We distinguish released code, controlled evaluation evidence and internal research. A reproducible test can establish behavior under its exact conditions. It cannot by itself certify a customer's production system.
Framework-independent evidence and provenance binding, deterministic policy/version checks, receipts, state invalidation, selective recovery planning and fail-closed results. Public conformance vectors and acceptance tests support repeatability.
Released source and provenance →A valid cryptographic receipt does not establish that upstream evidence was truthful.
Reports 12/12 hard gates and 8/8 frozen scenarios for a defined outcome-aware recovery authority boundary using NVIDIA OpenShell as the enforcement substrate.
Review exact conditions and artifacts →Independent EvidenceBound research. Not NVIDIA validation, endorsement or a blanket production guarantee.
Demonstrates a human-approval and revocation path changing the exposed tool capability from three safe tools to four and back to three. The execution path rechecks authority.
Open controlled demo →The demo records externalSideEffect=false; it is not a live infrastructure deployment.
Private tooling supports staged evidence-package checks, independent-journal comparison, version-bound handoffs and human-reviewed draft findings for bounded synthetic cases.
See the offered assessment scope →This is not proof of client adoption, externally authenticated collection or a completed third-party audit.
Every report specifies the operation, evidence source, timing, permitted environment and transition tested. Untested conditions are marked unverified.
Cryptographic binding and persistence attest to bytes and configured trust roots, not to the truth or completeness of upstream observations.
EvidenceBound does not claim ISO certification, legal compliance, full security review or regulatory approval.
Private evaluation utilities must not be described as installed customer agents, production security gates or independently audited components.
Start with released Core source and conformance vectors for the policy/evidence runtime, use the frozen RAB-1 result for bounded authority-recovery behavior, and inspect the WebMCP demo for conditional human-approved tool visibility. These are distinct evidence types, not independent proof of customer deployment.
Important: none of these artifacts establishes the truth of external evidence, a customer's production risk posture, or a formal certification. An assessment reports only the conditions and effects it can actually observe.
Browse the technical research, conformance vectors, public benchmark, controlled demo and assurance-context mapping.
Request an explicitly authorized, bounded check in staging or an agreed synthetic environment. The introductory fixed scope is $1,000 USD; broader testing is quoted separately.