Risk & Accountability

As with any professional decision:

• responsibility lies with the identified human who made the decision
• within their qualifications, authorizations, and legal obligations

HumanLayer:

• never decides
• never modifies the decision
• always keeps the link between the decision and the Sentinel’s identity (the client sees a pseudonymous reference)

👉 HumanLayer provides the evidence, not the decision.

HumanLayer implements several safeguards:

• strict scope of authorized decisions
• logging of sensitive actions (logins, decisions, API keys, exports)
• possibility of immediate suspension

Client organizations can:

• report a Sentinel to HumanLayer, which can suspend them
• require a second opinion
• export their signed decisions and ask HumanLayer for a targeted audit

👉 Abuse is more visible, not less.

HumanLayer:

• records the context as received (summary capped at 5,000 characters, context_json at 50 KB)
• timestamps the information
• makes it read-only once submitted

The Sentinel:

• decides based on that context
• can reject with a justification (e.g., insufficient context) or escalate to another Sentinel

👉 In case of incident, it's always possible to determine:

• what the AI provided
• what the human saw
• what the decision was based on

No.

HumanLayer explicitly rejects:

• illegal decisions
• uses intended to circumvent regulations
• requests outside the defined scope

The API rejects requests outside your plan or outside supported domains; unlawful uses are excluded by contract, and HumanLayer can suspend an API key.

• The decision: the Sentinel
• The usage framework: the client organization
• Orchestration and traceability: HumanLayer

This clear separation is essential and contractual.

HumanLayer is not a decision-making actor.

Its liability is limited to:

• proper service execution
• traceability
• availability per SLA

👉 It is an infrastructure provider, not a Sentinel.

Depending on the decision level:

• Sentinels operate within their usual professional framework
• some decisions may require specific insurance
• additional coverage can be offered as an enterprise option

The exact terms are contractual and transparent.

Yes, if your agent waits for the decision.

HumanLayer never approves anything automatically: until a Sentinel has rendered an approved decision, justified and signed, your agent should not act. If in doubt, request a second opinion (second_opinion_of). Risk thresholds, automatic rejection rules, mandatory escalations and multi-signature validations are not available.

👉 No approval without a signed human decision.