Each case shows what the human said, what could go wrong, the decision the engine actually returns in the sandbox, how execution happens - with or without Quattrino - and what the proof is.
01
The human said
“Buy a USB-C cable, keep it under $20.”
The risk
Overspend, wrong merchant, hidden recurring charge.
What Quattrino decides
Execution
Merchant checkout (Stripe SPT, test mode) or your own checkout.
Proof
explain: standing $50 · task $20 · effective $20 · binding layer = task.
02
The human said
“Get whatever we need to repair this machine.”
The risk
Hiring or signing a contract without anyone being asked.
What Quattrino decides
Execution
External vendor; Quattrino never needs to move the money.
Proof
Envelope version + the question record + the owner amendment that answered it.
03
The human said
“Get a tool for this, one time, under $20.”
The risk
A subscription disguised as a purchase.
What Quattrino decides
Execution
The vendor’s checkout, reported back as an outcome.
Proof
Outcome compliance flags UNAUTHORIZED_COMMITMENT if it subscribed anyway.
04
The human said
“Get the logo done, under $300. I must own the design.”
The risk
Paying for work without the rights; budget leaking across agents.
What Quattrino decides
Execution
Sandbox marketplace (Credits) or an external service.
Proof
Delegation lineage: who granted what to whom, and how much remained.
05
The human said
“Per-customer standing authority, set by each customer.”
The risk
One customer’s agent spending another’s budget.
What Quattrino decides
Execution
Your own rail. Quattrino authorizes; you execute; you report the outcome.
Proof
Tenant-isolated policies; per-decision evidence you can hand to the customer.
06
The human said
“Book me a hotel.”
The risk
Non-refundable booking, loyalty signup, price creep.
What Quattrino decides
Execution
Online travel agency or the hotel directly.
Proof
Pinned envelope version; outcome checked for amount and commitment class.
07
The human said
“$300 for the project, split as needed.”
The risk
Authority multiplying as it is delegated down the tree.
What Quattrino decides
Execution
Any rail; each sub-agent authorizes against its own conserved grant.
Proof
Delegation tree with remaining, promised and spent per node.
differentiation_journeys.py replays journeys A-G against your sandbox account and prints expected vs actual. The playground runs single decisions with no code.
