A useful summary tells you what the records support and what they leave unresolved. An incomplete dataset makes that distinction especially important: an assistant can produce a fluent account of events that the input never established.
This exercise uses invented records. No customer data or actual model response appears below. Start with the AI Control Layer routine, then use the evaluation cases to repeat the exercise.
Define the question before reading the answer
Use September 13, 2026 as the fixed reference date. For this exercise, an item is overdue only when it is unfinished and its due date is before that date. Due today is a separate category. Missing dates mean unknown, not overdue.
Ask: “Summarize the schedule position of these four items. Cite item IDs. Separate overdue, due today, completed and unknown. Do not infer causes or assign blame.”
| ID | Task | Status | Due date |
|---|---|---|---|
| DEMO-1 | Prepare test checklist | In progress | 2026-09-12 |
| DEMO-2 | Review sample export | To do | 2026-09-13 |
| DEMO-3 | Confirm mock handover | Done | 2026-09-11 |
| DEMO-4 | Write restore notes | To do | Missing |
There are four records, including three unfinished items. Only one is known to be overdue. The completed item's past due date does not prove it was completed late; there is no completion timestamp.
Mark every unsupported claim
Consider this deliberately flawed sample:
“Three tasks are late because the team has not reviewed them. The handover was completed two days late. Everything should finish tomorrow.”
The first sentence overstates the overdue count and invents a cause. The second invents a completion time. The last predicts delivery without supporting evidence. Better wording cannot rescue those claims; they need removal or additional authorized evidence.
Review each sentence with three questions: Which record supports it? Does that record establish the whole statement? What would I need to know to say more?
Write a supported version
A defensible answer is:
“DEMO-1 is overdue under the stated rule. DEMO-2 is due today. DEMO-3 is marked done; the records do not show when it finished. DEMO-4 has no due date, so its schedule position is unknown. Three items remain unfinished. Causes and expected completion dates are not provided.”
This is an expected answer for the exercise, not a model performance claim. Different wording is acceptable if it preserves the facts and uncertainty.
Ask for the smallest missing input
If the decision depends on DEMO-4, ask its authorized owner to confirm the due date. If it depends on whether the handover was late, request the relevant completion timestamp. You do not need the whole team's message history to answer either question.
Do not silently widen the scope to another project. A request for more information must still follow the reader's permissions. The app control guide explains why app scope and ordinary document review are different checks.
Keep the review record
Save the exact input, prompt, output and reference date. Record factual mistakes separately from style preferences. A useful rubric has four rows: correct classification, correct totals, cited evidence and no invented facts. Treat an unsupported cause or completion claim as a failure even if the prose reads well.
If you change the prompt, rerun the original case and one variation: remove another due date, change a completed item to unfinished, or add a conflicting note. Record the variation rather than overwriting the original fixture. The small evaluation-set guide shows how to organize those runs.
Close the exercise by deleting temporary copies from the tool you used where supported. Retain only the sanitized evidence needed for your review. A passing summary exercise establishes behavior on that case; it does not certify the connected application's access controls.