Product capability · Decision Authorization

Prove what was authorized—and what actually happened.

Execution evidence and recovery

Authorization, pre-effect consumption and observed execution are distinct facts. OntoGuard’s correlated evidence and standalone verifier make those distinctions inspectable under the configured trust model.

Why it matters

Support audit and incident investigation without treating a log entry or an ALLOW as proof of actual business effect.

Evidence chain

Signed authorizationPre-effect consumeTrusted executorExecution receiptReconciliation

Idempotency and recovery standing help distinguish unresolved outcomes from completed effects. System-of-record reconciliation detects discrepancies within observed coverage; it is not prevention of every alternate route.

In correctly configured production trust mode, required evidence must be independently anchored and verified before acknowledgement/protected execution; unavailable required proof cannot yield success. Reference-log qualification is not evidence of an external production log deployment. Historical integrity, current freshness, rollback and fork checks are distinct.

Proof disclosure uses approved scope and pinned trust identities. Browser checksum verification is not signature, independent-log or freshness verification.

PR #201 · TRACE adapter

Independently reviewed bounded integration

Merged September 20, 2026. Independent maintainer reproduction: 27 adapter tests. Authorized $250,000 path committed once with TRACE; altered $260,000 path refused, pending, zero commits and no TRACE.

Read PR #201 and review

PR #236 · Pre-commit enforcement

Independently reviewed bounded integration

Merged October 6, 2026; October 3 review by Carlos (carloshvp): 36 integration tests, 97 negative wrapper/direct-executor probes refused with zero commits and no protected effect, 4 positive probes executed the validated exact action. Imran Siddique approved and merged.

Read PR #236 and review

Bounded adapter and software harness, not independent validation of the complete runtime, SI, customer deployment or universal non-bypassability. Source: linked PRs.

Evaluate the runtime

Define the source or protected-action boundary, success criteria and integration responsibilities before a focused evaluation. No confidential data through the public site.

Current-release verification track — controlled evidence

Release-reported qualification, v1.5.2 (Python 3.13.5). The distribution reports 463 combined runtime and verifier tests passing in each of the source and installed-wheel campaigns. Its reference-log measurement reports anchor-added p95 23.1 ms and group commit 3,714 authorizations/s/tenant at a 20 ms commit window. These are reference measurements, not end-to-end customer latency or independent production-log performance. The record does not establish external production log deployment, Python 3.12 qualification, or signed OCI publication. Public PR reviews are separate evidence. Publication of release-specific metrics requires owner approval.

Historical verifier exercise: the current reference section includes eight preserved synthetic cryptographic vectors and two fresh mock-target execution packets. Offline recorded checks use test trust material, not an independent production log. Packet downloads are historical reference evidence. No current V16 freshness check was run.

Inspect current reference evidence

Semantic Intelligence

Actual controlled synthetic v332 run: 22 constituted concepts, 20 graph edges and 2 Needs Review rows, with exact source-cell projection; not a production-qualified demo.

Open controlled SI Explorer · Download projection

Decision Authorization

Fresh v1.5.2 reference packets and a separate controlled local mock-target execution, with test keys and recorded offline checks; not production deployment.

Inspect packets · Release evidence

What has been independently reviewed?

AgenTrust PR #201, merged September 20, 2026: 27 adapter tests in maintainer reproduction; the exact $250,000 payment committed once with TRACE, while the changed $260,000 path was refused with zero commits and no TRACE. PR #236, merged October 6, 2026 after the October 3 review: 36 integration tests, 97 negative wrapper/direct-executor probes refused without protected effect and 4 positive executions. These validate bounded integrations, not the whole proprietary runtime or all routes.

How does authorization differ from execution?

In a separate v1.5.2 controlled local mock-target run, the changed amount returned HTTP 409 with no observed mock commit, a tampered consume acknowledgement was refused, and an exact verified consume produced one recorded local effect. The pre-effect packet reports CONSUMED_OUTCOME_UNKNOWN (checkpoint 4); the exact-effect packet reports EXECUTED_EFFECT_FORMED (checkpoint 5) after an accepted signed outcome receipt. The former is not proof of no effect. Both historical checks used development test trust; V16 current freshness was NOT RUN. No real financial transfer, independent production log or customer deployment is claimed.

Inspect the ten selectable proof packets · Before outcome packet · Recorded exact-effect packet · Reference execution observations.

Integration status and scope

AgenTrust is an independent open-source project, originally sponsored by OPAQUE Systems. OPAQUE wound down in October 2026 after a technology-licensing agreement with Anthropic. Neither AgenTrust, OPAQUE nor Anthropic endorses or partners with OntoGuard.

AgenTrust independent-project notice · Constellation Research · OPAQUE announcement

Listed at AgenTrust’s verified tier (TRACE conformance level 0); the adapter is Apache-2.0.

Decision Authorization v1.5.2 includes an OpenShell-compatible supervisor adapter and an agent-action corridor. It is an integration capability; it does not imply an NVIDIA partnership, listing or certification.

Official integration index