JosephAA/toporisk-evaluation-artifacts

TopoRisk Evaluation Artifacts

English | 简体中文

TopoRisk studies how a limited verification budget should be allocated across an agentic workflow graph. Instead of ranking steps only by their local failure probability, the scheduler estimates how an error can reach terminal outputs and recomputes marginal value after every selected audit.

This repository is an artifact-first research preview. It contains code, aggregate measurements, and figures. The manuscript and its LaTeX sources are intentionally not included. The work has not undergone peer review.

The canonical implementation, tests, and GitHub Actions workflow are available in the TopoRisk GitHub repository.

What is included

• deterministic simulation and analysis scripts;
• controlled synthetic-DAG results and ablations;
• aggregate fixed-trace allocation summaries;
• aggregate online-pilot outcomes and verifier cost statistics; and
• figures generated from the packaged aggregate results.

What is excluded

• the manuscript PDF and LaTeX source;
• API keys, .env files, provider request identifiers, and account data;
• raw prompts, model responses, and conversational traces; and
• upstream benchmark records whose redistribution terms have not been cleared.

Task identifiers in the online summary are local study labels, not provider or customer identifiers.

Results at a glance

Evaluation	Observation
Controlled DAGs, 10% audit budget	Mean corrupted-sink fraction: 0.261 for local risk and 0.147 for TopoRisk
Controlled DAGs, 20% audit budget	Mean corrupted-sink fraction: 0.172 for local risk and 0.059 for TopoRisk
Fixed-trace allocation stress test	Mean paired residual-loss differences of -0.526 and -0.544 under two scorer/verifier assignments
Recovered topology, two checks	Dynamic TopoRisk is optimal on 44/44 graphs vs. 38/44 for static single-node marginal
Recovered topology, three checks	Dynamic TopoRisk is optimal on 44/44 graphs vs. 38/44 for static single-node marginal
Retail online pilot, DeepSeek agent	39/45 successful runs with and without verification
Airline online pilot, DeepSeek agent	41/45 without verification vs. 37/45 with full Qwen coverage
Airline online pilot, GLM agent	40/45 without verification vs. 39/45 with full Qwen coverage

The fixed-trace experiment uses constructed hazards over frozen contexts and is not an estimate of natural error frequency or deployment safety. The online null and negative results are retained. The GLM extension uses the same DeepSeek user simulator and Qwen verifier as the primary Airline experiment; it broadens agent coverage but does not establish provider-independent generality. The recovered-topology study uses real tool-call dependency structure but simulated faults; dynamic recomputation improves on the static single-node marginal ranking in 6/44 graphs at both tested budgets. This tests the allocation mechanism, not online safety.

Repository layout

results/controlled/          Synthetic-DAG results and ablations
results/fixed_trace_summary/ Aggregate heterogeneous-verifier stress test
results/online_summary/      Aggregate paired online-pilot outcomes
figures/                     PNG figures generated from packaged results
src/                         Simulation, analysis, plotting, and checks
docs/                        Provenance, scope, and reproduction notes

Reproduce the controlled study

Python 3.10 or newer is recommended. The simulator uses the standard library; Matplotlib is required only for figures.

python src/run_experiments.py --out results/controlled
python src/run_allocation_ablations.py --out results/controlled/ablations
python src/run_structure_sensitivity.py --out results/controlled/structure_sensitivity
MPLCONFIGDIR=/tmp/toporisk-mpl python src/make_figures.py
MPLCONFIGDIR=/tmp/toporisk-mpl python src/make_ablation_figures.py
python src/verify_release.py

The aggregate outputs from remote-model experiments are frozen so that their reported values can be inspected without making new paid API calls. Those aggregates cannot reconstruct the original conversations.

Intended use

These artifacts support analysis and reproduction of verification-allocation policies. They are not a production safety benchmark and do not establish that any model is safe for autonomous high-impact actions.

Citation

Until a manuscript identifier is available, cite this artifact release:

@misc{yuan2026toporiskartifacts,
  title        = {TopoRisk Evaluation Artifacts},
  author       = {Yuan, Ye},
  year         = {2026},
  howpublished = {Hugging Face dataset repository},
  url          = {https://huggingface.co/datasets/JosephAA/toporisk-evaluation-artifacts},
  note         = {Artifact-first research preview; manuscript not included}
}

Licenses

• Code under src/: Apache License 2.0 (LICENSE-CODE).
• Aggregate result tables, figures, and documentation: Creative Commons Attribution 4.0 International (LICENSE-DATA-DOCS).

No license is granted for excluded upstream benchmark content or for the unreleased manuscript.