Research
Methods, results, limitations, and downloadable artifacts
Research that treats the runtime as part of the system
Each publication separates measured outcomes from interpretation. Results are dated and scoped to their tasks, models, policies, and runtime configuration. Downloadable PDF and source artifacts are provided when a paper reaches its publication threshold.
Papers
Simulation Fidelity and Runtime-Driven Convergence in Output-Capped Tool-Using Agents
A controlled study of whether runtime intervention—not prompt wording alone—can help tool-using agents finish large, structured artifacts under bounded output.
runtime control artifact validation ablation
The HTML edition is the primary accessible version. The PDF has selectable text but is not yet a tagged PDF.
Artifact-Grounded Evaluation of Tool-Using Agents
An evaluation method that reports objective artifact validity, tool policy, judge output, and platform reliability as distinct outcomes.
evaluation reproducibility agent systems
The HTML edition is the primary accessible version. The PDF has selectable text but is not yet a tagged PDF.
A Protocol-First Runtime for Stateful Tool-Using Agents
The design of a pure orchestration core surrounded by explicit model, tool, state, execution, and streaming adapters.
architecture state protocols
Publication standard
Before a quantitative result appears here, its public package must include:
- an operational definition of every metric;
- the task set, sample size, repetitions, and stopping rule;
- model and runtime configuration sufficient to interpret the result;
- sanitized inputs or a documented generator;
- versioned analysis code and generated figures;
- uncertainty and known limitations;
- a record linking each public claim to its evidence.
The complete standard is documented under Reproducibility.