SynthWorld user guide

This file is now a compatibility index for repository links and historical anchors. Detailed user documentation lives at https://bluntmachetti.github.io/synthworld/.

Use these canonical entry points:

The headings below intentionally preserve earlier GitHub anchors while routing each topic to its canonical owner.

Choose your use case

See the documentation home for current journey routing and the generated benchmark catalogue for governed current state.

Try SynthWorld without installing it

Browse the published frozen tables on Hugging Face, then use Benchmarks for publication boundaries.

Install and create your first world

See Getting Started.

Run five foundational evaluation examples

See Evaluating a system.

Use case 1: safe connected identity fixtures

See Identity worlds and BENCHMARKS.md for the core fixture’s measured limits.

Use case 2: PII extraction

See Privacy and exposure and DATA_DICTIONARY.md for the exact extraction contract.

Use case 4: relationship inference

See Identity worlds, Evaluating a system, and BENCHMARKS.md.

Use case 5: breach-risk calibration

See Privacy and exposure and DATA_DICTIONARY.md.

Use case 6: agent identity and delegated authority

See Agent authority, Asteria Agentic v1, and the agent-authority contract.

Use case 7: exposure scenarios

See Privacy and exposure.

Use case 8: households and workplaces

See Identity worlds and BENCHMARKS.md.

The original guide carried one measured reference that is retained here for compatibility until it has a dedicated generated-performance reference page. examples/measure_households_cost.py records interpreter and platform with each run.

Reference run: Python 3.12.12, Linux x86-64, glibc 2.43; three timed repeats after a discarded warm-up.

People	Median runtime	Peak Python allocation
100	0.046 s	1.1 MiB
500	0.287 s	12.3 MiB
2000	2.003 s	116.6 MiB

uv run python examples/measure_households_cost.py --person-count 100 --repeats 5

Peak allocation is tracemalloc, not process RSS. The configured 2000-person ceiling is not a measured performance cliff.

Use case 9: identity-resolution ambiguity

See Identity resolution.

Use case 10: search-provider input without the answer key

See Privacy and exposure.

The public half rejects truth, it does not merely omit it

See Public vs evaluator.

Controlled failure modes, all planted deliberately

See Privacy and exposure and BENCHMARKS.md.

Scoring, and what it refuses to hide

See Privacy and exposure and Metrics.

Reference baselines

See Privacy and exposure and BENCHMARKS.md.

Use case 11: enterprise identity and access structure

See Enterprise Identity Planning, Enterprise identity and access, and the enterprise contract.

Author, validate, compile

See Enterprise identity and access.

Validation reports every error in a stage, not the first one

See the enterprise contract for normative validation behavior.

What compilation writes

See Enterprise identity and access and the enterprise contract.

What the seed moves, and what it does not

See Determinism, seeds, and keys.

Limits worth knowing before you author

See the enterprise contract.

Use case 12: projecting a compiled world to SCIM, OpenFGA, and AuthZEN

See Standards profiles and the enterprise contract.

SCIM

See Standards profiles.

OpenFGA

See Standards profiles.

AuthZEN

See Standards profiles.

Every projection reports what it lost

See Standards profiles and the normative enterprise contract.

Shared Signals / CAEP is a declaration, not an emitter

See Standards profiles.

Use case 13: enterprise authorization benchmarks

See Enterprise Identity Planning, Enterprise identity and access, Evaluating a system, and the relevant contract README.

Run the enterprise-agentic smoke pack

See Enterprise identity and access and CLI reference.

Check the shape before you score

See Evaluating a system.

Evaluate a prediction

See Evaluating a system.

Scoring the directory/RBAC oracle from Python

See Enterprise identity and access and the enterprise contract.

The identity-fabric pack is Python-only

See CLI reference and the enterprise contract.

Limits, stated plainly

See Enterprise identity and access and Benchmarks.

Reading evaluation results

See Metrics and Evaluating a system.

Safety boundary

See Safety boundary and Public vs evaluator.

Enterprise trees

See Enterprise identity and access and the normative enterprise contract.