Code

Runner / orchestrator / scoring code that produced everything in data/, plus the figure-build scripts that turn the analysis CSVs into the rendered PDFs in paper/figures/. Anonymized for double-blind review.

Layout

pyproject.toml       package metadata + dependencies
requirements.txt     pip-style mirror for non-poetry installs
.env.example         environment variables (only OpenRouter + tracing knobs)

decision_bench/      the Python package
  bench.py             CLI entry: `python -m decision_bench.bench` to
                       launch a single (model × substrate × condition) cell
  benchmarks_runner.py runner-side dispatcher
  benchmarks/          per-substrate adapters
    base.py
    gaia.py
    tau_bench.py
    bfcl.py
    swe_bench_pro.py    appendix-only, not in headline results
    terminal_bench.py   deferred substrate scaffold
  call_model.py        the `call_model` tool wired into orchestrators
  read_profile_tool.py the `read_profile` tool wired into orchestrators
  read_file_tool.py    auxiliary file-read tool
  client.py            OpenAI-compatible TracingClient over OpenRouter
  config.py            env-driven config (OpenRouter key + tracing knobs)
  registry.py          model-id ↔ canonical-name + vendor mapping
  accounting.py        per-task / per-call trace recorder + budget caps
  smoke.py             substrate sanity-check entry
  analysis/            tagger + scorer + analysis utilities
    tagger.py            deterministic step-skill tagger (no LLM)
    profile_static.py    builder for C2 cards
    profile_judge.py     builder for C3 cards
    gaia_score.py / gaia_split.py / loaders.py / analyze.py
  skills/taxonomy.py   7-skill taxonomy + step-rule definitions
  profiles/            33 profile cards (3 variants × 11 models;
                       same content also at ../../profile_cards/)

tools/                  one-off CLIs and pipeline glue
  build_c1_template.py / build_c2_profile.py / build_c3_profile.py
  decisionbench_analyze.py    end-to-end analysis driver (zips → CSVs)
  launch_stage2_sweep.py      run all Stage-2 cells in-process
  launch_swe_pro_pilot.sh     appendix-only SWE-Bench-Pro pilot
  swe_pro_reparse.py          re-score SWE-Bench-Pro patches
  terminal_bench_eval/        deferred Terminal-Bench scaffolding

figure_scripts/         the figure-build code referenced in REPRODUCING.md
  build_main_figures.py     Figs 2, 4, 5, 7, 8, 9 + appendix recovery_ratio
  build_pareto_addons.py   Fig 3
  build_ceiling_fig.py     Fig 6
  build_overview_diagram.py Fig 1
  _fig_style.py            shared color / font constants

(per-variant profile cards live in ../profile_cards/ at the bundle root)

What was removed for the anonymous release

A web-facing proxy service and operator account-management code lived upstream of this runner in our deployed stack; none of it is required to reproduce the experiment, so it was removed. Specifically: the HTTP API surface and any code referencing payment processing, per-user state, multi-account API-key rotation, and automatic Hugging Face upload have been dropped. What remains is the minimum needed to (a) launch a single cell against OpenRouter, (b) sweep the full (model × benchmark × condition) Stage-2 matrix in-process, (c) re-aggregate the released run zips into the analysis CSVs, and (d) re-render every figure in the paper.

Anonymization placeholders

ANON-* strings appear where original deployment values were removed. None of these are required to reproduce the released numbers from the released run zips:

Where	Placeholder
.env.example runtime knobs	reduced to OpenRouter key + tracing-dir + app-name only
figure_scripts/*.py REPO path	resolved at runtime via Path(__file__).parents[] walking up looking for analysis/; override with DECISION_BENCH_REPO env var

The git history (.git/) is not included.

Running

See ../REPRODUCING.md for the full per-claim reproduction recipe.

# install
python -m pip install -e .
cp .env.example .env       # then set OPENROUTER_API_KEY

# launch one cell (single-shot CLI; no condition orchestration — uses the
# substrate's solo runner). For Stage-2 (condition × model × benchmark),
# use the sweep launcher below.
python -m decision_bench.bench gaia --model claude-opus-4.7 --max-tasks 32

# launch the full Stage-2 sweep in-process (one cell at a time)
python tools/launch_stage2_sweep.py \
       --version stage2 \
       --jobs-dir ./stage2_jobs/ \
       --max-usd-per-job 30 \
       --out stage2_manifest.jsonl

# aggregate released zips → analysis CSVs
python tools/decisionbench_analyze.py \
       --in-dir ../data/stage2_runs \
       --out-dir ../data/analysis/regenerated