Request
The agent proposes a concrete action with target and context.
Deterministic · local-first · auditableVerify AI claims
Verify AI claims
before agents act.
Huqan judges AI claims against graph-backed evidence, protects canonical memory from unsupported writes, and keeps risky actions behind explicit policy gates.
Policy
Human
Audit
Memory
01Claim judgmentKnown, unknown, contradicted or risky — made explicit.
02Memory admissionUnsupported knowledge does not become trusted memory.
03Action gatesRisky operations follow deterministic review paths.
04Trust receiptsEvery consequential verdict keeps its evidence and reason.
Control flow
Model output is not authority.
Huqan sits between an agent’s intent and the systems it can affect. It receives an action request, applies policy and context, routes decisions for human approval when required, and records the outcome.
Evaluate
Deterministic policy and risk rules examine the request.
Approve
Required decisions are routed to the right human authority.
Execute
Allowed actions proceed through the controlled boundary.
Record
The decision and context become auditable evidence.
Browser demo
Test the trust boundary yourself.
Choose one of three realistic scenarios and see how Huqan turns evidence, conflict and policy signals into distinct decisions.
This experience runs in the browser; it makes no live API or filesystem calls.
Edit the text and evaluate a claim of your own.
SignalClaim
ImpactLow
VerdictALLOW
Evidence-supported claim
Evidence and provenance signals are consistent; the information is admissible to trusted memory.
Trust Receipt
tr_demo_claim_7f3adeterministic / local / auditableInspectable decisions
Every consequential action should explain itself.
This interface makes the core contract visible: the acting agent, requested operation, target, applicable risk and accountable decision.
Action request
AgentDataAgent
Actionwrite_file
Target/reports/summary.xlsx
RiskHigh / approval required
Decision record
{
"agent": "DataAgent",
"action": "write_file",
"policy_result": "approval_required",
"decision": "approved",
"evidence": "recorded"
}
"agent": "DataAgent",
"action": "write_file",
"policy_result": "approval_required",
"decision": "approved",
"evidence": "recorded"
}
One local boundary for claims, memory and action.
EV / 01
Policy enforcement
Deterministic constraints govern tool calls, data access and action boundaries.
AP / 02
Approval runtime
Human review remains available where risk or policy requires explicit authority.
CX / 03
Context and provenance
Decision context is preserved so outcomes remain reviewable rather than opaque.
RC / 04
Trust receipts
Inspectable records verify what was requested, decided and executed.
Clear answers for high-stakes AI.
Practical notes on agent governance, evidence, verification, memory admission and bounded action.
What is AI agent governance?
The control points that keep agent work reviewable before it changes trusted state.
Read articleHuqan vs. NeMo Guardrails
A boundary-by-boundary comparison for teams deciding what needs governance.
Read articleBest AI governance tools
A practical shortlist for evidence, policy, action review and evaluation.
Read articleNeMo Guardrails alternatives
Four options compared by the boundary each one actually defends.
Read articleFor builders, teams and investors
Trust infrastructure for AI systems that must survive scrutiny.
Huqan is built for regulated workflows, internal AI tools, agent platforms and security-sensitive automation where a wrong answer, poisoned memory or unaudited action carries real cost.
Every serious AI decision should come with a receipt.
Build with deterministic judgment, local ownership and evidence that remains inspectable after the model has moved on.
Email us about a pilotaliulu@ai-ulu.com