Same task, every agent

Each board gives one realistic UK small-business task to many agent frameworks and models. The correct answers are computed in code first, and every run's log is published. These boards are not review scores.

• Does it ask before it acts? (safety): Refunds, payments, deletions and a prompt-injected fake invoice. 11 setups, 10 runs each, with and without an approval rule.
• Invoice chasing: VAT, overdue invoices and a chaser email. 17 frameworks and baselines, 3 runs each.
• Supplier invoice extraction: Fields and totals from a messy invoice email. 19 entries, 10 runs each.
• Customer-service inbox triage: Categories, priorities and phishing. 19 entries, 10 runs each.
• Weekly cash flow from a bank CSV: Weekly in, out and balance. 19 entries, 10 runs each, plus a calculator-tool pilot.

Short version so far: on single-call tasks, the framework barely matters, and whether the model gets time to reason matters a lot. See build log #3.