mini 包。
+    task_description: AISB B2 智能体系统 是 AISB BenchStore catalog pilot 条目。公开包只用于 public-dev 工作区初始化、公开 evaluator
+      检查和论文/benchmark track 提交脚手架；正式 leaderboard 分数必须由组织方 hidden replay 产生。
+capability_tags:
+- agent_systems
+- tool_use
+- task_success
+aisb_direction: B2
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: benchmark
+benchmark_mode: agent_task_success
+requires_execution: true
+requires_paper: true
+integrity_level: hidden_reference
+snapshot_status: partial
+support_level: preview
+discovery:
+  collection: AISB
+  collection_priority: 100
+  recommendation_weight: 500
+  featured: false
+  featured_reason: null
+display:
+  placement: grid
+  card_size: m
+  badge: AISB
+  accent_priority: normal
+official_links:
+  homepage: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  github: https://github.com/ResearAI/NLPCC-2026-Task9-AISB
+  benchmark_overview: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B2_AgentSystems
+release_gate:
+  release_ready: false
+  status: registry_pending
+  full_real_replay: passed
+  reason: Public-dev and scorer-private replay paths exist locally, but final registry promotion, release
+    assets, and license review remain pending.
+  next_required_step: Validate the published BenchStore release asset on a clean host, then complete
+    registry and license review.
+primary_outputs:
+- results.json
+- paper/paper.pdf
+- paper/claims.json
+- logs/experiment_log.jsonl
+- logs/iterations.jsonl
+cost_band: medium
+time_band: 1-4h
+difficulty: hard
+data_access: public
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 5
+    gpu_count: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+source:
+  benchmark_dir: benchmarks/aisb/B2_AgentSystems
+  bench_yaml: benchmarks/aisb/B2_AgentSystems/bench.yaml
+  source_repo: ResearAI/NLPCC-2026-Task9-AISB
+  source_commit: b6b7527a226ca7328697156608571f2cb7dcda3a
+download:
+  provider: github_release
+  repo: ResearAI/NLPCC-2026-Task9-AISB
+  tag: aisb-v0.0.1
+  asset_name: aisb.b2.agent_systems-v0.3.0.zip
+  url: https://github.com/giao-123-sun/DeepScientist/releases/download/aisb-v0.0.1/aisb.b2.agent_systems-v0.3.0.zip
+  source_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B2_AgentSystems
+  archive_type: zip
+  local_dir_name: aisb.b2.agent_systems
+  sha256: 640fe257640743a51c7e65bede73c9ea19be97b4d85b232689c23d477d051216
+  size_bytes: 25925
+  published_at: '2026-04-30T00:00:00Z'
+  asset_status: published
+dataset_download:
+  primary_method: packaged_public_dev
+  sources:
+  - kind: public_dev
+    access: public
+    url: null
+    note: Public-dev files are included in the published BenchStore package.
+  - kind: hidden_reference
+    access: private
+    url: null
+    note: Organizer-only hidden references are excluded from catalog packages.
+credential_requirements:
+  mode: none
+  notes:
+  - No credential is required for packaged public-dev evaluation.
+  - External model APIs are optional participant choices and not bundled.
+environment:
+  python: '>=3.11'
+  docker: true
+  key_packages:
+  - pytest
+  - PyYAML
+  notes:
+  - Use docker/Dockerfile when available for official replay parity.
+  - Do not copy organizer-private hidden references into participant workspaces.
+launch_profiles:
+- id: b2_public_dev_smoke
+  label: B2 public-dev smoke test
+  description: Initialize an AISB public-dev workspace, validate submission artifacts, and run the public
+    evaluator.
+  startup_contract:
+  - Use only files inside the installed benchmark workspace and the generated submission directory.
+  - Treat hidden references as organizer-only and unavailable to the agent.
+  - Produce benchmark outputs, paper artifacts, claims, and logs before packaging.
+  commands:
+  - python scripts/agent_tools.py workspace init B2 --dest .work/benchstore/B2
+  - python scripts/agent_tools.py evaluate B2 --bench-dir .work/benchstore/B2 --submission .work/benchstore/B2/submission
+  - python scripts/agent_tools.py submission validate .work/benchstore/B2/submission
+  - python scripts/agent_tools.py submission package .work/benchstore/B2/submission --output .work/benchstore/B2/submission.zip
+  - python scripts/agent_tools.py submission replay .work/benchstore/B2/submission --track B2
+  one_click_prompt: Start AISB B2 from the installed BenchStore package. Initialize the public-dev workspace,
+    inspect AGENT.md and README.md, run the baseline or your method, evaluate locally, validate/package
+    the submission, and record any release-gate risks in logs/experiment_log.jsonl.
+paper:
+  venue: AISB
+  year: 2026
+  track_fit:
+  - paper_track
+  - benchmark_track
+  requires_claims_json: true
+  url: null
+benchmark_track:
+  metric: task_success_rate
+  benchmark_source: tau2-bench airline mini
+  packaged_data_status: real tau2-bench airline mini split
+  official_score_requires_hidden_replay: true
+benchmarks:
+  primary:
+    name: tau2-bench airline
+    source: github.com/sierra-research/tau2-bench
+    tasks: 2
+    metric: task_success_rate
+    sota: no public absolute scores
+    public_score: 0.0
+  alternatives:
+  - name: GAIA
+    source: huggingface.co/datasets/gaia-benchmark/GAIA
+    note: gated; secondary candidate
+  - name: Terminal-Bench
+    source: github.com/terminal-bench
+    note: CLI agent benchmark; future candidate
+papers:
+  highlights:
+  - title: 'tau2-bench: A Benchmark for Tool-Agent-User Interaction'
+    source: github.com/sierra-research/tau2-bench
+    why: Primary benchmark; airline domain used for public-dev
+  - title: Recursive Multi-Agent Systems
+    arxiv: '2604.25917'
+    url: https://arxiv.org/abs/2604.25917
+    venue: April 2026
+    why: +8.3% avg accuracy; 1.2x-2.4x inference speedup
+  - title: 'DR3-Eval: Towards Realistic and Reproducible Deep Research Evaluation'
+    arxiv: '2604.14683'
+    url: https://arxiv.org/abs/2604.14683
+    venue: April 2026
+    why: 5-dimension deep research evaluation framework
+  full_references_url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/aisb/B2_*/references/papers.json
+nlpcc_tracks:
+- track: T1
+  name: AI/CS Reasoning & Engineering
+  url: https://github.com/ResearAI/NLPCC-2026-Task9-AISB/tree/main/benchmarks/nlpcc/T1
+private_hidden_reference_strategy:
+  policy: organizer_only
+  public_package_includes_hidden_reference: false
+  public_catalog_discloses_private_paths: false
+  score_release: aggregate_only
+  notes:
+  - Hidden answers, labels, canaries, and private evaluator internals are never packaged.
+  - Official scores must record hidden_reference_version without disclosing hidden content.
+score_card:
+  required_fields:
+  - team_id
+  - system_name
+  - direction
+  - track
+  - benchmark_package_version
+  - evaluator_version
+  - hidden_reference_version
+  - docker_image_digest
+  - data_split_version
+  - benchmark_score
+  - paper_score
+  - cas_score
+  - final_score
+  - security_verdict
+  - license_verdict
+  - used_reference_scope
+  - official_score
+  - created_at
+  track_scoring:
+    paper_track: Final_A = 0.0 * S_benchmark + 1.0 * S_paper
+    benchmark_track: Final_B = 0.7 * S_benchmark + 0.3 * S_paper
+    cas: integrity gate, not bonus
+risk_flags:
+- hidden_reference_not_public
+- not_release_ready
+- registry_promotion_pending
+risk_notes:
+- This catalog entry is not release_ready.
+- Download URL, checksum, and size point to the published GitHub release asset.
+- Official scores require organizer-side replay against private hidden references.
+recommended_when: Use this entry for AISB public-dev integration, local agent workflow checks, and catalog
+  UI validation.
+not_recommended_when: Do not treat this entry as an official release or public leaderboard claim until
+  release gates pass.
+image_path: ../image/aisb.b2.agent_systems.svg