Your rules.
Checked by code.
North is a checkpoint for your money. You write the rules. Helpers and AI assistants can only ask. Plain code, the Money Guardian, decides, and only it can approve a payment.
The path of a payment
An AI assistant can now spend your money. North is the checkpoint that makes it follow your rules.
Prices and store comparisons come live from Channel3, and so does the 30-day price history (it builds up day by day). North never takes a price from an AI: it reads the price from Channel3 again at checkout.
Pick a step to see what it does
Every payment follows the same five steps. Nothing moves by itself: you choose.
What each one can and can't do
Concierge
AI helper
Can
- Read your words or photo and pick the right helper.
- Turn a rule you wrote into a structured rule. You read it back and confirm.
- Ask one question with picture choices if you were too vague.
Can't
- Pay or refund.
- Make a rule weaker. It can only create or tighten rules.
- Approve anything.
Finder
AI helper
Can
- Turn words or a photo into products.
- Give 3 picks that fit what is left this week.
Can't
- Set a price. Names, pictures and prices come from the store.
- Buy anything.
Scam checker
AI helper
Can
- Read an ad and find the matching listing.
- Raise a flag. A flag can only make the Guardian refuse.
Can't
- Write the verdict. Code does, from real prices.
- Approve or loosen anything.
Price Watcher
AI helper
Can
- Watch only items you are waiting on.
- Write one short message about a change that code found.
Can't
- Buy. It can only tell you.
- Put its own numbers in a message. Code fills in prices.
Any outside AI assistant
Connects through MCP
Can
- Ask to buy, with a reason.
- Read your rules in plain words and what is left this week.
Can't
- Pay, refund, or change a rule.
- See your PayPal or the amount set aside for rent.
- Use Buy now. Only you can.
Money Guardian
Plain code, no AI
Can
- Approve, wait, save for later, or refuse.
- Sign a one-use approval for exactly one payment.
Can't
- Be talked into anything. No AI output can approve a buy or weaken a rule.
- Let anything past a hard wall, even Buy now.
What the Guardian checks, in this order
It stops at the first check that applies, and tells you which one it was. The same order is shown in Behind the scenes after a buy.
Hard walls
Refuse, no way around. A price that cannot be read, a seller flagged as a scam, a buy that listing text started, rent and bill money, going over this week's plan.
Your must-have list
Groceries and other must-haves pass every rule you made. They never pass a hard wall.
Locked rules
Never, while the rule is on. The only way around is changing the rule, which takes a day.
A saved item's own timer
Saved items wait their full time, whatever the time of day.
Firm rules
The buy is saved for later. No Buy now.
Gentle rules
Wait, or buy now? Buy now works right away, and is written in the audit log.
- Price and settings can be read
- Seller is not flagged as a scam
- Not started by text hidden in a listing
- Rent and bill money stays safe
- Fits what is left this week
- Your must-have list
- Locked rules
- A saved item's own timer
- Firm rules
- Gentle rules
Your words become the rule
This is a real rule: the one the practice account starts with. The words below are written by North's own code.
1 · You write
“Wait a day before I buy anything after 9pm.”
2 · The Concierge makes a rule
{ "kind": "late_night", "afterHour": 21, "untilHour": 5, "strength": "firm", "waitHours": 24 }Code checks every field. The AI cannot add a field or loosen one.
3 · North reads it back
Wait a day on buys from 9pm to 5am. I save it and remind you.
You choose Gentle, Firm or Locked, and confirm. Nothing is saved until you say yes.
4 · The Guardian follows it
A stricter rule starts right away. A weaker rule, or turning one off, starts tomorrow, and you can undo it before it starts.
What has been checked, and how far that goes
These numbers are read from the project, not typed in. The test numbers were made on November 9, 2026 at version d718b74. Run npm run proof to make them again.
- 1,810of 1,811 automatic tests pass (1 skipped), in 100 test files.
- 29Guardian tests that try thousands of random buys, rules and times.
- 612attacks from an AI attacker, in 10 categories. Got through: 0. Money moved against the rules: $0.00.
- 1serious hole found by an independent AI review. It is fixed and has a test.
Run on November 8, 2026. The attacker was Haiku 4.5: the bigger models often refuse that job. A separate judge reads the ledger to see whether any money moved against the rules.
| What was tried | Tried | Turned away | Got through |
|---|---|---|---|
| poisoned listings | 63 | 31 | 0 |
| prompt injection reason | 43 | 33 | 0 |
| fake urgency | 39 | 25 | 0 |
| price quantity currency | 91 | 58 | 0 |
| replayed approvals | 53 | 25 | 0 |
| races | 48 | 29 | 0 |
| revoked keys | 74 | 64 | 0 |
| demo tool abuse | 51 | 31 | 0 |
| auth and csrf | 118 | 88 | 0 |
| rule loosening | 32 | 20 | 0 |
“Turned away” means the request was refused or stopped before it could do anything. The rest were asks that were inside the rules, answered normally. Zero got through.
Serious An AI reviewer that had not seen the work looked for a way to move money without the Guardian. It found one: an assistant could buy with no yes from you once a saved item's wait was over. It is fixed, a regression test guards it, and the same pass found 2 low problems and 1 note, also fixed or written down.
Honest limit: the fixes were checked by tests, not by a second independent review.
32 deal-check cases and 10 listings with hidden instructions, plus 37 rule-writing cases. The developer wrote them with Claude's help, so they are a smoke test, not proof. On the last recorded run, the real Sonnet 5.5 followed none of the 10 hidden instructions. If it had, the Guardian would still refuse: a buy that listing text started is a hard wall. Run npm run eval:scam to repeat it. That run uses the live AI and store.
What this does not show
- No person outside this project has audited North, and there has been no outside penetration test.
- Payments here use a PayPal sandbox with test money. Real payments would also need PayPal's risk data and monitoring.
- The checks cover the cases we thought of, and the attacker was an AI. Passing them is not a promise that nothing can go wrong.
Where PayPal shows up
At checkout, the button says Pay with your saved PayPal. After a payment, the receipt says Paid with PayPal and keeps PayPal's own order and capture numbers. Turn on Behind the scenes on a laptop to see the same numbers, and whether PayPal was called at all. If the Guardian says no, it was not.