Origin
Run reference check

Synthetic authority analysis

See the gap between access and use.

Inspect standing permissions, exercised authority, and what a compromised identity could reach. Then widen a delegation edge and watch the result change.

Where the reference check asks what a selected policy would allow, this analyzer measures how much authority a seeded synthetic agent fleet holds versus what it exercised; it evaluates no policy and signs nothing. Every fleet and event on this page is synthetic and seeded. These numbers describe the demonstration, not a customer deployment. Definitions and denominators accompany the published benchmark.

Metric definitions ↗Published synthetic benchmark ↗Explore the analyzer ↓

05 · Delegated authority

Widen one edge. Watch the root's blast radius move.

A correct capability token can only narrow as it is delegated, so an ancestor reaches exactly its own grant. Break that on any edge and authority flows back up the tree.

Delegation treeSix synthetic identities. A widened child grant adds its scope to every ancestor's effective authority.platformopsfinancepayroll-botpii-botledger-botDelegation treeSix synthetic identities. A widened child grant adds its scope to every ancestor's effective authority.platformopspayroll-botpii-botfinanceledger-bot
  • passattenuation holds — Every child narrows. The root reaches exactly its own grant: 5 of 9 sensitive resources.
  • noteblast radius at the root — 5 / 9 sensitive resources reachable — BRI 55.6%
Recorded · one take · synthetic data
Recorded 2026-09-28 on originphysicalai.com (57d5ccf), using synthetic demo data. The in-page analyzer uses seed 20260818, 400 roots and 2,074 synthetic identities. 63.2% of granted scopes were never exercised in the window. Widening one delegation edge changes the root’s sensitive-resource reach from 55.6% to 66.7%. A fresh seeded corpus has 90 planted widened edges: 90/90 caught, 0 false positives against synthetic ground truth. The separate published benchmark uses 2,000 roots. Duration 0:29. Captions and a cursor dot are overlaid; the run is one continuous take. Recording method and provenance.

63.2% of granted scopes never exercised in the window · seed 20260818, 400 roots, 2,074 synthetic identities · describes the demonstration, not a customer deployment

90/90 planted widened edges caught, 0 false positives · against synthetic ground truth · separate seeded corpus

55.6% → 66.7% root's sensitive-resource reach after widening one delegation edge · synthetic, seeded