How to catch omitted constraints in AI-assisted work

Define what counts as a constraint

Start with the decision being supported and the consequence of getting it wrong. Then collect constraints from the assignment, source documents, policies, data definitions, dates, audience expectations, system permissions, and downstream format. Separate mandatory requirements from preferences and assumptions.

The NIST AI RMF Core asks organizations to document intended purpose, context, requirements, targeted scope, human oversight, and knowledge limits. Those categories are useful because a requirement can disappear before a model is used. A vague handoff, inaccessible attachment, stale source, or unstated review rule can all create the same final symptom.

Give every constraint a stable ID. Record its exact source, whether it is mandatory, the work product it affects, the person who can interpret it, and the evidence that will demonstrate compliance. Do not convert ambiguous language into a firm rule without naming the interpretation.

Sources: [1]

Build a constraint-to-test matrix

Use a matrix with six columns: constraint ID, source passage, operational interpretation, affected output, verification test, and disposition. Add a seventh column for uncertainty when the source is incomplete or conflicting.

The NIST AI RMF Playbook Map guidance recommends examining purpose, deployment setting, requirements, user expectations, impacts, and contextual factors. A matrix turns that broad context into checks that a knowledge worker can execute before delivery.

Not every constraint can be verified automatically. Dates, totals, and required sections may support deterministic checks. Tone, legal interpretation, commercial judgment, and materiality usually require an accountable reviewer. Label the test type so a green automated check is not mistaken for full approval.

Sources: [2]

Work a synthetic proposal example

Imagine a team preparing a response to a fictional facilities-services request. The source pack contains a 15-page request, a pricing workbook, a clarification email, and an internal approval policy. The AI-assisted draft is polished but omits four constraints:

The team maps each item to a test. C-01 is checked against the submission calendar and timezone. C-02 is reconciled to the workbook schema. C-03 is checked on the rendered file, not the editable source. C-04 requires a human reviewer to trace every past-performance statement to an approved record.

The draft initially passes spellcheck and arithmetic review but fails three constraint tests. The correction record shows what changed and who approved the interpretation. That evidence is more useful than asking the model to “double-check everything.”

• C-01: the clarification email moves the deadline from 17:00 to 15:00 local time.
• C-02: the price must include optional weekend coverage as a separate line.
• C-03: the executive summary has a two-page limit.
• C-04: any statement about past performance requires an approved internal source.

Detect where constraints disappear

Run the matrix through four handoff points: source intake, task framing, generation, and final assembly. At each point, ask whether the constraint is present, transformed, or lost. A requirement may appear in the prompt yet disappear when another tool summarizes the source. It may survive generation but be removed during layout or channel conversion.

Test negative cases deliberately. Include a superseded date, a qualification in a footnote, a rule that applies only to one region, a missing attachment, and two sources that disagree. The purpose is to expose which stage owns clarification and when work must stop.

Record met, not met, not applicable, or unresolved. Avoid a generic pass status. An unresolved constraint should name the person or source needed, the delivery consequence, and the temporary disposition, such as hold, narrow, or qualify.

Verify the final artifact

Review the rendered deliverable and its decisive source packet. Confirm that every mandatory constraint has a disposition, the cited source is current, calculations use the intended definitions, qualifications remain visible, and unresolved items are surfaced to the decision owner.

Sample a few non-mandatory assumptions too. Repeated omission of preferences may reveal a weak intake process even when the formal requirements are met. Preserve the matrix beside the delivered artifact so later reviewers can distinguish an omitted requirement from a consciously accepted tradeoff.

Two Prune can create a bounded simulation with conflicting sources, hidden constraints, and a defined deliverable, then retain the submitted work and available workflow evidence for human interpretation. It does not certify compliance, approve a business decision, or prove performance outside the assessment. Missing evidence should remain a confidence limitation rather than being treated automatically as failure.