Intelligence work often arrives as a polished document. Its conclusions may be clear, its language careful, and its sources listed at the end. Yet a reviewer who needs to test one sentence still faces a difficult question: which exact piece of evidence supports this exact claim?
A bibliography is not enough. A link to a long report is not enough. Even a quoted paragraph can hide an important mismatch in time, geography, quantity, or certainty. Verifiable intelligence requires a path from the wording in the final report to the precise evidence and analytical steps beneath it.
Begin with the atomic claim
A report sentence can contain several propositions. Consider a sentence that says a company opened two facilities in 2024, increased capacity, and became the regional market leader. Those statements may rely on different sources. One may be directly supported, another may be inferred, and the last may not appear in the supplied material at all.
Review becomes more reliable when compound sentences are decomposed into atomic claims. An atomic claim is small enough to assess against evidence without losing the context that gives it meaning. It keeps material qualifiers such as time, place, amount, modality, and attribution visible. Removing those qualifiers may make retrieval easier, but it can also change the claim.
The faithful report wording should therefore remain authoritative. A separate retrieval-oriented form can resolve references or add context, but that derived wording must stay visible, editable, and linked to the original. Search convenience should never silently rewrite what the analyst actually asserted.
Anchor evidence precisely
Once a claim is clear, evidence needs an exact anchor. A useful anchor identifies the source version, page, text span, and quoted passage used in the assessment. It should resolve against the same accepted document bytes later, not whichever copy happens to be available at review time.
This precision matters because documents change. A corrected edition can alter pagination. A downloaded file can be replaced under the same name. A quotation without version context can become impossible to reproduce. Version-pinned sources and exact anchors let another reviewer return to the same place and inspect the same material.
An anchor also makes disagreement productive. Reviewers can distinguish a retrieval problem from an interpretation problem. They can see whether the cited passage is relevant, whether it covers every material component, and whether surrounding context changes its meaning.
Keep verdicts inside their boundary
Evidence assessment is not a universal truth machine. A verdict should describe the relationship between a claim and the supplied sources. It can record that the material supports, partially supports, contradicts, or leaves the claim ambiguous. It can also record that sufficient evidence was not found.
That last state needs careful language. not_found means no sufficient supporting or contradicting evidence was found in the supplied sources. It does not mean the claim is false in the world. Source identification also does not establish source credibility. Those are separate judgments that require separate methods and accountable human reasoning.
Keeping this boundary explicit prevents a narrow audit result from acquiring more authority than it deserves. It also gives downstream readers a useful answer: not certainty at any cost, but a clear account of what the reviewed corpus can and cannot substantiate.
Record the transformation, not only the result
AI-assisted analysis can help segment claims, retrieve candidate passages, and prepare draft assessments. Each transformation introduces choices. A system may resolve a pronoun, select search terms, rank passages, or interpret whether a qualifier is material. Those choices should be recorded as part of provenance.
The goal is not to preserve private reasoning or imply that a model is infallible. The goal is to retain enough structured context to understand how an output was produced: which claim version was used, which source versions were searched, which candidates were considered, and which evidence a reviewer accepted or rejected.
This turns provenance into an operational property rather than a decorative citation layer added at the end.
Human review is the accountable boundary
Retrieval can miss evidence. A model can overstate a passage. A perfectly valid citation can still fail to support the full claim. Human review remains necessary for decisions that depend on semantics, context, and materiality.
Hadalith Trace is designed for that boundary. Its initial scope is a completed report and a closed, user-supplied set of version-pinned PDF sources. It supports an evidence-audit workflow in which reviewers can inspect exact anchors, search the supplied corpus, recover missed evidence, and finalize scoped assessments.
The resulting record does not eliminate judgment. It makes judgment inspectable. Every conclusion can retain a route back to the source, every transformation can have a place in the record, and every final verdict can remain accountable to the evidence actually reviewed.