A change to a shared service
Imagine replacing a model-provider adapter. Quality Fabric can run the selected behaviour tests. cert-fabric can examine whether the service respects its declared boundaries and whether substitution preserves the required contract. The result should retain missing checks, exclusions and the actual test executor.
A security question needs different evidence. Guardian’s experimental range investigates whether revoking a test credential really prevents its later use. That experiment informs protection design; it does not establish that this service change is secure.
Where the work stands
Quality Fabric and cert-fabric have service code, interfaces and test suites. The inspected sources include a path for cert-fabric to delegate test execution to Quality Fabric. This is evidence of implementation, not a fresh validation of a deployed service.
Guardian’s broader external security platform remains a design proposal. Its Phase-0 spike implements a smaller experiment with real local credential state and signed outcome records; separate-host authority and broader learned defences remain further work.
The shared principle is useful for an evolving AI system: keep the scope and origin of evidence visible, so a narrow success cannot quietly become a universal assurance claim.
Technical detail: interfaces, evidence and boundaries
Quality Fabric: execute and report
A test plan enters a service API, is organised for execution and reaches framework-specific adapters. The inspected pytest adapter collects structured results, coverage and artifacts; the execution routes expose submission, status and results. A separate Model Context Protocol (MCP) server makes selected Quality Fabric operations available to AI agents.
The wider project includes prediction, test generation and structural-analysis interfaces. Their presence does not establish predictive accuracy, comprehensive coverage or uniform support across every framework. Quality Fabric originated in work to validate Maestro and developed as a standalone testing project.
cert-fabric: make the boundary check explicit
A service manifest identifies contracts, test locations and applicable checks. Static gates and test outcomes contribute to a verdict covering independence, boundary integrity, substitutability and operational requirements. A numerical maturity score cannot override a failed hard gate.
The test-execution implementation attempts Quality Fabric and records which executor produced the result. When permitted, an unavailable executor can lead to an explicitly recorded local Python fallback; non-Python work cannot silently take that path. Required-executor policy can block the run. Verdict tests also check that unjustified narrowing of test scope cannot receive an unqualified pass.
Here, “certification” means this project’s declared engineering criteria. It is not an external accreditation or a compliance certificate, and it does not authorise deployment.
Guardian Fabric: separate judgement from authority
The platform design separates report-only sensors, detection, mechanically constrained response authority and an audit record. Model-assisted investigation would advise; operator-controlled policy would govern consequential action. This describes the proposed architecture, not the protection supplied by the spike.
The Phase-0 implementation uses a local credential store, deterministic scoring and a signed, hash-linked log. Tests distinguish recorded revocation from a credential that still works. A small adaptive attacker changes its choices from rewards. This is a limited adaptation experiment, not evidence of large-scale co-evolution or a complete security service.
The design’s separate trust domain is not delivered by in-process object separation. External log anchoring, separate-host settlement and broader learning environments need their own implementation and evaluation.
What this work teaches
Test execution, architecture checks and security experiments answer different questions. Their results need scope, provenance and failure states. A passing score should not conceal a failed prerequisite; an available interface should not stand in for a verified outcome. The next step is to evaluate these mechanisms in explicit deployment contexts, including failed dependencies, misleading evidence and recovery.