Aula de cursos

Goal: inspect a sophisticated generative AI answer rather than celebrate its length. Current models can produce hypotheses, tables, operational plans and sometimes interactive simulations. That output is a starting point, not a validated result.

The coffee-shop report

A small coffee shop wants to reduce customer waiting. An assistant attributes delays to payment at the counter, layout and preparation, and strict first-in-first-out processing. It recommends QR payments, rearranging equipment and batching similar drinks. It lists observations to collect and proposes a simulator. This is a useful report, but what exactly has the assistant established?

First problem: plausible is not observed

It calls small premises, limited staffing and a specific order/payment sequence “known facts.” Those are plausible descriptions of many coffee shops, yet the assistant has never visited this shop. Customers may order through an app; multiple baristas may work independently. The right label is assumption or generalization, not local observation.

Second problem: serving more is not the same as making everyone wait less

Batching may improve throughput while delaying complex orders. Measure at least total completed orders per hour, average waiting time and high-percentile waiting time. A policy can improve one measure while worsening another. Ask which notion of service success matters and to whom.

Third problem: a convincing simulator may model the wrong system

A dashboard with sliders for counter, bar and delivery times need not simulate queues. To model congestion, one may need arrivals over time, product mix, staffing and parallel stations, variability, priority rules and order dependencies. A stage-time calculator is not a calibrated operational simulation.

2026 laboratory

1. Ask any available generative assistant for three improvements to a fictional coffee shop’s waiting times. Save its complete answer.
2. Audit each claimed fact. Which is actually known? Which is an industry generalization? Which is a guess?
3. Change the success metric from orders per hour to the 90th-percentile waiting time. Examine whether its preferred solution changes.
4. If your system supports generated code or apps, ask for a minimal queue simulator with visible fictional inputs and documented assumptions. Otherwise request pseudocode and a spreadsheet calculation.
5. Test boundary cases: no arrivals, more arrivals than service capacity, a station offline, and a complex order ahead of five simple orders.

You have already given a coherent answer. Do not provide another list of ideas. Audit your reasoning. Identify unobserved assumptions about the shop, distinguish throughput from waiting-time distribution, explain when batching can disadvantage customers, and specify the variables a serious queue model needs. Design a low-cost observation to disconfirm your recommendation.

Submission: original answer, audit, revised hypothesis and reproducible test plan. Your score depends on materially useful corrections rather than the number of tokens you generate.