Observation
Something recorded. Traceable to a source record, and wrong only if the source is wrong.
Design Pattern · Sep 10, 2026
Typed Utterances and Honest Uncertainty
Every statement the system makes carries a type: observation, inference, recommendation, automated action, human-approved action, unresolved uncertainty, or prohibited action. Missing data is rendered as missing, and standing constraints hold at every autonomy level.
Under the pillar Zero Dashboard Experiences
Thesis
A reasoning system that does not label the epistemic status of its own statements is asking the reader to guess which sentences are facts.
The argument
The types are not decoration. Each one carries a different obligation on the system and a different response from the reader.
Inference
A conclusion drawn from observations. Carries confidence and can be argued with.
Recommendation
A proposed action with a downside if wrong, alternatives, and a required authorizing role.
Automated action
Something already done under an autonomy level, with the policy and model version in force recorded.
Human-approved action
Something done after review, with what the reviewer changed captured as part of the record.
Unresolved uncertainty
An honest deadlock, often two models disagreeing inside their bands, surfaced with a proposed test rather than a chosen side.
Prohibited action
Something the system was asked or inclined to do and refused, recorded so the refusal is auditable.
Missing is not zero
The second half of this pattern is the treatment of absence. A missing reading is drawn as a gap, not as a flat line. A partial data export marks the affected fields missing rather than carrying forward the last known value. An assembled evidence packet names the element it could not find instead of inferring it.
This choice has a visible cost: it makes the operation look worse. In the prototype's synthetic data, documentation completeness falls five points because partial nightly exports were marked missing rather than defaulted. A less careful system would have shown the higher number and been wrong. Which number an organization prefers to see is a governance decision more than a technical one.
Standing constraints
Some limits do not vary with autonomy level, workflow or configuration: no diagnosis, no prescribing, no independent determination of medical necessity, no fabricated time, and no claim submitted without an authorized human review. A drafted sentence that asserts something no source record contains is blocked before a reviewer sees it.
Constraints of this kind are only credible if their enforcement is recorded. A blocked action that leaves no trace is a policy claim, not a control.
What a legacy vendor would say
That typed output is jargon users will ignore; that displaying uncertainty erodes confidence in the product; and that showing missing data as missing makes the customer's numbers look worse than a competitor's, which loses deals.
The commercial objection is the honest one, and it explains why defaulting is common. It is also the reason to treat this pattern as a procurement question rather than a design preference.
What would settle it
Test whether the labels change behavior. Present matched users with typed and untyped output and measure how they treat an inference: verification rate, override rate, and whether a wrong inference is caught. If the labels do not change behavior, they are ceremony.
Open questions
Whether seven types is the right number or whether three would be used more faithfully. How to keep the labels from becoming invisible through repetition. Whether an organization that adopts missing-is-not-zero can be compared with one that does not, on any shared measure at all.