Enterprise authorization with OPA and an AuthZEN-style adapter

This frozen external-consumer experiment asks whether one authorization policy can consume released SynthWorld 0.16.0 artifacts through an AuthZEN-style request boundary and preserve the scoreable authorization behaviour across two fictional organizations. Britannia is the calibration topology. Aurelia uses a different input schema that the mapping algorithm was not written against, but these two author-related inputs are not a representative generalization sample.

The experiment is recorded in Discussion #147 and published as the immutable release enterprise-authorization-opa-authzen-0.16.0-2. It is an unsupported evidence record, not a maintained SynthWorld adapter, AuthZEN conformance result, product certification, vendor comparison, or claim that the policy is suitable for production.

Experiment boundary

The experiment installed idcognito-synthworld==0.16.0 as a released dependency. It did not modify SynthWorld or import repository examples or tests. Its workflow kept benchmark generation, the system under test, submission sealing, and scoring as distinct capabilities:

Britannia and Aurelia topology YAML
    -> experiment-owned deterministic mapping
    -> released SynthWorld compilation
    -> physically split public and evaluator artifacts
    -> experiment-owned OPA policy and AuthZEN-style adapter
    -> raw responses and typed sealed submissions
    -> isolated offline SynthWorld scoring

The projector received only the split public deliverable and wrote the derived policy and request inputs used by OPA, the adapter, and the system-under-test runner. None of those capabilities had an evaluator mount; the offline scorer was the only evaluator reader. A deliberate regression exposed evaluator paths to both the projector and runner; the isolation checker rejected both configurations.

Run identity

The experiment directory was not a Git repository. Its frozen source is therefore identified by the source-file-set digest and final manifests, not by a source commit. The release tag points to the SynthWorld repository revision that hosts the evidence; it is not an experiment-source revision.

Method

One deterministic mapper converted each organization topology into the released enterprise identity and authorization authoring surfaces. One combined RBAC/ABAC/ReBAC policy then processed every public evaluation cell through an experiment-owned AuthZEN-style adapter in front of OPA.

Each request bound the subject identifier and type, action name, resource identifier and type, evaluation tick, profile, and topology lane to the selected public cell. Sixteen fixed-cell-ID mutation controls changed those values independently and were all refused. The runner retained raw requests and responses, normalized them into typed submissions, and sealed each topology’s public inputs, policy, adapter, package and product identities, raw results, and submission before scoring. The isolation checker ran separately and its retained evidence is present in the reference ZIP, but it was not seal-bound: a missing controls mount caused both published seals to record isolation_evidence: null. The recorded isolation result is therefore independently auditable rather than cryptographically chained into the submission seal.

The reproduction workflow uses Docker Compose, digest-pinned images, and a hash-pinned Python lock. The two complete deterministic runs compared 175 files byte-for-byte. A separate clean extraction of the published reproduction ZIP also completed successfully.

Results

Every metric retains its own denominator; no aggregate score is computed.

The baseline was iteratively improved while aggregate metrics were visible, and one author wrote both the scenario overlay and the OPA policy. The near-perfect figures therefore support that the released contracts were learnable and internally self-consistent in this experiment; they are not a blind first-attempt or independent-implementation result.

Britannia contributed 324 cells and the single final-decision error; Aurelia contributed 325 cells and reproduced every effective and final decision. The binding metric remains separately visible: canonical account-to-principal binding is evaluator-only, so plausible same-kind substitutions are not always decidable from the public evidence. A superficially correct final decision must not conceal that mechanism-level miss.

Frozen assets

Publication revision 2 is the authoritative release. Revision 1 accidentally included unmanifested Python bytecode caches in both archives. Revision 2 removes only those generated files; experiment inputs, source, retained evidence, results, and internal checksums are unchanged. Both revisions remain immutable so the correction history is auditable.

Verify the release attestation and assets with gh release verify and the retained checksums:

gh release verify enterprise-authorization-opa-authzen-0.16.0-2 \
  --repo bluntmachetti/synthworld
gh release download enterprise-authorization-opa-authzen-0.16.0-2 \
  --repo bluntmachetti/synthworld
sha256sum -c SHA256SUMS
unzip enterprise-authorization-opa-authzen-0.16.0-2-reproduction-kit.zip
cd enterprise-authorization-opa-authzen-0.16.0-2-reproduction-kit
./run.sh

First execution may need network access to obtain the pinned images and packages. After that run, ./reproduce.sh --offline performs the complete two-run byte-identity comparison using the retained local cache.

What this experiment does not establish

• AuthZEN protocol or implementation conformance. The adapter is merely shaped around the published decision-request model.
• A comparison among authorization strategies. It exercises one policy; the negative controls are intentionally broken variants, not alternatives.
• A blind first attempt. Three policy defects were corrected after aggregate metrics exposed them; no case labels, per-case verdicts, canonical identities, ownership truth, or expected outcomes were supplied to the policy.
• Independent interpretation by benchmark and policy authors. One author wrote both the scenario overlay and the OPA policy, so a symmetric misunderstanding could remain invisible even though the negative controls prove that the metrics move.
• Individual person-to-agent accountability. The topology mapping establishes team-level responsibility but does not model an explicit person assignment for every agent.
• Arbitrary topology portability. Aurelia shows that the mapping algorithm survived a schema it was not written against, but both topology documents share an authoring style and vocabulary and cannot represent a generalization sample.
• A SynthWorld rendering feature. SynthWorld 0.16.0 had no HTML renderer for these enterprise artifact types, and this experiment did not substitute an experiment-owned viewer.
• A seal-bound isolation proof. The isolation checker passed and its evidence ships separately, but the published seals contain a null isolation-evidence field.
• A maintenance commitment. Experiment authors own future adaptations, policies, infrastructure, and support.

These limitations define the useful next experiment: keep the released benchmark and isolation boundaries, add explicit person-to-agent accountability, and compare multiple independently designed authorization strategies over the same frozen worlds rather than optimizing another single policy against this result.