A search system returns no useful passage. The tempting conclusion is that the source material contains no evidence for the claim. That conclusion is convenient, fast, and often unjustified.
Retrieval is a method for finding candidates. It is not an assessment of reality, and it is not even a complete assessment of a closed corpus. Search can fail because the wording differs, the claim is compound, the relevant passage uses an abbreviation, or important context sits across a page boundary. An empty result describes what a particular retrieval process found under particular settings. It does not decide the claim.
This distinction is essential in evidence audit. no_candidates_retrieved should remain a process state. not_found should be a human-reviewed evidence verdict with a clearly stated scope.
Retrieval has a limited job
The job of retrieval is to reduce a large source set to a manageable group of passages for inspection. Lexical search can surface exact terms and rare phrases. Semantic search can find related language. Ranking can combine those signals and place likely evidence near the top.
Each method has blind spots. Exact search may miss paraphrases. Semantic search may favor topical similarity over a decisive detail. A query built from the full report sentence may overlook evidence for one material component. A query simplified too aggressively may retrieve many passages while dropping the date, jurisdiction, or modality that makes the claim specific.
Good retrieval improves review, but it does not replace it. Candidate ranking says where to look first. It does not establish support, contradiction, ambiguity, or absence.
Preserve the empty result honestly
When no candidates are returned, the system should record that outcome without promoting it into a substantive judgment. The record can identify the claim representation, source versions, retrieval configuration, and candidate count. That information is valuable for diagnosis and later evaluation.
It should also open a review path. A reviewer may reformulate the query, search a distinctive name or number, limit the search to one document, inspect nearby pages, or look for evidence expressed through tables and headings. If relevant text is found, the reviewer should be able to select the server-backed span and create an exact anchor outside the original candidate set.
This manual recovery is not an exception that weakens the workflow. It is evidence about the retrieval system. The recovered passage can be linked to the claim, while the original miss remains recorded for future improvement. Provenance should show both how the evidence was eventually found and why it was absent from the first candidate list.
Reserve verdicts for evidence assessment
After an adequate review, a reviewer may conclude that the supplied sources contain no sufficient supporting or contradicting evidence. That is a valid not_found verdict, but its wording must retain the boundary: no sufficient evidence was found in the supplied sources.
It does not mean the claim is false. It does not mean evidence exists nowhere else. It does not establish that the supplied sources are complete or authoritative. It also does not score their credibility. The verdict describes the relationship between one claim and one pinned corpus reviewed through a documented process.
Separating retrieval state from evidence verdict prevents hidden certainty. It stops an infrastructure limitation from becoming an analytical conclusion. It also helps teams compare cases fairly, because a search miss and a reviewed absence no longer share the same label.
Make partial evidence visible
Manual review may reveal that the empty result came from the structure of the claim. A sentence can assert an actor, action, amount, date, and cause at once. The corpus may support some components but not others. Searching for the whole sentence can fail even when several components have clear evidence.
Decomposing the sentence into atomic claims and material components gives the reviewer a better map. Exact anchors can then show which passage supports which component. The final result may be partial support or ambiguity rather than not_found. This is more informative than a forced binary answer and more faithful to the supplied evidence.
Human judgment remains necessary here. Software can validate that an anchor points to the correct source version and text range. It cannot settle every question of relevance, implication, qualifier matching, or materiality through structure alone.
Design for recoverable misses
Hadalith Trace is designed around a closed, user-supplied set of version-pinned PDF sources. Within that scope, retrieval assists an evidence-audit workflow. Reviewers retain the ability to search the full pinned corpus, inspect exact source locations, recover evidence outside suggested candidates, and finalize the assessment themselves.
The principle is simple: absence of retrieval is information about the process, not a verdict about the claim. Record the miss. Preserve the search context. Give the reviewer a route to recover evidence. Only then, after inspecting the supplied material, should the workflow permit a scoped conclusion.
Systems earn trust not by pretending they never miss, but by making their misses visible, recoverable, and impossible to confuse with proof.