Claude Fable 5

The last train had gone. Mara stood under the dead departures board, her ticket to the coast folded small in her palm. Behind her, the shutter of the kiosk rattled as the night guard locked it.

She counted the benches. Four. The far one held a glove—child-sized, red wool, the thumb worn through. Someone had set it upright on the armrest, fingers open, as if waiting to be collected.

Mara picked it up. Damp. The cuff had a name stitched in uneven thread: Theo. The same loops, the same wobble, as the label inside her own son's coat, the one still hanging in the hall at home.

She turned the glove over twice.

The guard called out that the doors were closing. Mara tucked the glove into the lost-property slot by the window, where morning staff would find it. Then she unfolded her ticket, tore it once, and dropped the halves into the bin.

She walked back toward the street, the way she had come.

Why price alone can't be blamed

Two changes happened at once—higher price and shorter hours—so they are confounded. Fewer coffees could result from either (or both): shorter hours mechanically cut the number of selling opportunities, independent of price. Without separating them, any causal claim about price is unidentified.

Simple experiment

Hold hours fixed and vary only price (or vice versa), ideally randomizing across days:

import random
days = list(range(28))
random.shuffle(days)
treatment = {d: ("high_price" if i < 14 else "old_price")
             for i, d in enumerate(days)}
# Keep opening hours IDENTICAL on all days.
# Record coffees sold per day, then compare group means.

Compare average daily sales between high-price and old-price days (a t-test or simple difference in means). Because hours are constant and price is randomized, any gap reflects price. Optionally run a 2×2 design (price × hours) to estimate both effects and their interaction.

Check: also compare sales per open hour in the historical data—if that rate barely fell, hours, not price, drove the drop.

type Appointment = { id: string; startsAt: string }; // ISO 8601, UTC

function localDateKey(iso: string, timeZone: string): string {
  // en-CA yields YYYY-MM-DD
  return new Intl.DateTimeFormat("en-CA", {
    timeZone, year: "numeric", month: "2-digit", day: "2-digit",
  }).format(new Date(iso));
}

export function groupByLocalDate(
  appts: Appointment[],
  timeZone: string,
): Map<string, Appointment[]> {
  const groups = new Map<string, Appointment[]>();
  for (const a of appts) {
    const key = localDateKey(a.startsAt, timeZone);
    (groups.get(key) ?? groups.set(key, []).get(key)!).push(a);
  }
  return groups;
}

Time zones: Store instants in UTC; pass the user's IANA zone explicitly (never rely on the server's local zone). Intl handles DST and offset history. Avoid toISOString().slice(0,10), which always gives the UTC date.

Midnight test:

test("23:30 Pacific stays on the 15th", () => {
  const a = { id: "1", startsAt: "2024-03-16T07:30:00Z" }; // 23:30 PDT, Mar 15
  const g = groupByLocalDate([a], "America/Los_Angeles");
  expect([...g.keys()]).toEqual(["2024-03-15"]);
  expect(groupByLocalDate([a], "UTC").has("2024-03-16")).toBe(true);
});