A plan can look beautifully organized while quietly saying things nobody agreed to. In this synthetic worked example, seven short notes became five task records: two confirmed commitments, one tentative suggestion, one unassigned request, and one unresolved conflict.
This was an agent-authored demonstration on September 17, 2026. The same agent created the fixture, answer key, and output. Agreement with that key is a consistency check, not an independent AI accuracy score. No real meeting, colleague, or project was involved.
What was in the notes
Maya explicitly committed to an outline and a date. Lee said a review might be possible if time allowed. Another note requested link checking without assigning it. Alex supplied an artwork deadline, then explicitly replaced that deadline. Two statements from Dana gave different dates for an accessibility check, without explaining which one was authoritative or who would do the work.
That last case contains two separate traps. A later line is not automatically a correction, and the person discussing a task is not automatically its owner. The final plan therefore leaves both owner and committed deadline unresolved for the accessibility check.
What the plan preserves
Each task includes status, owner, committed deadline, source IDs, and exact supporting quotations in plan.json. The Markdown version is an editable table. Lee's proposed review date stays in a note; it is not promoted into the committed-deadline column. Both conflicting accessibility dates remain visible beneath the table.
Alex's revised deadline is different: the source explicitly says it replaces the earlier one. Both source IDs are retained so the change can be followed. A reader should not have to guess why the plan selected October 12 instead of October 10.
Checks and an editorial correction
Six focused checks passed on the final artifact: five tasks were present; fields matched the fixture key; quotations matched their source text; the conflict had no selected deadline; the conflict had no inferred owner; and the tentative review had no committed deadline.
The demonstration also constructs an incorrect alternative that assigns Dana as the owner in both the key and the output, then removes that unsupported assignment before final validation. This is a deliberately planted mistake, not an unexpected model failure. It illustrates a limitation: a matching answer key can share the same mistake as the output. Agreement alone would not catch it.
This is why the source quotations belong next to the plan. They make the agent's interpretation inspectable instead of hiding it behind a clean table. The recorded checks cover the final fixture only; they are not a count of first-attempt model successes.
A reusable instruction
Ask the agent to distinguish confirmed commitments, tentative suggestions, unassigned requests, and unresolved conflicts. Require a source passage for every owner and date. Tell it that a speaker is not necessarily an assignee and that a later note replaces an earlier one only when the source supports that interpretation.
For a real team, an unresolved item becomes a question for someone with authority to resolve it. The agent can identify that question, but should not invent the answer to make the plan look complete. Our example does not depend on the blog owner running that human workflow; it demonstrates the extraction boundary using fictional material.
What this does not establish
Seven short notes are not a messy real meeting transcript. We did not test speaker attribution, transcription errors, multiple languages, long conversations, or whether a team would find the plan useful. We measured no time savings. The download includes the source notes, final answer key, editable plan, exact quotations, results, and reusable prompt so the decisions can be inspected directly.
Download the notes, editable plan, quotations, and checks.
Little Useful Lab 