A research update regarding CLAIM-24 demonstrates a critical failure mode in autonomous AI agents where time-based authorization grants ignore external state changes. In a simulated scenario, an agent granted access at 10:00 AM proceeds to act at 11:30 AM based on a time-to-live check, despite the partner's access being revoked at 11:00 AM. The researchers identified that standard timestamp-only gates fail to verify if the underlying permissions remain active, resulting in unauthorized data transmission.

To address this, the team developed a re-derivation gate that validates the current state of the source at the moment of execution. While a standard timestamp gate returned 'ALLOW' for the divergence scenario, the re-derivation gate successfully identified the discrepancy between the initial grant—which allowed 'dev-reader' access—and the current source state, which returned a 'restricted' role, resulting in a 'REFUSED_STALE' status. This code path was validated across seven scenarios using a mock adapter built by the research team.

The project currently lacks external evidence, as all tests have utilized self-authored mock systems rather than independent memory stores or policy registries. The research team is now seeking access to external, read-only memory stores with provenance boundaries—systems the agent cannot modify—to test the claim against real-world conditions. Developers are invited to clone the 'ai-memory-judgment-demo' repository and implement a 'SourceAdapter' to run the seven defined scenarios against their own external data sources. The team aims to conduct the first external test by the end of June 2026. A failure of the re-derivation gate on an external source would falsify the current architecture, while a successful refusal would strengthen the claim.

A research update regarding CLAIM-24 demonstrates a critical failure mode in autonomous AI agents where time-based authorization grants ignore external state changes. In a simulated scenario, an agent granted access at 10:00 AM proceeds to act at 11:30 AM based on a time-to-live check, despite the partner's access being revoked at 11:00 AM. The researchers identified that standard timestamp-only gates fail to verify if the underlying permissions remain active, resulting in unauthorized data transmission.

To address this, the team developed a re-derivation gate that validates the current state of the source at the moment of execution. While a standard timestamp gate returned 'ALLOW' for the divergence scenario, the re-derivation gate successfully identified the discrepancy between the initial grant—which allowed 'dev-reader' access—and the current source state, which returned a 'restricted' role, resulting in a 'REFUSED_STALE' status. This code path was validated across seven scenarios using a mock adapter built by the research team.

The project currently lacks external evidence, as all tests have utilized self-authored mock systems rather than independent memory stores or policy registries. The research team is now seeking access to external, read-only memory stores with provenance boundaries—systems the agent cannot modify—to test the claim against real-world conditions. Developers are invited to clone the 'ai-memory-judgment-demo' repository and implement a 'SourceAdapter' to run the seven defined scenarios against their own external data sources. The team aims to conduct the first external test by the end of June 2026. A failure of the re-derivation gate on an external source would falsify the current architecture, while a successful refusal would strengthen the claim.