WP42 — The Autonomous White-Collar Worker
Measuring the economic frontier of Claude Code, computer-using agents and autonomous knowledge work. UQO Working Paper Series No. 42 (Department of Administrative Sciences, Université du Québec en Outaouais). Author: Simon-Pierre Boucher. Built 2026-11-02 →.
Research question. What share of economically valuable white-collar deliverables can an autonomous Claude Code worker complete to a professionally acceptable standard, at what cost, with what reliability and supervision needs, and how much of it is explained by the level of computer access (file-only → shell → macOS layer → browser/office/portal → self-repair)?
Headline results (pilot, 150 runs, claude-opus-5-5, 2026-11-03)
Scenario-weighted acceptance (deterministic completion ∧ no critical violation ∧ blind quality ≥ 70): A0 0.00 · A1 0.87 · A2 0.90 · A3 0.90 · A4 0.83.
Shell premium A1−A0 = +0.87 [0.67, 1.00], p = 0.004. Preregistered primary contrast A3−A1 = +0.03 [0.00, 0.10], p = 1.0 (precise null, ceiling).
Cost per accepted deliverable 0.87–1.11 USD (Sonnet 5.5: 0.40 USD at 0.85; Fable 5.1: 3.30 USD at 0.85). Exploratory recovery experiment:
after an induced interruption at 120 s, one bounded recovery session raises acceptance from 0.05 to 0.85 (20/20 self-check passes).
Human arms H0/H1 designed, not executed. Total API spend 261.89 USD. Monograph: release/wp42_uqo.pdf (189 pp.); journal version: release/wp42_uqo_journal_version.pdf (41 pp.).
Layout
| Path | Content |
|---|---|
CLAUDE.md |
the research mandate (verbatim) |
RESEARCH_PLAN.md, PROJECT_STATUS.md |
operational design, phase status |
protocol/ |
preregistration, amendments, ethics (human arms, not executed), power analysis |
literature/ |
110 verified references (references.bib), structured review, verification log |
benchmark/ |
authoring spec, taxonomy, public scenario bundles (scenarios/<TASK_ID>/), shared evaluator helpers, public self-check |
src/wp42/ |
controller: store, launcher, arms, bundle, portal, scheduler, evaluate, judge, metrics, export, dashboard, watchdog, CLI |
data/raw/ |
SQLite store + immutable raw streams per run (git-ignored) · data/validated/ Parquet/CSV datasets |
analysis/ |
econometrics, robustness, figures, tables (all generated by scripts) |
manuscript/ |
LaTeX monograph · journal_version/ short version |
dashboard/ |
private monitoring page (python -m wp42.dashboard --serve 8642) |
tests/ |
unit tests (.venv/bin/python -m pytest -q tests) |
Hidden oracles (ground truth, deterministic evaluators, rubrics, validation reports) live outside the repository in ~/wp42-oracles/
and are never copied to executor nodes. The API key lives in ~/Desktop/Cluster/secrets/wp42/.env (controller) and <node>:~/wp42/.env.
Running the laboratory
cd wp42 && export PYTHONPATH=src
.venv/bin/python -m wp42.cli sync-tools # portal package → nodes
.venv/bin/python -m wp42.cli assign --tasks ACC-01 FIN-01 ... --arms A0 A1 A2 A3 A4 --reps 2 --seed 4242 --stage pilot
.venv/bin/python -m wp42.cli run --parallel 12 # executes, collects, evaluates, judges (envelope enforced)
.venv/bin/python -m wp42.cli status | ledger | export --tag pilot
.venv/bin/python -m wp42.cli kill-all # emergency shutdown
.venv/bin/python -m wp42.dashboard --serve 8642Workers run as claude --bare -p on M3U96a/M3U96b/M4M64a/M4M64b (Claude Code 2.1.314) in ~/wp42/runs/<run_id>/, with the arm's tool
flags, 4 USD / 1,800 s / 150-turn caps, no Agent tool and a depth guard. See docs/PHASE_0_DISCOVERY.md for the measured environment.