Out of five links saved on an ordinary day, there is almost always one that resists. An article behind a paywall. A video with no automatic transcript. A page deleted between the moment you shared it and the moment the system went to fetch it. A scanned PDF whose text is only an image.
At that exact point, a generative system has two options, and the choice between them defines the product almost entirely.
The easy option
The first option is to let the model write anyway. It knows the headline, it knows the outlet, it has probably met the subject during training. It will produce a plausible paragraph, well turned, consistent with the rest of the episode. Nobody will notice a thing.
That is exactly what makes this option dangerous. The paragraph is indistinguishable from the others, except that it rests on nothing you actually saved. A listener cannot know. They will find the gap later, on some detail, and from then on they will doubt everything else. Trust is not lost in proportion to the number of errors.
What we do instead
A source whose extraction fails never becomes text. It is marked with a readable reason — paywall, content unavailable, unsupported format, transcript missing — and it leaves the editorial pipeline.
Then it comes back, in exactly one place: the outro. Twenty seconds at the end of the episode recall the sources used, and name the ones that were dropped. Word for word, it gives sentences like this:
A fifth source was dropped: the article could not be extracted cleanly.
In the app, that same source appears in your library with its status, its reason, and something you can do about it. Retry the extraction. Paste the text by hand. Delete it. Failure is a displayed state, not a hidden error.
The edge case that moved the rule
One exception was introduced after a real case. A video had no usable transcript, but its author published alongside it a document listing their sources and claims. That document did extract cleanly.
The story aired, on the basis of that document, and the script says so explicitly: what is reported comes from the sources the author published, not from the video itself. The general rule is unchanged. What airs always rests on a readable document, and the listener always knows which one.
Why this is a selling point
Naming your gaps seems counterintuitive for a product. An episode that announces a dropped source looks weaker than an episode that announces nothing.
The observed effect is the reverse. A system that says what it could not do makes everything else it asserts credible. It is the same logic as the sentence-by-sentence check, applied one step earlier: at the input, not the output. A briefing that never gets things wrong because it says nothing unverifiable is a briefing you can listen to while walking, checking nothing.