← All guides
Recommendation lineage

What to log when software makes a recommendation

An operational console surfaces a clean confidence score telling a steward which site inspection to defer next quarter. Six months later, equipment fails on the deferred node—and the post-mortem hits a wall because all the database kept was "score: 0.24, status: approved." Which sensor tables fed that score? What validation test checked whether winter noise skewed it? Which person held delegated authority to act on it?

Author: Escherbridge. Published . Last reviewed: , America/Denver.

Both GAO-21-519SP (June 2021) and NIST AI RMF 1.0 (NIST AI 100-1, January 2023) address that audit hole directly. GAO highlights why oversight fails without explicit metadata: "inputs and operations are not always visible." Its opening practices require engineers and operators to "Define clear goals and objectives for the AI system to ensure intended outcomes are achieved" (Practice 1.1) and "Define clear roles, responsibilities, and delegation of authority for the AI system to ensure effective operations, timely corrections, and sustained oversight" (Practice 1.2), structured across governance, data, performance, and monitoring.

NIST reinforces the same cadence across its four functions—GOVERN, MAP, MEASURE, and MANAGE—noting that risk checks run continuously across the "AI system lifecycle dimensions." Read the full framework breakdown in AI accountability: NIST AI RMF and the GAO framework or the companion essay A recommendation still needs a name on it.

Five fields every recommendation record needs

When automated tooling proposes a decision with real operational consequences, storing the five fields implied by GAO and NIST turns an opaque nudge into an auditable trail: the stated goal the model was optimizing for, the input snapshot or query reference, the evaluation check run against ground truth, the named human who approved or overrode it, and how ongoing drift is monitored afterward.

Scope and boundaries

This guide is a practical engineering logging pattern derived from public GAO and NIST documents—not legal advice, not a certification, and not an implication that PrimusNeo is a federal system or claims full agency compliance. It includes zero fabricated deployment figures, project failure percentages, or CHAOS rates.

A situation like this

A field coordinator inherits an alert feed that ranks asset maintenance weekly, discovers false negatives piling up, and finds zero historical record of what objective function, feature set, or supervisor authorized the ranking rule.

Sources

  1. U.S. Government Accountability Office, Artificial Intelligence: An Accountability Framework for Federal Agencies and Other Entities, GAO-21-519SP, June 2021. https://www.gao.gov/products/gao-21-519sp
  2. NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
  3. Framework sibling: AI accountability: NIST AI RMF and the GAO framework. Essay: A recommendation still needs a name on it /guides/ai-rmf-accountability · A recommendation still needs a name on it: /blog/a-recommendation-still-needs-a-name

Also in this section

The useful next step is to write down the goal, the data, the check, and the name, or to read the framework companion. If something on this page is wrong, write to us.