Turn the customer message into structured ticket facts.
Let the model propose. Let the gate show its work.
A small agent can draft the move. The request, memory recall, policy rule, and proof chain stay visible while the decision resolves.
The model drafts an unsupported action from the request alone.
Remembero returns the facts and active policy rule.
The deterministic decision decides whether action is allowed.
Please refund this order today.
request(t_418, refund)amount(t_418, 249)disputed_identity(t_418)
Instruction: You are a refund triage agent. Return exactly one action token: APPROVE_REFUND or ESCALATE_HUMAN. Do not add explanation. Customer request: Please refund this order today. Memory tool result: none
Not run yetInstruction: You are a refund triage agent. Return exactly one action token: APPROVE_REFUND or ESCALATE_HUMAN. Do not add explanation. Customer request: Please refund this order today. Memory tool result: request(t_418, refund) amount(t_418, 249) disputed_identity(t_418)
Not run yet- Query
- action_blocked(t_418, approve_refund)
- Rule
- action_blocked(Ticket, approve_refund) :- proposed_action(Ticket, approve_refund), requires_human(Ticket).
- action_blocked(t_418, approve_refund)
- proposed_action(t_418, approve_refund)
- requires_human(t_418)
- disputed_identity(t_418)
Awaiting model output
Incoming request
Refund ticket T-418 · $249 · identity disputed
Please refund this order today.
Structured facts
request(t_418, refund)amount(t_418, 249)disputed_identity(t_418)
Prompt-only agent
Small model · no tools
- Build no-memory packet
- promptAvailableBrowserModel(packet)
simulated · ungoverned
Grounded agent
Same model · memory + rules
- Build memory packet
- promptAvailableBrowserModel(packet)
- proposal: approve_refund
- evaluateGroundedAgentCase(proposal)
The model never gets mutation authority.
The model reads the request and proposes what it wants to do.
Grounded rules and memory evaluate the proposal and decide: approve, block, or escalate.
Only the gate's decision can produce an action. Every decision is backed by facts and rules you can inspect.
Model proposals use the deterministic simulator (LanguageModel not run). Remembero memory, rule evaluation, and proof are real and run in this browser.