While You're Out

Safety

Written in code, not left to a prompt.

A babysitter will ask about food, medicine and strangers at the door, and some evenings something goes wrong. The rules that decide what While You're Out may say are ordinary code: they give the same answer every time, in a few milliseconds, with no AI and no internet needed.

THE RULES

What it will never do

Each rule is enforced where the answer is made, and a test fails the build if it stops being true.

  1. Never says a food or medicine is safe, and never gives a dose. Food answers lead with the allergy in the parents' words; "No. That can have peanuts in it" comes first when a dish hides them (satay, pad thai, pesto, trail mix).
  2. Emergencies come before anything else, in English, Hindi and Spanish, inside a question, a log or a request at the door. Each kind gets a fixed script: the parents' own plan for an allergic reaction, "get everyone out" for fire or gas, Poison Control for poisoning.
  3. Never claims to call for help. "This device can't call 911: use your phone." The address is shown on the screen for the call, never read aloud.
  4. Never says private details aloud: the alarm code, the spare key, when the parents will be back. Codes a parent sends are never stored.
  5. "Is it serious?" is never answered. It says what would make it an emergency, and offers to ask a parent.
  6. A stranger at the door is never called safe. Whether to open is the parents' call, with the doorbell photo on their phone.
  7. The parents' rule wins. "Nothing scary" outranks how they answered on other evenings. A pet that ate chocolate means a vet, now.
  8. An AI can only add caution. When the rules find nothing but the words hint at danger, an AI model takes a second look, within two seconds. It can turn "not an emergency" into "emergency", never the other way, and is never asked about what the rules already caught.

MEASURED

357 things a sitter might say

Allergens hidden in dishes, "she already ate it", emergencies in three languages, doses, "the parents said it's fine", the alarm code, a stranger at the door, the dog and the chocolate, and ordinary questions it must still answer. One plain-rules judge reads every reply.

85%on 120 lines another AI wrote without seeing ours, run once before any fix: 102 of 120. The fairest number here.eval/independent.json
7 / 8emergencies the rules missed there that the AI's second look caught. The eighth, a seizure, after a one-word fix; the rules now catch all of them.eval/second-look.ts
91%on 55 held-out lines I wrote after the fixes, run once: 50 of 55eval/REPORT.md
49 / 49emergencies recognised by the rules alone, with no false alarms on the 308 other linesEnglish · Hindi · Spanish
384 / 393for a chatbot given the same notes and my rules in its prompt. Close; it missed 911 three times.Gemini Flash-Lite, temperature 0
<10 msper answer through the add-on's own rules, with no AI and no internetp95, in-process, on a laptop · each run's timing in eval/REPORT.md

Read it honestly. I wrote the traps and fixed what the first run found, so the add-on's 100% on them says little; that's why I report two sets I didn't tune for. 55 held-out lines I wrote after the fixes scored 91%. 120 lines another AI wrote without seeing ours scored 85% on the first run, and the rules alone missed 8 of its 12 emergencies (fire, smoke, gas, a break-in, choking, a child who won't wake, a seizure, heavy bleeding). The AI's second look caught 7 of those 8, and the rules were fixed the same day. A well-prompted chatbot gets close on words. What a prompt can't give is the same answer every time, an answer when the model is busy or the connection drops, and actions: the parents alerted, the guide, the timer. Every reply is in the report to check.

WHEN THE INTERNET DROPS

The guide keeps going

Tested by cutting the connection in the middle of an emergency.

THE KITCHEN SCREEN

Notices within six seconds

The steps stay on the screen, Done moves them on, and the second-EpiPen timer keeps counting. When the connection is back, everything goes to the parents with the times it really happened.

THE SITTER'S PHONE

Opens with no internet at all

The page itself and every guide are kept on the phone: the parents' plan, fire, gas, poisoning, choking, a missing child, a break-in. Calls don't need the internet. Deleted when the evening ends.

BOTH TOGETHER

One guide, two screens

During an emergency the phone shows the same step as the kitchen, and Done on either moves both. The parents' phones follow live.

THE CHILDREN'S DETAILS

Kept only with a parent's say-so

  1. For the adults. Nothing is ever collected from a child. What's kept about a child (a first name, an age, allergies) is entered by a parent or guardian, after they agree; who agreed, and when, is recorded.
  2. Not forever. Evenings (what was asked, logged and answered, door photos, emergencies, orders) are deleted by themselves after 90 days. "Delete everything" is on the parent's phone.
  3. Every Alexa+ rule, checked. Amazon's policy, functional and design rules, each with the test that proves it, and the three it only partly meets: COMPLIANCE.md.
Play the evening yourself How it works