“The export button is broken - when I click it, nothing happens and there’s no error.”
Reproduces on Chrome 121 + Safari 17. Same step in dashboard. Audio confirms.
Summaries are excellent for catching up and useless for settling arguments. This guide separates the two jobs: why summarisation loses the detail that matters, and what a citable record contains instead.
“The export button is broken - when I click it, nothing happens and there’s no error.”
Reproduces on Chrome 121 + Safari 17. Same step in dashboard. Audio confirms.
None of these are model quality problems. They are consequences of the format.
A summary optimises for readability, so it drops the version number, the literal error string, the conditional that made the approval conditional. Those are precisely the details a dispute turns on.
Six people speak, one paragraph is produced. "The team agreed" hides who agreed, who objected, and who was silent. Consensus gets manufactured in the summarising step.
A summary reads with the same certainty whether a point was stated plainly or inferred from a half sentence. Nothing in the output distinguishes the two, so the reader cannot calibrate.
This is the whole idea: the record is a list of claims, and every claim carries its own proof.
One sentence, one assertion, so it can be individually accepted or challenged.
Who said it, by name, not by role or by team.
Their words, verbatim, long enough to carry the qualifier.
The second the quote begins, so playback is one click away.
The screen state at that moment when the claim is about something visible.
Whether the claim was stated, implied, or contested, surfaced rather than smoothed.
Not every call. The ones where scope, money, dates, or acceptance are decided.
A note is a pointer. If the recording is gone, the pointer dangles and the note becomes an assertion again.
Publish the readable summary if you like, but generate it from the cited record so the two cannot diverge.
A record that flattens dissent will be contradicted by the first person who remembers dissenting.
Long enough to defend the decision, short enough to honour privacy commitments. Then make deletion actually remove derived copies.
Nobody wants to read citations to find out whether a call was useful. Keep the summary. The argument here is about which artifact is authoritative: derive the readable version from the cited record, so improving the prose can never quietly change what the meeting decided.
That ordering is also what makes an answer safe to automate. When retrieval comes first and composition is constrained to what was retrieved, an unsupported question produces a refusal instead of a confident invention.
Or it could sit in a Drive folder nobody opens again. The difference is whether it has citations attached.