Irreversibility Budget Artifact

This directory contains the simulator, trace analysis, benchmark, and figure generation code for the irreversibility-budget paper.

• sim/fleetsim.py: discrete-event fleet simulator for RQ1-RQ5 and all arms.
• sim/results/: checked-in 300-run simulator outputs used by the paper.
• sim/bench.py: hierarchical escrow-ledger microbenchmark.
• sim/bench_results.json: checked-in benchmark output used by the paper.
• sim/traces/: real-trace classifier, analysis, parsed traces, and plot.
• figures/plot_figs.py: regenerates the three paper figure PDFs from sim/results/results.json.
• figures/*.pdf: generated outputs copied here for reproducibility. The paper tree keeps its own copies of the three PDFs that main.tex includes.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

For tests, coverage, and dependency auditing:

pip install -r requirements-dev.txt
python -m coverage run -m pytest
python -m coverage report
python -m pip_audit -r requirements.txt

From this directory:

python3 sim/fleetsim.py --runs 300 --out sim/results
python3 figures/plot_figs.py
python3 sim/bench.py --out sim/bench_results.json
python3 sim/traces/analyze.py
python3 sim/traces/plot.py

sim/traces/analyze.py uses the moved parsed JSONL trace artifacts by default. Pickle cache loading is disabled unless --use-pickle-cache is passed, because pickle is executable data. To re-parse raw cloned datasets, run with --fresh and set either IRREVERSIBILITY_TRACE_RAW_ROOT or both IRREVERSIBILITY_TAUBENCH_DIR and IRREVERSIBILITY_AGENTDOJO_DIR.

The parsed trajectories under sim/traces/ are derived from two public benchmarks, re-published here in parsed form so the analysis runs offline:

• tau-bench (retail + airline): github.com/sierra-research/tau-bench
• AgentDojo: github.com/ethz-spylab/agentdojo

See sim/traces/results.json → provenance for the exact paths used.

If you use this artifact, please cite the paper:

The Irreversibility Budget: Fleet-Level Risk Accounting and Admission Control for Agent Operating Systems. In Proceedings of the 2nd Workshop on OS Design for AI Agents (AgenticOS 2026).

MIT — see LICENSE.

@inproceedings{Mohammadi2026IrreversibilityBudget,
  author    = {Mohammadi, Bardia and Bindschaedler, Laurent},
  title     = {The Irreversibility Budget: Fleet-Level Risk Accounting and Admission Control for Agent Operating Systems},
  booktitle = {Proceedings of the 2nd Workshop on {OS} Design for {AI} Agents ({AgenticOS} 2026)},
  year      = {2026},
  address   = {Prague, Czech Republic},
  publisher = {ACM},
  url       = {https://github.com/mpi-dsg/irreversibility-budget}
}