One Feature, One Risk Register Slice

Decision memo

Decision: Conditional release, pending a defined evaluation threshold.

Feature: Suggested replies for an English-language support queue. An agent must choose and press Send; autonomous sending is outside scope.

Decision owner: Product lead, with a named safety owner able to stop the release.

This is a hypothetical example, not a report from a deployed system. Its purpose is to show how a risk register can stay attached to one feature decision instead of becoming a catalog of everything an AI product might do.

Decision lens Filled example Consequence for this decision
Accountability Product lead owns go/no-go; safety owner can stop release Ownership exists before evidence is read
Context English support queue; human chooses whether to send Claims do not extend to other languages or autonomous replies
Evidence Use a versioned, representative queue sample; define an error rubric, uncertainty treatment, and threshold before seeing results No performance result is asserted in this memo; release waits for the stated test
Treatment Keep human confirmation; narrow the queue or avoid release if the threshold is missed Mitigation is tied to the feature boundary
Reassessment Reopen the decision when the model, prompt, queue, user group, oversight, tolerance, or evidence materially changes A passed decision is not permanent permission

The empty-looking sentence in the evidence row is deliberate. “We will evaluate it” is not a measurement result. A decision memo should expose that gap rather than replace it with an invented score or an unlabeled impression.

Why slice at the feature boundary

The NIST AI Risk Management Framework Core organizes risk work around Govern, Map, Measure, and Manage. Govern is cross-cutting; Map establishes context and impacts; Measure uses quantitative, qualitative, or mixed methods and should document uncertainty; Manage prioritizes risks and responses such as mitigation, transfer, avoidance, or acceptance. NIST also says the functions are iterative and are not a checklist or prescribed sequence.

The memo above is an editorial adaptation of those ideas, not a NIST form, certification, or compliance claim. Its rows keep four questions visible in one place: who owns the choice, what context the claim covers, what evidence would support it, and what happens when the evidence is weak.

A feature slice is small enough to force specificity. “Customer-support assistant” hides meaningful differences between drafting, retrieval, summarization, and autonomous action. A human-confirmed reply suggestion has a different context and treatment from a reply sent without confirmation. Combining them can make a single reassuring control appear to cover both.

The slice still needs connections to the wider system. Shared data, models, safeguards, and user populations can create risks that no single memo contains. Someone must own those joins and reopen affected decisions when a shared dependency changes.

The debatable tradeoff is this: smaller slices improve accountability, but slicing too finely can hide system-wide interactions behind a stack of tidy local decisions.

The dependency edge should also carry the check that makes its invalidation signal actionable. For example, a slice might depend on a prompt version and a safeguard policy, but “reopen on change” is incomplete unless the register names the comparison, owner, and evidence expected after the change. That turns a graph edge into a review contract rather than a notification.

I would keep the previous decision visible as superseded evidence, then require a new, bounded check for each affected slice. This avoids both failure modes: silently reusing stale approval and reopening every unrelated feature whenever a shared component changes. The rule is a systems-design extension of the RMF’s iterative approach, not a claim that NIST defines this data model.