Datasets:

yuxianglai117
/

CLASHarness

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
record_type: string
task_id: string
system_prompt: string
run_config: struct<provider: string, model: string, history_mode: string, multimodal_input_mode: string, multimo (... 286 chars omitted)
  child 0, provider: string
  child 1, model: string
  child 2, history_mode: string
  child 3, multimodal_input_mode: string
  child 4, multimodal_input_mode_effective: string
  child 5, tool_mode: string
  child 6, tool_schema_mode: string
  child 7, reasoning_effort: string
  child 8, reasoning_effort_effective: string
  child 9, chat_template_kwargs: string
  child 10, replay_chat_reasoning: string
  child 11, native_tool_choice: string
  child 12, max_image_edge: string
  child 13, gemini_media_resolution: string
step: int64
attempt_index: int64
provider: string
model: string
input: struct<system_prompt_sha1: string, last_user_text: string, image_refs: list<item: struct<mime: strin (... 81 chars omitted)
  child 0, system_prompt_sha1: string
  child 1, last_user_text: string
  child 2, image_refs: list<item: struct<mime: string, bytes_len: int64, sha1: string, dims: list<item: int64>>>
      child 0, item: struct<mime: string, bytes_len: int64, sha1: string, dims: list<item: int64>>
          child 0, mime: string
          child 1, bytes_len: int64
          child 2, sha1: string
          child 3, dims: list<item: int64>
              child 0, item: int64
  child 3, history_len: int64
response: struct<answer_text: string, thinking_text: string, finish_reason: string, finish_mes
...
ring, stde (... 481 chars omitted)
  child 0, item: struct<step: int64, action: string, model_debug: list<item: string>, stdout: string, stderr: null, e (... 469 chars omitted)
      child 0, step: int64
      child 1, action: string
      child 2, model_debug: list<item: string>
          child 0, item: string
      child 3, stdout: string
      child 4, stderr: null
      child 5, executed_action: string
      child 6, clean_stdout: string
      child 7, clean_stderr: string
      child 8, error_tags: list<item: null>
          child 0, item: null
      child 9, compat_status: string
      child 10, execution_trace: list<item: null>
          child 0, item: null
      child 11, exit_code: int64
      child 12, reward: double
      child 13, done: bool
      child 14, step_evaluation: struct<success: bool, current_score: double, score_delta: double, passed_checks: list<item: string>, (... 35 chars omitted)
          child 0, success: bool
          child 1, current_score: double
          child 2, score_delta: double
          child 3, passed_checks: list<item: string>
              child 0, item: string
          child 4, failed_checks: list<item: string>
              child 0, item: string
      child 15, current_score: double
      child 16, score_delta: double
      child 17, passed_checks: list<item: string>
          child 0, item: string
      child 18, failed_checks: list<item: string>
          child 0, item: string
score: double
compact_fallback: bool
agent: string
to
{'task_id': Value('string'), 'agent': Value('string'), 'success': Value('bool'), 'score': Value('float64'), 'steps': Value('int64'), 'duration_s': Value('float64'), 'trajectory': List({'step': Value('int64'), 'action': Value('string'), 'model_debug': List(Value('string')), 'stdout': Value('string'), 'stderr': Value('null'), 'executed_action': Value('string'), 'clean_stdout': Value('string'), 'clean_stderr': Value('string'), 'error_tags': List(Value('null')), 'compat_status': Value('string'), 'execution_trace': List(Value('null')), 'exit_code': Value('int64'), 'reward': Value('float64'), 'done': Value('bool'), 'step_evaluation': {'success': Value('bool'), 'current_score': Value('float64'), 'score_delta': Value('float64'), 'passed_checks': List(Value('string')), 'failed_checks': List(Value('string'))}, 'current_score': Value('float64'), 'score_delta': Value('float64'), 'passed_checks': List(Value('string')), 'failed_checks': List(Value('string'))}), 'final_response': Value('string'), 'error': Value('string'), 'error_type': Value('string'), 'provider_impacted': Value('bool'), 'compact_fallback': Value('bool'), 'stop_reason': Value('string'), 'date_anchor_mismatch': Value('bool'), 'mode': Value('string'), 'metadata': {'mode': Value('string'), 'domains': List(Value('string')), 'difficulty': Value('int64'), 'scenario_slug': Value('string'), 'round_id': Value('string'), 'round_type': Value('string'), 'generator_id': Value('string'), 'template_id': Value('string'), 'instruction': Value('string'), 'canonical_instruction': Value('string'), 'check_results': List({'name': Value('string'), 'passed': Value('bool'), 'score': Value('float64'), 'weight': Value('float64'), 'message': Value('string'), 'role': Value('string'), 'ok': Value('bool'), 'error': Value('string'), 'judge_invoked': Value('bool')})}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              record_type: string
              task_id: string
              system_prompt: string
              run_config: struct<provider: string, model: string, history_mode: string, multimodal_input_mode: string, multimo (... 286 chars omitted)
                child 0, provider: string
                child 1, model: string
                child 2, history_mode: string
                child 3, multimodal_input_mode: string
                child 4, multimodal_input_mode_effective: string
                child 5, tool_mode: string
                child 6, tool_schema_mode: string
                child 7, reasoning_effort: string
                child 8, reasoning_effort_effective: string
                child 9, chat_template_kwargs: string
                child 10, replay_chat_reasoning: string
                child 11, native_tool_choice: string
                child 12, max_image_edge: string
                child 13, gemini_media_resolution: string
              step: int64
              attempt_index: int64
              provider: string
              model: string
              input: struct<system_prompt_sha1: string, last_user_text: string, image_refs: list<item: struct<mime: strin (... 81 chars omitted)
                child 0, system_prompt_sha1: string
                child 1, last_user_text: string
                child 2, image_refs: list<item: struct<mime: string, bytes_len: int64, sha1: string, dims: list<item: int64>>>
                    child 0, item: struct<mime: string, bytes_len: int64, sha1: string, dims: list<item: int64>>
                        child 0, mime: string
                        child 1, bytes_len: int64
                        child 2, sha1: string
                        child 3, dims: list<item: int64>
                            child 0, item: int64
                child 3, history_len: int64
              response: struct<answer_text: string, thinking_text: string, finish_reason: string, finish_mes
              ...
              ring, stde (... 481 chars omitted)
                child 0, item: struct<step: int64, action: string, model_debug: list<item: string>, stdout: string, stderr: null, e (... 469 chars omitted)
                    child 0, step: int64
                    child 1, action: string
                    child 2, model_debug: list<item: string>
                        child 0, item: string
                    child 3, stdout: string
                    child 4, stderr: null
                    child 5, executed_action: string
                    child 6, clean_stdout: string
                    child 7, clean_stderr: string
                    child 8, error_tags: list<item: null>
                        child 0, item: null
                    child 9, compat_status: string
                    child 10, execution_trace: list<item: null>
                        child 0, item: null
                    child 11, exit_code: int64
                    child 12, reward: double
                    child 13, done: bool
                    child 14, step_evaluation: struct<success: bool, current_score: double, score_delta: double, passed_checks: list<item: string>, (... 35 chars omitted)
                        child 0, success: bool
                        child 1, current_score: double
                        child 2, score_delta: double
                        child 3, passed_checks: list<item: string>
                            child 0, item: string
                        child 4, failed_checks: list<item: string>
                            child 0, item: string
                    child 15, current_score: double
                    child 16, score_delta: double
                    child 17, passed_checks: list<item: string>
                        child 0, item: string
                    child 18, failed_checks: list<item: string>
                        child 0, item: string
              score: double
              compact_fallback: bool
              agent: string
              to
              {'task_id': Value('string'), 'agent': Value('string'), 'success': Value('bool'), 'score': Value('float64'), 'steps': Value('int64'), 'duration_s': Value('float64'), 'trajectory': List({'step': Value('int64'), 'action': Value('string'), 'model_debug': List(Value('string')), 'stdout': Value('string'), 'stderr': Value('null'), 'executed_action': Value('string'), 'clean_stdout': Value('string'), 'clean_stderr': Value('string'), 'error_tags': List(Value('null')), 'compat_status': Value('string'), 'execution_trace': List(Value('null')), 'exit_code': Value('int64'), 'reward': Value('float64'), 'done': Value('bool'), 'step_evaluation': {'success': Value('bool'), 'current_score': Value('float64'), 'score_delta': Value('float64'), 'passed_checks': List(Value('string')), 'failed_checks': List(Value('string'))}, 'current_score': Value('float64'), 'score_delta': Value('float64'), 'passed_checks': List(Value('string')), 'failed_checks': List(Value('string'))}), 'final_response': Value('string'), 'error': Value('string'), 'error_type': Value('string'), 'provider_impacted': Value('bool'), 'compact_fallback': Value('bool'), 'stop_reason': Value('string'), 'date_anchor_mismatch': Value('bool'), 'mode': Value('string'), 'metadata': {'mode': Value('string'), 'domains': List(Value('string')), 'difficulty': Value('int64'), 'scenario_slug': Value('string'), 'round_id': Value('string'), 'round_type': Value('string'), 'generator_id': Value('string'), 'template_id': Value('string'), 'instruction': Value('string'), 'canonical_instruction': Value('string'), 'check_results': List({'name': Value('string'), 'passed': Value('bool'), 'score': Value('float64'), 'weight': Value('float64'), 'message': Value('string'), 'role': Value('string'), 'ok': Value('bool'), 'error': Value('string'), 'judge_invoked': Value('bool')})}}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

🚀 Dataset Overview & Highlights
⚡ Quick Start: Downloading & Running the Benchmark
🏆 Benchmark Leaderboard (600 Episodes, High Effort)
📦 Dataset Components, Layout & Record Schemas
⚖️ The Three Conflict Dimensions & Generation Pipeline
🔍 Provenance, Audit Notes & Known Issues
📜 License & Citation

CLASHarness Corpus

Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline

The official 600-episode multi-agent conflict evaluation benchmark and recorded 15-model trajectory corpus for CLASHarness (arXiv:2610.19128).

📄 Paper (arXiv) • 💻 GitHub Repository • 🚀 Overview • ⚡ Quick Start & Evaluation • 🏆 Leaderboard • 📦 Dataset Structure & Schema • ⚖️ Conflict Dimensions • 📜 Citation

🚀 Dataset Overview & Highlights

This repository hosts the compiled task specifications (tasks-v8), full-benchmark episode index (splits-v8), and recorded frontier model evaluation trajectories (eval-v8) for CLASHarness — a controllable generation and evaluation harness for multi-agent coordination under conflicting, time-sensitive, multimodal evidence.

🔗 Paper, Code & Evaluation Harness:

Paper: CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline (arXiv:2610.19128)

GitHub Repository: github.com/google-research/CLASHarness (8-stage scenario generator S1–S8, interactive multi-agent CLI runtime, 20 conflict seed recipes, and analysis suite)

Verifiable by Construction ($v^\star$ + Live Sandbox Execution) ◈ Every benchmark episode pre-commits an atomic ground-truth decision value ($v^\star$), executes a reference solver trajectory inside a stateful multi-agent CLI sandbox, and compiles dual-track grading rubrics ($\Phi = \Phi_{\text{det}} \cup \Phi_{\text{sem}}$) directly from recorded side effects.
Full 600-Episode Evaluation (--split total) ◈ Built from 20 conflict seeds across 15 conflict archetypes (3 multi-round scenario versions per seed × 10 rounds per version = 600 episodes). All 600 episodes are evaluated directly without a train/test split, stratified by reference trajectory length into Easy (236), Mid (224), and Hard (140).
Multimodal & Temporal Conflict Grounding ◈ Tests whether agents cross-examine conflicting peer claims against rendered UI screenshots, dashboards, invoices, and real on-disk state rather than blindly trusting verbal messages.
15-Model Frontier Trajectory Corpus ◈ Includes full step-by-step execution transcripts, tool calls, environment mutations, and check-level grades across frontier proprietary APIs (GPT-6 Astra, Claude Opus 5, GPT-5.6 Sol, Claude Opus 4.8, Gemini 3.7 Flash) and open-weight models (Kimi K3, DeepSeek V4 Pro, GLM-5.2 FP8).

⚡ Quick Start: Downloading & Running the Benchmark

Use the downloader in the CLASHarness GitHub repository to fetch the components directly into the data/ directory structure expected by the evaluation harness and analysis scripts.

1. Clone the Code Repository & Download the Dataset

git clone https://github.com/google-research/CLASHarness.git
cd CLASHarness

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]" huggingface_hub

# Inspect available dataset components
python tools/download_data.py --list

# Download the default release set (tasks-v8 + splits-v8 + eval-v8, 8.5 GB) into ./data/
python tools/download_data.py --repo yuxianglai117/CLASHarness

# Verify SHA-256 integrity against CHECKSUMS.txt
python tools/checksum_corpus.py --root data --verify CHECKSUMS.txt

2. Evaluate an Agent on All 600 Benchmark Episodes

If you only need to evaluate your own model on the 600 benchmark episodes, you only need tasks-v8 and splits-v8 (5.7 GB):

# Fetch the 600 compiled task specs and episode index
python tools/download_data.py --repo yuxianglai117/CLASHarness --component tasks-v8 --component splits-v8

# Run evaluation across all 600 episodes (--split total)
export GEMINI_API_KEY="your-api-key"

python examples/train_and_eval.py \
  --agent llm \
  --llm-provider gemini --model gemini-3.6-flash \
  --llm-judge-provider gemini --llm-judge-model gemini-3.6-flash \
  --task-data-dir data/multiagent_conflict_data_v8_Gemini_3_6_flash \
  --split total \
  --mode multi \
  --max-steps 20 \
  --parallel-workers 10 \
  --collect-data \
  --out-dir /tmp/clasharness-eval

3. Reproduce Paper Tables & Figures from Recorded Trajectories (`eval-v8`)

python analysis/table1_v8.py
python analysis/difficulty_bands.py
python analysis/mcnemar_v8.py
python analysis/cost.py

🏆 Benchmark Leaderboard (600 Episodes, High Effort)

Model Standings Across Difficulty Bands

Model	All Pass (%)	Pass Rate by Difficulty (%)			Mean Score (%)	Avg. Cost ($ / Tokens)
Model	All Pass (%)	Easy	Mid	Hard	Mean Score (%)	Avg. Cost ($ / Tokens)
GPT-6 Astra	28.7_±1.9	35.2_±3.1	28.1_±3.1	18.6_±3.3	66.8_±1.2	$3,003 (663M)
Claude Opus 5	25.2_±1.8	32.2_±3.0	25.0_±2.9	13.6_±2.9	64.2_±1.2	$2,631 (511M)
GPT-5.6 Sol	24.5_±1.7	31.4_±3.0	24.1_±2.9	13.6_±2.9	63.3_±1.2	$2,165 (454M)
Claude Opus 4.8	22.8_±1.7	30.1_±3.0	23.7_±2.8	9.3_±2.4	62.1_±1.2	$2,933 (571M)
Kimi K3	21.8_±1.7	26.3_±2.8	23.2_±2.9	12.1_±2.7	55.9_±1.4	$2,173 (706M)
Claude Sonnet 5	20.8_±1.7	28.0_±2.9	21.4_±2.7	7.9_±2.3	59.6_±1.3	$1,031 (497M)
Gemini 3.7 Flash	20.3_±1.6	26.7_±2.9	21.9_±2.8	7.1_±2.2	58.1_±1.3	$307 (395M)
GPT-5.6 Terra	20.0_±1.6	27.1_±2.9	20.1_±2.7	7.9_±2.3	60.2_±1.2	$731 (382M)
GPT-5.6 Luna	19.8_±1.6	23.7_±2.7	22.3_±2.8	9.3_±2.5	60.4_±1.2	$86 (450M)
DeepSeek V4 Pro^†	19.8_±1.6	24.2_±2.8	20.5_±2.7	11.4_±2.7	58.1_±1.3	$32 (70M)
Gemini 3.6 Flash	18.5_±1.6	24.6_±2.8	21.4_±2.7	3.6_±1.6	58.6_±1.3	$565 (363M)
GLM-5.2 FP8^†	14.7_±1.4	18.2_±2.5	16.5_±2.5	5.7_±2.0	48.7_±1.3	$120 (82M)
Gemini 3.1 Pro	12.0_±1.3	15.7_±2.4	12.1_±2.2	5.7_±2.0	48.7_±1.3	$578 (280M)
DeepSeek V4 Flash^†	8.5_±1.1	8.5_±1.8	11.2_±2.1	4.3_±1.7	37.5_±1.3	N/A (95M)
Gemini 3.5 Flash	3.0_±0.7	3.8_±1.2	2.7_±1.1	2.1_±1.2	25.9_±1.1	$712 (445M)

_{^†Text-only models evaluated using OCR transcripts of visual assets. Task difficulty is stratified by ground-truth reference trajectory length across all 600 episodes (Easy: 236 rounds, ≤ 9 steps; Mid: 224 rounds, 10–11 steps; Hard: 140 rounds, ≥ 12 steps).}

📦 Dataset Components, Layout & Record Schemas

1. Component Manifest

Component	Size	Description
`tasks-v8` ★	5.7 GB	Compiled task specifications (`specs.json`) for all 600 episodes of the primary benchmark (generated with `gemini-3.6-flash`). Required to run evaluation via `examples/train_and_eval.py`.
`splits-v8` ★	58 KB	Episode index files (`datasets/total.txt` listing all 600 evaluated episodes). Required by `examples/train_and_eval.py --split total`.
`eval-v8` ★	2.8 GB	Primary 20-step evaluation panel: 14 models (10 API models swept across `low`/`medium`/`high` reasoning effort + 4 open-weight models at `high`).
`eval-sonnet5`	1.9 GB	Cross-generator ablation trajectories (Arm B, generated by `claude-sonnet-5`).
`tasks-sonnet5`	7.1 GB	Compiled task specs for Arm B (`claude-sonnet-5`).
`splits-sonnet5`	58 KB	Episode index files (`datasets/total.txt`) for Arm B.
`eval-budget30`	440 MB	Step-budget ablation trajectories at 30 steps.
`eval-budget40`	577 MB	Step-budget ablation trajectories at 40 steps.
`eval-blind`	219 MB	Modality ablation trajectories (text-only arm with visual assets withheld).
`scenarios-v8`	12.3 GB	Raw materialized scenario packages for the primary corpus.
`scenarios-sonnet5`	15.3 GB	Raw materialized scenario packages for Arm B.

★ = Default Release Set (8.5 GB): Fetched by default via python tools/download_data.py --repo yuxianglai117/CLASHarness. Sufficient to run the full 600-episode benchmark and reproduce all primary paper tables.

2. Directory Layout

multiagent_conflict_data_v8_Gemini_3_6_flash/
  tasks/<task_id>/specs.json            # tasks-v8 (600 compiled task specifications + embedded base64 images)
  datasets/total.txt                    # splits-v8 (600-episode index read by --split total)
eval_results/
  data_3_6_flash_v8/                    # eval-v8 (recorded model evaluation trajectories)
    aligned_setting_{low,medium,high}/  # 10 API models across reasoning efforts
    open_source_high/                   # 4 open-weight models at high effort
      <Vendor>/<model>/<run>/<scenario>.jsonl

Each model directory holds run directories (r01, r02, ...). analysis/ reads r01 throughout, with one recorded exception — aligned_setting_high/Gemini/3_7_flash reads r02 — declared as RUN_OVERRIDE in analysis/corpus.py.

3. Record Schemas

Task Specifications (`multiagent_conflict_data_v8_Gemini_3_6_flash/tasks/<task_id>/specs.json`)

Field	Description
`instruction`, `canonical_instruction`	Natural-language coordinator prompt and canonical task specification
`initial_state`	Initial workspace files, communication channel history, active locks/approvals, and base64-encoded visual assets (`__bytes_b64__`)
`visible_constraints`, `hidden_constraints`	Explicit and latent operational constraints governing conflict resolution
`decision_requirements`	Pre-committed ground-truth decision requirements ($v^\star$) and authoritative source provenance
`ground_truth`	Reference solver trajectory (`reference_solution`) and compiled dual-track evaluation checks (`evaluation_checks`: deterministic state checks $\Phi_{\text{det}}$ + semantic LLM-judge checks $\Phi_{\text{sem}}$)
`domains`, `difficulty`, `template_id`	Domain tags, difficulty band (`easy` / `medium` / `hard`), and seed archetype metadata

Evaluation Trajectories (`eval_results/**/*.jsonl`, one JSON object per episode)

Field	Description
`task_id`, `agent`	Episode identifier (`<scenario>_v0X_r0Y`) and evaluated model identifier
`success`, `score`	All-or-nothing episode pass verdict (`bool`) and fraction of satisfied checks in `[0, 1]`
`steps`, `duration_s`	Total CLI actions executed and wall-clock duration in seconds
`trajectory[]`	Step-by-step trace: `action`, `executed_action`, `stdout`, `stderr`, `exit_code`, `reward`, `current_score`, `passed_checks`, `failed_checks`
`final_response`	The agent's closing summary message
`metadata`	Includes `difficulty`, `scenario_slug`, `round_id`, `instruction`, and per-check grading breakdown (`check_results`)

4. Benchmark Specifications

Dimension	Specification	Description
Conflict Seeds	20	Diverse domains across system admin, finance, cloud deployment, healthcare, and IAM
Scenario Packages	60	3 multi-round scenario packages generated per seed
Evaluation Episodes	600	10 evaluation rounds per package (evaluated in full via `--split total`)
Difficulty Bands	236 / 224 / 140	Easy (≤ 9 steps) / Mid (10–11 steps) / Hard (≥ 12 steps) across all 600 episodes
Conflict Plugins	23	Covering challenges, coordination protocols, and subagent interactions
State Protocols	7	`approval`, `booking`, `containment`, `deploy`, `lock`, `pairing`, `review`
Execution Modes	5	`multi` (default), `mock`, `martin`, `real`, `hybrid`
Evaluated Models	15	Frontier proprietary APIs + open-weight vision-language models

⚖️ The Three Conflict Dimensions & Generation Pipeline

The Three Conflict Dimensions in CLASHarness

Dimension	Operational Failure Mode	Coordinator Requirement
◈ Temporal Staleness	A subagent claims task completion at `T = 0`, but subsequent telemetry or logs show a regression.	Cross-check real on-disk state and prefer recent authoritative ground truth over stale verbal claims.
◈ Cross-Modal Discrepancy	A subagent's verbal text report directly contradicts a rendered chart, status dashboard, or UI screenshot.	Visually inspect and cross-examine text assertions against rendered visual evidence before acting.
◈ Actionable Delay	Operations require waiting for asynchronous delays to elapse and verifying physical file mutations.	Handle asynchronous latency and verify physical state mutations rather than prematurely declaring completion.

CLASHarness Hierarchical Generation Pipeline

For full architectural details on the 3-tier, 8-stage (S1–S8) verifiable generation pipeline, see Multi-Agent_GENERATION_PIPELINE.md in the GitHub repository.

🔍 Provenance, Audit Notes & Known Issues

Every scenario, document, chart, and conflict in this corpus is 100% synthetic — generated by gemini-3.6-flash (primary v8 corpus) or claude-sonnet-5 (Arm B) and materialized by the pipeline in the CLASHarness GitHub repository. No real user data, company data, or scraped content is included.

Post-Submission Audit Notes, Known Issues & Diagnostic Redaction Details (Click to Expand)

The paper's numbers are computed on the corpus exactly as published here. The items below were audited after submission; none of them altered any file in this dataset, and supplementary audit manifests are shipped under analysis/ in the GitHub repository so any view can be recomputed.

1. Eighteen Unwinnable Pre-Repair Rounds

Eighteen of the 600 rounds were unwinnable when these runs were scored due to a materializer visibility defect that graded them on a value absent from solver-visible materials (across 18 rounds × 10 API models × 3 efforts = 540 round-runs, there is 1 pass). The paper reports the full 600 with those rows retained because dropping them rescales every model by the same 582/600 factor and changes no ranking or statistical test. The defect is fixed in the generator code, and analysis/dead_rounds_exclusion.json lists the 18 rounds.

2. Post-Submission Check Fairness Audit

A check is fair if the instruction asks for it, or if it can reasonably be inferred from the materials the player is shown. A post-submission audit labelled all 1,194 required message criteria (Claude Opus 5.5, three votes each, with quoted evidence) and all 515 required protocol checks (Claude Opus 5.5 and, independently, Gemini 3.1 Pro). Re-grading the recorded high-effort episodes of 13 of the 14 panel models with flagged checks demoted to diagnostic yields:

Checks demoted	Checks	Claude Opus 5 (%)	Kendall τ-b vs. published	Opus 5 vs. Opus 4.8, p
None (as published)	0	25.2	1.00	0.120
Message criteria whose required value conflicts with the authoritative material	200	27.2	0.99	0.057
Protocol checks both judges call unasked or contradicted	106	29.5	0.91	0.005
Both of the above	306	31.7	0.94	0.002
Protocol checks Opus 5.5 calls unasked, contradicted or ambiguous, plus message criteria it calls unsupported or unasked	485	36.3	0.88	0.004
Upper bound, not an estimate: the previous row plus message criteria it calls over-specified or ambiguous	775	42.7	0.68	0.031

In every row the top three (Claude Opus 5, GPT-5.6 Sol, Claude Opus 4.8) and bottom three (GLM-5.2 FP8, Gemini 3.1 Pro, Gemini 3.5 Flash) of the 13 keep their places. analysis/v8_check_fairness_audit.json lists all 775 flagged checks.

3. Reference Replay & Seeding Notes

Checks the reference solution fails (20 of 3,760): Replaying all 600 reference solutions twice with the v8 judge (gemini-3.6-flash) passed 100% of deterministic checks and flagged 20 semantic checks across 18 rounds (removing them changes one Claude Opus 5 round: 25.2% -> 25.3%).
Stale cross-round file seeding (50 of 600 rounds): Each round seeds from scenario snapshots rather than earlier rounds' reference mutations; reference replay found no grading failures caused by this.
Arm B (tasks-sonnet5, eval-sonnet5): analysis/sonnet5_gold_replay_audit.json lists 74 required checks across 59 rounds flagged by reference replay.
Empty replies on Claude Opus 5 (8 of 1,800 episodes): Recorded with "stop_reason": "generation_stall_empty_reply" and "error_type": "provider_failure" in payment_chargeback_dispute (at most 0.5 points at high effort).

4. Diagnostic Error Redaction

Failed episodes that encountered infrastructure transport errors had internal cluster addresses, workspace paths, or numeric quota IDs in their .error strings replaced with placeholders (<cluster-address-redacted>, <internal-diagnostic-trimmed>, input-tpm:<quota-id-redacted>). No successful episode, trajectory, input, or score field was touched, and no reported metric depends on .error text.

📜 License & Citation

Dataset License: Released under Creative Commons Attribution 4.0 International (CC BY 4.0).
Code & Harness License: Released under Apache License 2.0 at github.com/google-research/CLASHarness.

If you use the CLASHarness dataset or codebase in your research, please cite:

@article{lai2026clasharness,
  title   = {CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline},
  author  = {Lai, Yuxiang and Zhong, Jike and Liu, Chengzhi and Li, Ming and
             Wang, Haonan and Yang, Xiaofeng and Zhao, Liang and Tang, Youbao and
             Fan, Ruixi},
  journal = {arXiv preprint arXiv:2610.19128},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.19128}
}

Downloads last month: 70

Total file size:

8.46 GB

Paper for yuxianglai117/CLASHarness

CLASHarness: Benchmarking Multi-Agent Conflict in Computer Use via a Controllable and Verifiable Generation Pipeline

Paper • 2610.19128 • Published 4 days ago • 30