TP2A Prototype B · M3

AI Runtime Explorer

Deterministic runtime skeleton. Tidak ada live LLM/provider. Tujuannya membuktikan identity, tool allowlist, permission, run limits, structured output, escalation, dan audit boundary sebelum model live dihubungkan.

Agent registry

Reconciliation Agent v1

ACTIVE
Model mode
controlled-live-provider
Limits
6 tool calls · 15s

Permissions

case.readpayment.readreconciliation.readreconciliation.suggest

Explicitly forbidden

reconciliation.approveallocation.createpayment.updatesettlement.approvewaiver.approverestructure.approvelitigation.approve

Runtime simulation

Prove the boundary

Belum ada simulated run.

Expected: valid suggestion → COMPLETED + human review required. Approval/allocation → DENIED.

Tool Registry

get_case

Read minimum approved case context through Case Service.

Runtime guardrails

  • • Agent identity is separate from human and service identities.
  • • Every tool must be explicitly allowlisted on the agent version.
  • • Tool permission must also exist in the agent permission set.
  • • Human-only or business-state mutation tools are never executable by an AI agent.
  • • The runtime fails closed when a tool, permission, budget, timeout or output contract check fails.
  • • AI outputs are recommendations/evidence and never become business truth without the owning service and required human gate.
  • • Live model execution is separately feature-gated and disabled unless TP2A_ENABLE_LIVE_LLM=true with an explicitly configured provider endpoint and credential.
  • • The model receives minimum reconciliation context and cannot select a candidate outside the supplied receivables.