Evidence before confidence
How relevance, citations, conflicts and gaps are assessed, and why a limited answer or an abstention is sometimes the correct result of a search.
Preserve the actual question
Search should preserve names, identifiers, quoted phrases and source restrictions. Exact identifiers matter more than a plausible paraphrase. If a question is ambiguous, refining it can be more useful than a broader search.
Support each claim
Source identifiers and versions are bound to the passages used. A composed answer must cite supplied evidence. Missing support, contradictions or insufficient current evidence should produce a limited answer or abstention, not invented certainty.
Read outcome labels
Complete means the configured answer checks passed, not that the whole dark web was searched. Source-only provides relevant references without a composed answer. No evidence and failed requests are distinct: one describes a search result, the other an execution problem.
Evaluation transparency
No holdout accuracy, recall or benchmark score is published before a judged evaluation has run. Future reports must identify task counts, answerable denominators, versions, language coverage, failures and independent-source limitations. Fixture tests establish software behavior, not broad research quality.
Related help
- Which sources can contribute?
- From your question to evidence you can inspect
- What is protected, and what is not
- A citation starts the checking process
- Compare evidence, coverage and limits
- Useful answers require more than fluent text
Content revision 2026-09-07.