XSOC AIDA-Guard

Two agents with the same role ask to export the same records to the same target. The second has read the content below from an external channel. Type the injection; the Guard decides both.

Two gateways are in play: production NIE authorizes the agent's model calls to EXIOM, and the Guards run on this host. EXIOM governs what the model may say; AIDA-Guard governs what the agent may do; each decision lands in Providence.

The agent's task tells it to follow the briefing's handling instructions, so its model is easy to persuade by design: the scene shows enforcement that sits outside the model.

The agent's capability token is minted for one Guard. Copy it, as an attacker who lifted it from the agent would, and present it to a second Guard. The key a Guard holds is derived for that Guard alone.

One control advances the tenant epoch. Every token minted at an earlier epoch, the one that just succeeded among them, is refused on its next use, with EpochStale in the record. There is no revocation list.

Edit the Q3 figures. Each is sealed under CKKS as its own ciphertext before any computation touches it, and the FHE service computes the risk score on the ciphertexts. Name two approvers to release the score.

The mode is a classification lookup: regulated data runs in Mode A, so the computation runs on ciphertext. Releasing the result is a separate, gated, recorded step: it presents the gate host's endpoint attestation and a dual-control receipt naming both approvers, and the chain records it.

Both approvals arrive through this console's operator session; the receipt and the release record name both.