ExoAI-S reports bounded progress on Guardian Receipts research

FOR RELEASE — September 2, 2026

Arkansas, United States

ExoAI-S is reporting a bounded research result from one frozen, inert software-verifier task. Two independently retained proposals each passed all seven visible test groups. A pre-bound 19-group held-out suite then found one defect in the first proposal and two in the second.

Both outcomes were retained separately, and automation stopped before choosing or applying either patch. This is the intended human decision gate.

WHY IT MATTERS

The useful result is not that either proposal was perfect. It is that hidden checks exposed task-specific defects after visible checks passed, while the workflow preserved the evidence and did not promote a candidate automatically.

EVIDENCE BOUNDARY

Neither proposal passed the complete contract. This does not demonstrate a finished security product, general model superiority, customer demand, grant readiness, autonomous-agent containment, or a scientific breakthrough.

NEXT MILESTONE

Keep both original results sealed, make the human repair-or-reject decision, and test the proposed customer problem through advance-consented interviews.

ABOUT GUARDIAN RECEIPTS

Guardian Receipts is the working name for an experimental authorization-and-evidence design intended to bind one specific agent tool action to its delegation context and link the executor’s reported outcome to that permission. Current work is bounded and human-supervised. It does not monitor the open internet or execute unknown code.

Public research notes: https://exorcizeaisafely.safelyexoai.chatgpt.site/research-notes