Skip to content
SynthWorldDeterministic identity benchmarks
Esc
↑↓navigate↵open⌘Jpreview
On this page

Connected synthetic identities. Repeatable tests.

SynthWorld builds connected fictional identity worlds for testing identity, privacy, and access systems. Generate useful fixtures, or test whether an agent still has authority after credentials and delegations change.

Choose your starting point

  • Generate synthetic identity data — linked fictional personas and relationships for identity matching, privacy tools, and connected data workflows. Start with a repeatable seed; add evidence and ambiguity when you need them.
  • Test agent authorisation — seven concrete action cases for delegated access and agent lifecycle decisions. Compare teaching policies, omit the ReBAC authority view, and inspect the failure.

Both paths work locally. Synthetic identity generation needs no policy engine, agent runtime, or authorisation setup. Python 3.12 or newer is required.

Start with connected identity data

pip install idcognito-synthworld==0.18.0
synthworld generate --seed 20260719 --persona-count 10 --output world.json

The same version, seed, and configuration reproduce the same bytes. These are fictional fixtures, not anonymised records from real people.

Continue to identity worlds, identity resolution, or privacy and exposure.

See an authorisation failure before integrating anything

A valid credential does not guarantee an active delegation. The agent-authorisation quickstart shows a post-revocation request that a credential-only check allows, and how adding the delegation check changes the outcome. You can inspect the result in your browser before installing the demo.

The comparison uses local teaching policies. It does not contact your production system or prove that an execution path enforces its decisions.

Go further

Versions and evidence

Both first-use paths above target SynthWorld 0.18.0. Pin the package version for repeatable integrations. Other reference pages track current main unless they state a specific released version; check each capability’s availability.

Where a benchmark provides public inputs, those inputs are physically separated from expected answers. Published reference truth is inspectable; this prevents accidental answer leakage, not deliberate cheating. Frozen conformance fixtures are not evidence of real-world generalisation.

Source · PyPI · Releases · Hugging Face · Support

Was this page helpful?