Papers already leave out the dead ends and fixes behind a result, and abstracts cut what little survives.

This reads 'recovery ideas' as the material about how things go wrong and get fixed (dead ends, failed attempts, the repairs that follow) and asks why abstracts and excerpts drop it while full papers keep some of it. The corpus has no note that studies abstracts directly, so what follows is built from neighboring notes on how compressing research and experience loses this kind of content.

Start with what a paper already is. One note describes the published paper as a lossy compiler: publishing imposes a 'Storytelling Tax' that erases process, failed branches and tacit reasoning, plus an 'Engineering Tax' that leaves out implementation detail Can research papers preserve the experiments that failed?. An abstract is a second round of compression on top of that first one. It is written to state the result, not the route, so the detours and the fixes that made the result possible are the first things cut. Whatever recovery knowledge survives tends to sit in the body of the paper, where the authors explain what they tried.

The details that get cut are also the ones that change what a reader does with the paper. In one study of self-evolving agents, changing the raw experience an agent had accumulated changed its behavior a lot, while changing the condensed summary of that experience barely mattered. The summaries had lost the critical details, and the models leaned on immediate context instead Why do LLM agents ignore condensed experience summaries?. A recovery idea is a detail of that kind. As an illustration, one note describes models that finetune on documents repeated in a cycle and start recovering performance on a document before they see it again, an effect that grows with scale Do networks recover from forgetting before re-encountering documents?. A one-line summary of 'models forget when trained on new data' would flatten that finding into the old story.

You also can't easily search your way to what an abstract leaves out. A query only carries what you already know to ask for. One note shows that a model's partial answer can expose gaps the original question couldn't express, and that using the draft answer as the next query pulls in what was missing Can a model's partial response guide what to retrieve next?. Another shows that retrieving isolated chunks destroys the document's structure, and that building a global summary first lets scattered evidence be found by its role in the document rather than by surface similarity Can building a document map first improve retrieval over long texts?. A 'what went wrong and how it was fixed' passage has a role but not a topic. It usually doesn't resemble the abstract's vocabulary, so nothing in the abstract points to it.

The loss is also hard to notice. Another note finds that frontier models degrade documents silently, keeping the surface looking intact, while weaker models visibly delete content Does model capability change how documents degrade?. That note is about document editing, not summarizing. Still, it suggests a compressed version can read as complete when it isn't. The practical move, which is my inference and not something a note tests, is to treat an abstract as a first draft of what you're looking for. Guess what the recovery story might be, then check the full text for the passages that confirm or contradict your guess.