The cup note

Starbucks has asked its baristas to write short notes on cups. Head office wrote one rule. What customers got depended on who read it: a store manager judged on drive-thru times, a barista with twelve people in line, a regular who likes being known, a commuter who just wants the coffee.

A policy does not produce an outcome. It passes through people, and each of them bends it. This page replays a simulation of that chain, worked out in advance: six stores, twelve baristas and ninety-six customers, each one played by a language model, under 24 versions of the rule and the shift. Pick a version and run it.

What this run found

• The rule bent before it reached a single cup. Under the strict rule, only 15 of 48 manager decisions enforced every cup. The rest carved out a rush exception or softened it to a request, each for a reason that fits their store.
• Strict and adaptive buy different things. On a short-staffed rush with personal notes, strict kept notes on 60% of cups and adaptive on 15%. Strict cost 7.0 seconds a cup against 2.0. Customer satisfaction came out at 62 and 60: the note barely moved it.
• The line sets satisfaction; the note moves connection. Satisfaction averaged 77 on quiet, fully staffed shifts and 60 on short-staffed rushes, whatever the rule. Connection averaged 42 across the runs where the note had to be personal and 38 where any mark counted. A rollout that measures only compliance would miss both.

How to read this

Follow one rule down the chain. The memo at the top is the same for every store. The managers decide what it means for their team, the baristas decide what to write with a line in front of them, and the customers decide whether they cared. The five numbers at the bottom are what head office would see; everything above them is why.

Change one lever at a time. Run the default, then press Compare and switch Enforcement from Strict to Adaptive. The managers, cups and faces swap to the new run, and the line under the levers says what moved. The charts at the end put all 24 runs side by side: who kept the most notes, who kept customers happiest, and what each cost in seconds.

Read the quotes, not just the totals. The numbers say a policy worked or failed. The quotes say where it broke: a manager who tells the team one thing and thinks another, a barista who stops writing at the fourth cup, a customer who liked the note but hated the wait.

How it was made

Each person is one call to gpt-6-luna, with reasoning effort set to low, run in advance through OpenAI’s batch service. There are three rounds, and each one sees only what the person in it would know:

• Managers see the rule, the shift, their store and their own job: what they are judged on, how long they have run the store, how they feel about programs from head office.
• Baristas see their manager’s message, the shift, and the eight customers they serve with the line length at each. They decide each cup and how many seconds it costs.
• Customers see who they are, where they are, how long they waited and what is on their cup. They never see the rule or the manager.

The people are the same in every run, so the differences come from the rule and the shift, not from new faces. Wait times are worked out by plain arithmetic from the line and the seconds each note added; the model never sets them.

This run made 2,736 calls across 24 scenarios (144 manager, 288 barista and 2,304 customer), used 2,008,340 tokens and cost $0.20. 0 answers failed the schema check. Generated 2026-10-29, prompt version v1.

What this is not

It is not Starbucks data. The stores, managers, baristas and customers are invented, and a language model wrote every word they say. It is not a forecast either: it shows how a rule can bend as it passes through people, not how often it will. Use it the way you would use a tabletop exercise before a rollout to thousands of stores: to find the questions worth testing in real stores, not to answer them.