One local retrieval pass, top three results.
Run a hunt to populate the baseline evidence.
Search, read, score coverage, refine, repeat.
Every new hop appends evidence instead of replacing it.
The saved search is a local lexical heuristic, not an embedding or language-model call. Query and document title plus text are tokenized and stemmed. Each query term contributes an exact count capped at one point five. A missing term at least five characters long may receive point five five when a document term shares a four-character prefix. The term score is weighted point six four; substring matches in document text add point three six. For a single exact occurrence with a text match, the score is one. A fuzzy-only example with no text substring gives point three five two. Both bars share five hundred pixels per score unit. Scores are capped at one; only scores strictly above point zero eight enter search results. These values rank lexical evidence, not probability that an answer is true. One-shot search returns the top three matching documents. The iterative hunt starts from the same query, then searches at most five times, returning at most three results per hop while excluding previously seen document IDs. This gives a fifteen-ID ceiling, not a guarantee of fifteen results. Both bars use thirty pixels per document ceiling. New queries combine up to three original terms longer than five characters with three weighted terms from retrieved text, excluding original stems. This can expose new evidence but can also drift away from the original question. No new documents are fetched; all hops search the loaded corpus. A hunt can stop when no results remain or the generated query is empty. Complete coverage stops it only after at least three hops, even if coverage was already complete sooner. Coverage joins retrieved titles and text and tests four fixed patterns: Singapore or supplier; quality, coating, variance or failure; dual sourcing, Osaka or secondary; and Q3, third quarter or revenue. Two matching categories give point five and four give one. Both bars share four hundred eighty pixels per coverage unit. These categories are tailored to the bundled topic, not generated from an arbitrary question. A document containing the keywords can satisfy coverage without supporting a claim. Composite confidence is a weighted average of retrieved scores by per-hop unique counts, clipped at one; it is not calibrated answer certainty. Selecting a hop changes the displayed accumulated evidence, not the saved canonical run. Graph edges all radiate from a query node, so they show retrieved document scores rather than inter-document reasoning paths. Source grounding, citations and factual adequacy still need independent review. This narration makes no external model or backend request.