When should a research agent ask about a missing unit instead of guessing?

Identify which choice would change the result

NIST’s AI Risk Management Framework emphasizes context, intended use, limitations and differentiated human responsibilities. The decision records below are an original proposed way to apply that general concern to a financial-research input. NIST does not prescribe these cases or establish that a particular chatbot handles them correctly.

An ambiguity is material here if its alternative interpretations produce different numeric results or permissions. Harmless wording choices need not stop every task. A conceptual explanation can sometimes show both interpretations conditionally, while a calculation presented as the user’s own result needs the missing input resolved.

Write three explicit ambiguity records

Case A says return 8 on an initial $1,250 with no other context. If 8 means 8%, the hypothetical gain is $100; if it means $8, the gain is $8 and return is 0.64%. Record status: unit unresolved. Ask whether 8 is a percentage return or a currency amount, and for the relevant measurement period.

Case B says position 100 while an explicitly fictional unit price is $25. If 100 means shares, market value is $2,500; if it means dollars, the quantity is four shares. Record status: quantity unit unresolved. Asking shares or dollars is necessary before presenting one exposure as the intended answer.

Case C says loss limit 2% with a fictional $100,000 portfolio and $10,000 holding. Two percent of portfolio value is $2,000; two percent of the holding is $200. Record status: denominator unresolved. Neither interpretation grants authority to place an order, and the numbers are not recommended limits.

Complete one input card after a fictional clarification

For Case A, stipulate a teaching reply: 8 means an 8% total return over one specified month, in the same currency, with no external cash flows. The completed card records initial value $1,250; measure 8%; period one stated month; currency unchanged; flow assumption none; permitted output explanatory calculation only.

The conditional calculation is then gain = 1,250 × 0.08 = $100 and ending value = $1,350. The card also records that this was an invented clarification, not an actual user response or observed investment result. Cases B and C remain unresolved until their own missing definitions are supplied.

Continue useful work without converting a guess into authority

While waiting, an agent can explain the alternatives and identify the source fields needed for reconciliation. It should label each branch as conditional and avoid overwriting the original ambiguous text. Asking for a unit should not become a request for an entire account statement, private strategy or unrelated personal information.

If an ambiguity affects an external action, stop that action until the appropriate instruction is clear. A resolved unit does not automatically resolve permissions, suitability or factual accuracy. This guide describes a proposed review method; it is not a claim that Thesis Lab supports brokerage execution or automatically enforces every input-card field.

Method, sources, and limits

The three fictional records and calculations were checked on 27 October 2026. They demonstrate interpretation differences, not a measured improvement in agent reliability. AI assisted the draft; separate AI verification is required before publication and is not qualified human financial review.