Authority and autonomy – suiss One docs

One is designed to do more for you over time, but it never earns authority by guessing. Authority is something you give explicitly, step by step, in Access. One can be more careful than its authority allows; it can never go beyond it.

Where authority comes from

The authority One can use is the intersection of the Grants you give One and the Mandate of the One instance doing the work. Both are Access records.

One cannot grant itself anything, widen a Grant or stretch a Mandate.
Conversations, memories, routines, relationship levels and model confidence never create authority.
Anything that needs your own authority (a new Grant, widening one, an approval) is prepared by One and carried out by you, from your own instance in Access.
Every sub-agent and execution instance is a separate agent instance in Access. Its authority can only be narrower than One's, it has its own trail, and it can be stopped on its own.

Autonomy per action

How much One may do without asking is set by the behaviour scope of the Mandate. It is not one global slider: it can differ per capability, per resource (account, recipient, merchant), per context (time, place), per budget and per risk.

Choice	What One does
Do it yourself	If Access allows the action, One does it, unless one of the "ask anyway" factors below applies
Do it if I said so before	One acts only if the earlier go-ahead (which request, which wording, when) is recorded against the Mandate; otherwise it asks
Ask me first	One prepares the action and you confirm it on the Access approval surface
Leave it to me	One prepares the information and the link, and does not act

"Order my weekday coffee", "draft emails and ask me", and "leave bank transfers to me" can all be true at the same time.

Rules in your own words

You describe a rule in a sentence or pick a template. One turns it into a structured Grant and Mandate proposal (capability, resource, amount and budget, time window, end date, behaviour scope), and you see it as a plain preview in Access before confirming. The preview calls out anything One had to interpret, such as whether "weekdays" includes public holidays. Until you confirm, the proposal has no effect, and granting is always your own action.

When One asks anyway

Even when Access allows an action, One is designed to ask in these cases:

Factor	Example
Low or conflicting confidence	"Your usual americano?" when recent orders have varied
Novelty	A merchant, recipient, address, amount level or capability seen for the first time
Intent conflicts with habit	You said you feel unwell today, so the usual coffee order is not placed
A parameter came from outside	The recipient, amount or target URL came from an email or a web page
Irreversibility	The action cannot be undone and confidence is borderline
Price or terms changed	The price is clearly above your usual level, or the delivery date moved

Confidence only works in the direction of asking. High confidence never turns a denial or a required approval into an allow, and there is no fixed "act above 0.9" threshold.

What One never does on its own

These actions cannot be given to One or its sub-agents by any Grant, and nobody, including you, can relax this list:

Changing passwords, two-factor or recovery settings, including recovery email and phone.
Adding a payment method.
Widening a Grant or any authority.
Deleting the account, or destroying a retention class, a life area or all of memory.
A first transfer to a recipient that has never been paid before.
Sharing sensitive data with a new party for the first time.
Signing out sessions and authorizing new devices.
Installing software.

For these, One only hands off: it takes you to the right screen and explains what is needed. There is no switch for them in settings; the list is shown read-only. If the browser agent runs into one of these on a website, it stops and hands the task to you.

Forgetting a single memory record is not on this list. It is an everyday action you take in the app with a clear confirmation.

Acting while you are away

Background work is read-only by default. While you are away, One reads, watches, prepares and adds things to your brief; it does not change anything. A state-changing action happens only inside a Mandate that is explicitly named, has a budget, an end date and a fixed target (a site, a recipient, a merchant).

"When X happens, do Y" is an Access Trigger, always tied to a Mandate:

Field	Content
Event source	Calendar, email, a price watch, a Relay event, another One's typed reply
Condition	Written in the shared restriction language
Allowed action	A single action with a fixed target
Budget	An amount and/or count ceiling
End date	Required; there are no open-ended Triggers
Bound Mandate	The Trigger cannot exceed it

When a Trigger fires, One asks Access for a fresh decision. If the Mandate has been revoked, narrowed or has expired in the meantime, nothing happens, and your brief says the rule fired but the authority is gone. If Access requires an approval, the work waits for you; your absence never counts as consent.

Each Trigger appears as a card: "If X, do Y, at most Z, until this date", with its last run, remaining budget and end date. It can be paused or removed with one tap.

Approvals

What needs approval comes from Access, from the action's consequence tier, its reversibility and your rules. It is never left to the model's judgement, and there is no classifier-based automatic approval.
Some actions always need approval by default: purchases, sending messages, submitting forms, legal acceptance such as agreeing to terms, and changing account settings. Over time you can add narrow exceptions ("coffee under a set amount from this café"), never a blanket one ("all purchases").
Silence is not consent. Every approval request has an expiry. When it passes, the work stops ("stopped: no approval") and no default action is taken. Restarting it later is a new Access decision.
The agent does not write the approval. The exact intent is rendered by the Access trusted approval surface from typed fields. One's explanation sits next to it, visually separate. Higher-risk actions require a passkey unlocked by biometrics or PIN.
The approval channel never collects secrets. No password, PIN, card number, one-time code or recovery code is ever asked for as an answer.
Approval fatigue is measured. If you keep accepting the same kind of request unchanged, One may propose a narrower rule; you still confirm it yourself.
Approvals can be listed together but are bound one by one. "Approve all" is offered only for low-risk, reversible items.

Reversibility

Reversibility is a separate field from risk. Every capability declares one of three values:

Value	Meaning	Example	What you see
`reversible`	Undone without side effects	Adding a calendar entry	"Undo"
`compensable`	Made good with time or cost	Payment to refund, order to cancellation	"Request refund" or "Cancel", with the time and cost
`irreversible`	Cannot be undone	Sending an email or message, sharing data	No undo button

An undeclared value counts as irreversible. For irreversible sends such as emails and messages, the executor is designed to hold a short window by default in which you can still say "don't send"; you can turn it off or lengthen it per capability. For every purchase, the refund and dispute path is kept, so "Request refund" is always next to it in the activity log.

Which surface can approve what

Action	Voice (car, glasses)	Watch	Phone or desktop with strong authentication
Low-risk, reversible or compensable, within Mandate and budget	Can approve	Can approve	Can approve
Irreversible, higher risk, or over budget or Mandate	Cannot; sends it to your phone or desktop	Cannot; sends it to your phone	Can approve
Never-delegated actions	No	No	Only by your own direct action

This table is an upper bound. Your settings can only make it stricter, for example "don't approve anything from my watch".

Payments and loyalty

All payments go through Pay. One requests a payment, and the card or agent token comes from Pay; One never sees raw card data.

One works for you. Commissions, paid placement and suiss's own interests do not enter recommendations or rankings, and suiss products are ranked like any other provider.