NVIDIA × LHINOS — Human Authority for Agentic AI

01 · Understand in 90 seconds

Not another guardrail. A different decision.

LHINOS is not positioned as a replacement for NeMo Guardrails, OpenShell or NVIDIA runtime security. The proposed addition is an explicit human-authority decision for consequential business actions.

NVIDIA · existing technical controls

Can the system safely and technically perform this action?

NVIDIA already provides strong machinery for controlling model and agent behavior.

NeMo Guardrails can apply input, output, retrieval, execution and tool-related controls.
Current tool-calling rails can validate declared tool calls and their structure.
OpenShell applies policy to sandbox filesystem, network, process and inference behavior.

LHINOS · proposed complementary layer

Does this agent have legitimate human authority for this effect?

The decision is not “is the tool technically allowed?” but “is this specific consequential act covered by a valid human mandate?”

Named actor, action, target and purpose.
Bounded scope, threshold, validity and action budget.
Explicit human confirmation when authority is missing or exceeded.
Reconstructable evidence for the authority decision.

CapabilityWhat can the agent do?Model, tools, runtime, integrations.

Technical permissionWhat does policy permit?Guardrails, sandbox and security policy.

Human authorityWhat has a human actually authorized?Mandate, purpose, scope, time, budget, evidence.

02 · Run the proof

Give the agent capability. Do not give it authority by accident.

Configure one consequential action. The engine first respects the NVIDIA technical-policy result, then evaluates whether valid human authority exists. A one-shot grant is consumed after use.

LHINOS Navigator · Authority Request Builderinteractive prototype

Requested consequential action

Agent capability

NVIDIA technical policy result

Human authority source

Purpose

Unknown does not become ALLOW. The agent cannot self-issue or widen authority. This browser proof creates no external effect.

Decision tracenot evaluated

1

Agent capabilityTool and credentials exist.

CAPABLE

2

NVIDIA technical policyRuntime/security decision remains independent.

NOT RUN

3

LHINOS human authorityMandate, scope, purpose, time, limit and evidence.

NOT RUN

4

External effectNo real-world action is executed in this assessment room.

NOT EXECUTED

Authority Decision

NOT EVALUATED

Run the proof to produce a bounded decision and evidence receipt.

{
  "status": "NOT_EVALUATED",
  "external_effect": false
}

Proof ready. No external action has been executed.

03 · Inspect the boundary

Small integration surface. Hard semantic separation.

The smallest useful NVIDIA × LHINOS validation is one agent, one consequential tool call and one independent authority decision. No replacement of NVIDIA guardrails or runtime policy is required.

NVIDIA agent stack

Capability + technical policy

Agent reasoning, tool invocation, guardrails, sandbox and runtime restrictions continue to determine whether the action is technically admissible.

→

LHINOS authority decision point

Human mandate + evidence

Evaluate actor, action, target, purpose, scope, threshold, validity, budget and authority evidence. Return ALLOW, DENY or HUMAN_CONFIRMATION_REQUIRED.

→

Consequential effect

Execute only when both layers permit

Technical permission does not create human authority. Human authority does not bypass technical policy. Both must hold.

Proposed AuthorityRequest

{
  "actor": "agent:enterprise-ops",
  "action": "customer_record.update",
  "target": "crm:case-1842",
  "purpose": "approved_case_resolution",
  "scope": ["status", "next_step"],
  "consequence_class": "external_effect",
  "requested_budget": 1,
  "evidence_refs": ["human_mandate:..."]
}

Proposed AuthorityDecision

{
  "decision": "ALLOW",
  "grant": {
    "action_budget": 1,
    "scope": ["status", "next_step"],
    "valid_until": "...",
    "self_widening": false
  },
  "reason": "bounded human mandate matches",
  "evidence_receipt": "..."
}

01Stop capability without authority. A technically capable agent must still be deniable.

02Explicit human handback. Missing or exceeded authority produces a human confirmation boundary.

03Bound the grant. Actor, action, target, purpose, threshold, time and budget remain constrained.

04No self-widening. The agent cannot expand rights, duration or scope on its own.

05Consume one-shot authority. A one-action grant cannot be silently reused.

06Reconstruct the decision. Evidence explains why the effect was or was not eligible.

04 · Challenge it

Do not accept the pitch. Try to break the distinction.

ChatTHINOS is deliberately instructed to be critical rather than promotional. Ask where this overlaps with NVIDIA, what would falsify the fit, or what the smallest integration proof would require.

Technical challenge prompts

The right result may be “you already do this.”

If NVIDIA already expresses the same human-authority semantics end-to-end — including authority source, purpose, target, bounded scope, validity, one-shot budget and reconstructable evidence — that is a valid architectural falsifier.

Requested NVIDIA outcome: technical assessment, asynchronously.

No meeting is required at this stage. The useful internal answer is one of three: architecture fit worth testing, material overlap with existing NVIDIA capability, or not relevant.

Open 2-page brief Fit worth testing Likely overlap Not relevant

NVIDIA public documentation used for architectural context

These references are included to avoid claiming a gap that NVIDIA already addresses.

NeMo Guardrails · Tool Calling OpenShell · Security Controls

ChatTHINOS · NVIDIA Review ModeLive architecture explainer · no autonomous external actions

checking…

I am here to help you test the architectural claim, not to sell it. Run the Authority Proof or ask me where this overlaps with NVIDIA's existing controls.

Current proof state is supplied as context.No external action execution.