pki.sgit.ai / the bench / what you authorised and never asked for

What you authorised and never asked for

A grant is what your agent's tools can do. A mandate is what you expected it to do. Everything in the grant is authorised — you, or a default you never looked at, authorised it. The gap between the two is what you authorised and never asked for, and it is where the losses live.

This page cannot scan you. Agent exposure lives inside a laptop, a container, a workspace configuration and a set of connected accounts, and there is nothing to probe from outside; a site that claimed to scan it would be lying. So this page hosts the questions, the public profiles and the interpretation, and everything below runs in your browser: your answers are kept on your machine, and there is no button that sends anything anywhere. The measurement itself runs where the agent lives, with the probes.
Step 1 · name your tools

Which of these do you use?

Public profiles, contributed and challengeable. A profile is a configuration of a product listing its tools; two variants of one product are siblings that differ in a few rows. Measured means somebody ran the probes and the file is on the site; a claim means the rows are derived from what the surface is.

Or bring a measurement:

Step 2 · your grant appears

What those tools can do — before you have typed anything about yourself

Name at least one tool above.

What would prove this assessment wrong

An action inside the grant is never a surprise, however wrong it was. A surprise is an action outside the grant. A hallucination inside the grant is by design and is bounded; one that reaches outside the grant means the measurement was wrong, or containment was. So the assessment is falsifiable: zero surprises over a period means the measurement held and the assessment was calibrated; any surprise means it was wrong, by exactly that much. Very few security assessments can say what would prove them wrong. (The game next door measures something else — a prediction gap — and never calls it a surprise.)
What happenedWhose failureWhat to do
Inside the grant, outside the mandateThe agent did something it was permitted to do and nobody expectedThe grant's. It was too wideReduce the grant, or accept it by name
Outside the grantThe agent did something the measurement said it could notThe measurement's, or containment'sRe-measure. Your model of the world was wrong

The assessment expires, and that is the product

Every assessment carries a date and the profile versions it was computed against, and it goes stale when either moves. Three sources of movement, none of which you can watch and all of which the site can:

MovementEffect
A vendor changes a defaultEverybody on that profile has a wider grant this morning than last night
A product gains a capabilityNew rows appear in a grant nobody edited
An incident is publishedA control claimed as enforced is demoted, and any policy relying on it loses a condition

Free through level three; independence is what costs

LevelWhat it isFree or paid
1 – 2The concepts, the public profiles, your grant, the questions, the gap — this pageFree, open source, no account
3The reduction guidance and the re-check loop — this pageFree. Guidance that costs money is guidance nobody follows
3Running the probes where the agent lives — the runnerFree. Self-run, and labelled as such
3 – 4Your environment modelled rather than the public one: your tools, your assets, your mandate, your registerPaid. Work somebody does for you
4 – 5Somebody running the probes in your environment and vouching for the result; watching the ground move for a named estatePaid, recurring. Independence is the thing a customer cannot produce for themselves at any price

The boundary, stated rather than discovered: everything about the public products is free and open; everything about your particular estate is work somebody does for you. Nothing on the paid rungs is a feature an individual cannot have; it is a quantity an individual does not need.

What this does not prove

Specified by brief v0.33.64 (name the question, not the concept), shaped by the precedents: one small input, the verdict first as a statement rather than a grade, the evidence beneath, no account, a result worth showing somebody, and the schema before the site. The technical term stays excess authority (grant minus mandate) in the schemas; the public phrase is a sentence, because every coined noun tested either collided or needed explaining. CC BY 4.0.