How the roll-up gate works
How the Agent Action Admissibility gate works
This tool grades each proposed AI-agent action against your own domain rules before it runs. It reads 8 actions and rolls them into one verdict — the worst, never the average — then a gate that can only make it worse.
The 8 actions
01A-1002 · refundfix firstsecond refundsOrder assertion on order-4471 - rule marked review, promoted to blockingINADMISSIBLE
02A-1003 · payoutpaysTo routed to party-88, a SupportRep - range requires BuyerINADMISSIBLE
03A-1006 · grantparty-88 asserted as both Customer and SupportRep - disjoint classesINADMISSIBLE
04A-1008 · refundthird refundsOrder assertion on order-5120 - promoted to blockingINADMISSIBLE
05A-1005 · status_writehasStatus 'probably shipped' is not in the closed set {paid, shipped, refunded}REVIEW
06A-1001 · refundfirst refundsOrder assertion on order-5120 - no rule violatedADMISSIBLE
07A-1004 · payoutpaysTo party-41, a Buyer - within rangeADMISSIBLE
08A-1007 · status_writehasStatus 'shipped' - in the closed setADMISSIBLE
LEDGER CLEAN
every action admissible
HOLDS QUEUED
worst action is a closed-set or review-only violation
CONTRADICTIONS FOUND
any action violates a functional, disjoint-class, or range rule
Structural promotion gate · worsen-only
Functional and disjoint-class rules assert impossibility, not surprise - a violation of either is always promoted to blocking severity, even when the rule pack marked it review, because a contradiction cannot be a warning.
CONTRADICTIONS FOUNDWhy: R1 (at most one refund per order) was declared severity 'review,' but functional rules are structural, so the engine promotes the second refund on order-4471 to blocking regardless - a contradiction is never just a warning. 3 admissible - 1 review - 4 inadmissible, of 8 actions - admissible share 38% (context only).
Fix firstA-1002, the second refundsOrder on order-4471 - its own rule was marked review, but the promotion gate forces it blocking anyway.
No single action is the crisis — the clustering is. The roll-up takes the worst, the gate escalates when trouble compounds, and it names the one thing to fix first. Grades proposed actions against rules you write. It is a working aid, not legal, compliance, or security advice, and it never scores, ranks, or assesses people.
RedHub AI