A summary you cannot check is a rumour with better formatting

Here is a sentence from a meeting summary. It could be from any product in this category, including one you already pay for.

The team agreed to delay the launch to the second week of November, pending legal review.

It reads well. It is specific. It has a date in it. And there is no way to tell, from the summary alone, whether anybody said it.

Maybe someone did. Maybe two people discussed November and a third said "let's not commit to that yet", and a model resolved the ambiguity in the direction that made the cleanest sentence. A week later, nobody in the room remembers which. The summary is now the record, and the record cannot be checked.

The failure is quiet, which is what makes it expensive

A wrong summary does not announce itself. It is not a crash, it does not show up in an error rate, and the people best placed to catch it are the ones least likely to re-read a document about a meeting they attended.

What it does instead is get forwarded. Somebody who was not there reads the summary, plans around November, and discovers in three weeks that the commitment was never made. At that point the cost is not a wrong sentence, it is a schedule.

What we do instead

Every claim Sprig shows you points at the passage that produced it.

• The recap's key points each carry the moment they came from.
• An answer to a question across meetings names the meeting and the passage, so you can disagree with it.
• A commitment is attached to the sentence that created the obligation, with the person who said it.
• If Sprig cannot produce the passage, it does not make the claim. It says the question is not covered.

That last one is the whole discipline, and it is the one that shows up in the product as an absence. A model that says "not covered" looks worse in a demo than a model that answers everything. It is right more often.

What it costs

Three things, stated plainly, because a post about honesty that hides its own trade-offs is not worth reading:

1. The recap is shorter. Anything we cannot ground does not appear, so Sprig's output is less comprehensive than a tool that will summarise freely.
2. "Not covered" appears more than people expect. A question about something that was implied but never said gets an honest refusal rather than a confident guess.
3. The build is slower. Attribution has to survive every layer, and a feature that breaks the link between a claim and its passage is rejected even when the feature is good.

We think this is the right trade for a meeting record specifically, because the entire value of a meeting record is that it can settle a disagreement about what was said. A summary that can be wrong in undetectable ways cannot settle anything.

How to check us on this

Do not take the claim on trust. That would be funny.

Open any recap and pick a line in the summary. It should take you to the passage behind it. If it does not, that is a bug and we would like to hear about it at our contact page.