1% ---------------------------------------------------------------------------2% Abstract body (Agent H). Input inside \WPabstract; no \section here.3% Sources: analysis/tables/table1_outcomes_by_arm_pilot.csv (acceptance, quality,4% critical-error rate, cost per accepted), table2_contrasts_pilot.csv (contrasts),5% table9_model_sensitivity.csv (Sonnet/Fable), table10_recovery.csv and6% analysis/derived/recovery_experiment.json (recovery), docs/PHASE_5_AUDIT.md7% (violations, judge agreement, spend), protocol/amendments.csv (A-002, A-011).8% ---------------------------------------------------------------------------9We build and operate an experimental laboratory in which a controller launches independent Claude Code sessions (\texttt{claude-opus-5-5}, client 2.1.314) on four Mac nodes to perform complete professional assignments---reconciliations, valuations, appraisals, credit analyses, contract reviews, payroll closings, bid evaluations, KPI packs, executive triage and an empirical replication---in a synthetic white-collar economy with hidden oracles, planted errors and prompt-injection traps. Ten validated scenarios in ten occupations are run under five randomised levels of computer access: no shell (A0), shell (A1), shell plus the headless macOS automation layer (A2), full computer use with browser, office tools and an enterprise portal (A3), and the same with bounded self-repair (A4). Acceptance requires deterministic completion, no critical violation and a blind quality score of at least 70. Over 150 pilot runs, scenario-weighted acceptance is 0.00, 0.87, 0.90, 0.90 and 0.83 in A0--A4. The shell premium is $+0.87$ [0.67, 1.00] ($p=0.004$); the preregistered primary contrast, full computer use minus shell, is $+0.03$ [0.00, 0.10] ($p=1.0$), a precise null at this difficulty, where every arm with a shell sits at a ceiling. Seven runs contain a critical violation, six of them in A0. The marginal cost per accepted deliverable is 0.87--1.11 USD; \texttt{claude-sonnet-5-5} reaches 0.85 acceptance at 0.40 USD and \texttt{claude-fable-5-1} 0.85 at 3.30 USD. In an exploratory experiment that interrupts sessions at 120 seconds, allowing one bounded recovery session instead of stopping raises acceptance after the induced interruption from 0.05 to 0.85 ($+0.80$ [0.55, 0.95], $p=0.004$); 20 of 20 resumed sessions pass the public self-check. We operationalise the Autonomous Labor Frontier---acceptance within a budget and a supervision allowance---and show that, under the stated five-minutes-per-question assumption, the supervision margin rather than the budget is what separates the frontier from the acceptance ceiling. The human arms were designed but not executed, the benchmark is authored and validated by agents, and no displacement estimate is implied. Total API spend was 261.89 USD.10