AI Use Policy

Purpose

To ensure that the artificial intelligence (AI) features of the NIMS Logic platform are designed, operated and changed in a way that protects institutional data, keeps people in control of every action, and keeps AI risk identified, measured and managed.

Scope

This policy applies to every AI feature of the NIMS Logic platform operated by Gov Claim Technologies, LLC (dba NIMS Logic), including the Sentinel assistant, its knowledge retrieval over NIMS Logic documentation, and its evidence retrieval over incident files. It applies to all employees and contractors who design, build, operate or change those features.

Personnel use of third-party AI tools in their own work is outside the scope of this version of the policy.

Roles and responsibilities

• The CTO owns this policy and the AI features it governs, approves any new model, new AI feature, new AI write action or new data source for AI, and maintains the AI Risk Assessment and the AI Impact Assessment.
• The COO approves exceptions to this policy.
• Engineers building or changing an AI feature are responsible for keeping it within the requirements below.

Approved models and providers

• AI models shall be consumed only as dedicated deployments in NIMS Logic's own Azure AI Foundry resource, under NIMS Logic's Azure subscription, in a United States region.
• Model access shall use managed-identity authentication. No model API key shall be embedded in application code.
• Model versions, and each deployment's version-upgrade setting, shall be specified explicitly in infrastructure configuration. Where a deployment accepts the provider's new default version automatically, that is recorded in the AI Risk Assessment.
• Adopting a new model, or replacing the model behind an AI feature, requires CTO approval and a passing run of the AI evaluation suite against the platform's real session path before the change reaches production.
• Each model deployment shall carry a content-safety policy.

Institutional data and AI

• Institutional data shall not be used to train, fine-tune or otherwise alter any AI model. NIMS Logic operates no model-training pipeline.
• An AI feature shall read institutional data only as the requesting user, through the same authorization the rest of the platform enforces. AI features shall not run under a service account with standing access to institutional data.
• The scope an AI tool acts in shall come from the authenticated user's context and never from a value produced by the model.
• Institutional data may be held in an AI retrieval index only to find records for a user who is already authorized to see them. Every result retrieved from such an index shall be re-checked against the requesting user's permissions before it is shown or passed to the model, and an entry shall be removed from the index when its source record is deleted.
• Sentinel conversation transcripts are Confidential data under the Data Management Policy and are handled, retained and disposed of under it.

Human oversight

• An AI feature shall not change a record on its own. Every state-changing action an AI feature proposes shall execute only after the user explicitly confirms it, through a single-use confirmation bound to that exact action.
• A destructive action shall require a second confirmation.
• The set of AI write actions shall be limited to those approved by the CTO. As of this version they are recording an ICS 211 check-in and initiating an ICS 221 demobilization.
• AI features do not make automated decisions about individuals. Decisions remain with the authorized user.

Required safeguards

Every AI feature shall carry, at minimum:

• instructions in the base system prompt that treat all user text, tool results and record content as data and never as instructions, placed so that a change of model cannot remove them;
• AI output rendered as text and never as executable markup;
• limits on the length of a response and the number of tool calls in a single turn;
• a per-user request rate limit; and
• the ability to be switched off platform-wide through configuration, without an application release.

Audit and accountability

AI features shall write an audit record for each exchange, each authorization decision, each read of classified data and each executed write, attributed to the requesting user, in line with the platform's audit logging standard.

Evaluation

A change NIMS Logic makes to the model or model tier behind an AI feature shall pass the AI evaluation suite before it reaches production. Changes to an AI feature's base prompt or tools should be evaluated the same way. The suite runs real tasks through the platform's session path and is itself tested to confirm that it fails a deliberately wrong answer.

Operator-tuned prompt layers are appended after the base prompt and cannot replace its safety instructions. They take effect without a release, and this is recorded in the AI Risk Assessment.

AI risk and incidents

• NIMS Logic shall maintain an AI Risk Assessment and an AI Impact Assessment. Both are reviewed at least annually, and again before a new model, new AI feature, new AI write action or new source of data for AI reaches production.
• AI incidents are handled under the AI Incident Response Procedure, which operates within the Incident Response Plan.

Exceptions

Requests for an exception to this policy must be submitted to the COO for approval. Exceptions shall be documented.

Violations & enforcement

Any known violations of this policy should be reported to the CTO. Violations of this policy can result in immediate withdrawal or suspension of system and network privileges and/or disciplinary action in accordance with company procedures up to and including termination of employment.

Version history