Untrusted context admission, derivation control, and effect-bound revalidation in agentic systems

A computer-implemented method for governing external context artifacts in an agentic system, the method comprising:

• receiving an external context artifact for potential use by the agentic system in connection with a candidate execution operation;
• assigning an artifact identifier to the external context artifact and storing, in an artifact registry, an artifact record associated with the artifact identifier, the artifact record storing at least a source class, a current version indicator or digest, and an admission state;
• determining provenance evidence for the external context artifact and, using one or more machine-readable policy rules, determining a mutability state and a trust tier for the external context artifact;
• determining, based at least in part on the provenance evidence, the mutability state, and the trust tier, an authority-eligibility status indicating whether the external context artifact is eligible to influence authorization of the candidate execution operation;
• assigning an admission state selected from deny, quarantine, context-only, and authority-bearing to the external context artifact;
• storing the external context artifact or a transformed representation thereof in one of a plurality of workspaces selected according to the admission state;
• intercepting, at an execution boundary gate associated with the candidate execution operation, a request to authorize or perform the candidate execution operation;
• revalidating, at the execution boundary gate and in view of a then-current version indicator or digest, the admission state, the mutability state, the trust tier, the authority-eligibility status, and at least one applicable policy state associated with the external context artifact;
• and permitting the candidate execution operation only when revalidation confirms that a workspace-selected representation of the external context artifact remains eligible to serve as an authority-bearing input for the candidate execution operation, and otherwise blocking, downgrading, or holding the candidate execution operation.

One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause an agentic-system governance platform to:

• receive an external context artifact for potential use by an agentic system in connection with a candidate execution operation;
• create, in an artifact registry, an artifact record associated with an artifact identifier for the external context artifact, the artifact record storing at least a source class, a current version indicator or digest, and an admission state;
• determine provenance evidence for the external context artifact and, using one or more machine-readable policy rules, determine a mutability state and a trust tier for the external context artifact;
• determine, based at least in part on the provenance evidence, the mutability state, and the trust tier, an authority-eligibility status indicating whether the external context artifact is eligible to influence authorization of the candidate execution operation;
• assign to the external context artifact an admission state selected from deny, quarantine, context-only, and authority-bearing;
• store the external context artifact or a transformed representation thereof in one of a plurality of workspaces selected according to the admission state;
• and expose to the agentic system only a workspace-selected representation of the external context artifact, including preventing a representation assigned the context-only admission state from serving as an authority-bearing input to authorize the candidate execution operation.