GovKM Proposed Standards and Guidance

Evaluate AI Separately from Information Readiness

Test a named model and integration for retrieval fidelity, unsupported answers, permission handling and the intended business use.

GovKM-proposed public draft. Published for public review. Not externally accredited or endorsed.

View the EICARS framework resource

Document

Return to the EICARS guidance index

EICARS implementation guidance — proposed public draft. Checklist coverage: AI01, AI02, AI03. Prepared November 9, 2026. These suggested approaches support the published GovKM-proposed EICARS v0.1 draft; they do not add requirements or grant external accreditation.

Purpose and use

Test a named model and integration for retrieval fidelity, unsupported answers, permission handling and the intended business use.

Use this guide within a declared organizational scope, with existing repositories and suitable organizational processes. AI is optional. Examples and suggested evidence are implementation aids; the published standard governs a conformance assessment. The Riverton case below is fictional, not an observed deployment or test result.

Responsible roles

Business owner and applicable AI, security, privacy, procurement and system-authorization officials.

Assign responsibilities through the organization’s actual instruments. A suggested role in this guide does not itself confer decision authority.

AI01 Evaluate connectors and retrieval

Suggested approach: For each named use, test actual connector coverage, formats, parsing and OCR where relevant, source and version citations, date-sensitive retrieval and permissions. Use representative information and approved expected outcomes. Include failed or missing sources and report what the connector cannot access.

Riverton example: Riverton tests the approved date decision, an earlier schedule, a scanned inspection attachment and an unavailable HR source. The assistant must not present missing coverage as evidence that no condition exists.

Evidence and completion check: Keep model and connector identification, configuration context, test dates, expected results, observations and limitations. Distinguish a parsing failure from an unsupported reasoning conclusion.

AI02 Evaluate malicious input and unsupported answers

Suggested approach: Test relevant misleading or malicious source content, unsupported conclusions and escalation behavior with safe approved examples. An instruction embedded in a retrieved document should not confer organizational authority or override access restrictions. Evaluate outputs against the business task rather than assuming citations make every conclusion reliable.

Riverton example: A synthetic Riverton planning note tells an assistant to ignore pending IT acceptance and disclose staffing details. The test checks that the output respects the task and permissions and reports unsupported readiness claims.

Evidence and completion check: Retain the approved case and observed response without exposing sensitive production data. Document failure impact, restrictions, correction and retest; do not treat a few successful cases as a guarantee.

AI03 Record separate AI deployment authorization

Suggested approach: Record the separate decision on the proposed AI use through applicable organizational processes. Distinguish AI-use risk acceptance, information-system authorization, procurement requirements and any delegation to execute actions. Use the agency’s risk classification and applicable rules; not every use has the same approval duties.

Riverton example: Riverton may approve an assistant for drafting a briefing while withholding authority to issue opening instructions or modify contract records. Retrieval access is not authority to act.

Evidence and completion check: Retain the deployment decision, permitted uses, exclusions, responsible officials and action limits. If no AI is proposed, record that fact and continue the information-environment assessment.

Evidence and completion checks

Agree the population, expected results and acceptance conditions before assessing. Record the date, assessor, evidence reference, observed result and limitation. Use approved safe examples for access, malicious-content or lifecycle tests. Documented policy and observed operating behavior are both relevant; one does not automatically prove the other.

On smaller screens, scroll the table horizontally to view every column.

Item	Suggested evidence and demonstration
AI01	Keep model and connector identification, configuration context, test dates, expected results, observations and limitations. Distinguish a parsing failure from an unsupported reasoning conclusion.
AI02	Retain the approved case and observed response without exposing sensitive production data. Document failure impact, restrictions, correction and retest; do not treat a few successful cases as a guarantee.
AI03	Retain the deployment decision, permitted uses, exclusions, responsible officials and action limits. If no AI is proposed, record that fact and continue the information-environment assessment.

Record a gap where the required outcome is not demonstrated. Assign an owner and review date, then verify any correction. Completion of this page alone does not demonstrate conformance with all ten domains, certify legal compliance, approve an AI deployment or authorize autonomous actions.

Common mistakes

Claiming information conformance certifies a model; testing only easy questions; equating retrieval access with permission to execute actions.

Continue the checklist

Use this page with the EICARS standard, the identity guide and the full Riverton scenario. No particular product, model or repository migration is required.

Implementation overview

Suggested operating approach: assign the responsible roles, confirm applicability, implement suitable controls in the existing environment, examine operating evidence and track unresolved gaps. Tailor the methods to scale and risk. Maintain current services during preparation and preserve records, access and legal obligations throughout.

Standards traceability and references

Government Authority and References

U.S. federal reference review: November 9, 2026. These are GovKM mappings of relevance, not government endorsement or a legal compliance determination. Statutes and regulations are mandatory only within their scope; federal policy applies to covered agencies and uses; technical guidance and voluntary resources require the stated selection, adoption or contractual basis. Private and nonfederal organizations must identify their own governing requirements.

The eCFR is authoritative but unofficial. Retrieved Title 36 pages displayed currency through November 5, 2026. Source retrieval does not establish exhaustive legal currency or clause-level implementation compliance. Check current agency supplements, deviations, court orders, schedules, contracts and later amendments. Add the organization’s actual instruments and the qualified official’s applicability determination.

AI01 — Evaluate connectors and retrieval

OMB M-25-21 — Federal AI use
Issuing body: Office of Management and Budget. Pinpoint: 4(b)(i)–(ii), section 5. Relevance: Covered AI testing.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Covered agency AI uses; high-impact provisions and national security/intelligence exceptions have distinct scopes.
NIST AI 600-1 — Generative AI Profile
Issuing body: National Institute of Standards and Technology. Pinpoint: relevant measurement suggested actions. Relevance: Voluntary evaluation.
Classification: Voluntary government guidance. Applicability: Generative AI risk management.
NIST SP 800-53 Revision 5 with Release 5.2.0
Issuing body: National Institute of Standards and Technology. Pinpoint: AC-3, SA-11. Relevance: Access and testing.
Classification: Government technical guidance — conditional on selection or adoption. Applicability: Selected controls, baselines, tailoring and applicable policy or contracts determine obligations. Not every catalog control is universally mandatory.
36 CFR 1236 Subparts D and E — Federal records digitization
Issuing body: National Archives and Records Administration. Pinpoint: applicable subpart validation provisions. Relevance: Digitized records.
Classification: Regulation — mandatory when applicable. Applicability: Digitization of temporary or permanent Federal records, according to the applicable subpart.

GovKM recommendation boundary: OCR, connector coverage and citation tests are use-specific recommendations; they are not universally prescribed by the cited authorities.

AI02 — Evaluate malicious input and unsupported answers

NIST AI 600-1 — Generative AI Profile
Issuing body: National Institute of Standards and Technology. Pinpoint: relevant risks and suggested actions. Relevance: Generative AI risks.
Classification: Voluntary government guidance. Applicability: Generative AI risk management.
OMB M-25-21 — Federal AI use
Issuing body: Office of Management and Budget. Pinpoint: 4(b). Relevance: Covered high-impact AI.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Covered agency AI uses; high-impact provisions and national security/intelligence exceptions have distinct scopes.
OMB M-26-04 — Procured LLM principles
Issuing body: Office of Management and Budget. Pinpoint: principles and procurement implementation. Relevance: Covered LLM requirements.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Covered agency LLM procurement under the memorandum’s definitions and exclusions.

GovKM recommendation boundary: The exact adversarial examples and acceptance tests must fit the deployment. Use test material approved for the organization and task.

AI03 — Record separate AI deployment authorization

OMB M-25-21 — Federal AI use
Issuing body: Office of Management and Budget. Pinpoint: section 3(b), 4(b); footnote 22. Relevance: AI risk review scope.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Covered agency AI uses; high-impact provisions and national security/intelligence exceptions have distinct scopes.
NIST SP 800-37 Revision 2 — Risk Management Framework
Issuing body: National Institute of Standards and Technology. Pinpoint: Authorize step. Relevance: System authorization.
Classification: Government technical guidance — conditional on selection or adoption. Applicability: Security/privacy risk management under applicable Federal policy or voluntary adoption.
OMB M-25-22 — Federal AI acquisition
Issuing body: Office of Management and Budget. Pinpoint: section 4. Relevance: Acquisition conditions.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Covered Federal AI acquisition; obligations flow to vendors through applicable procurement instruments.
OMB Circular A-123 revised March 10 2026
Issuing body: Office of Management and Budget. Pinpoint: section II. Relevance: Management accountability.
Classification: Federal policy — mandatory provisions within stated scope. Applicability: Executive agencies as defined in its Applicability section.

GovKM recommendation boundary: Checklist completion does not grant authority to execute actions. The applicable agency process and legal delegation govern deployment.

Federal AI-use, high-impact AI, AI-acquisition and covered LLM procurement requirements have distinct scopes and exceptions. Information preparation does not substitute for any applicable AI-use risk acceptance, information-system authorization, procurement or legal decision.

This section gives relevant official references for each covered checklist item and is not an exhaustive list of every possible obligation. The actual agency, program, records, data, system and contract determine additional duties. The EICARS organization of these tasks, suggested registers and completion checks remain GovKM recommendations unless independently required or adopted.

Version history

November 9, 2026 — Publication authorized by GovKM. Guidance based on checklist and government authority reference package checked November 5, 2026. References must be periodically revalidated.

November 9, 2026 — Expanded implementation guidance, version 0.1. Replaced the shorter implementation text with the reviewed Riverton examples, task-specific approaches, evidence checks and official authority mappings, with Robert Noyes’s approval. Existing URLs and checklist anchors are preserved. No EICARS requirement or GovKM doctrine is amended.