How to Stop AI Hallucination With Evidence Anchoring

A lawyer citing a case that does not exist is not a technology story, it is a structure story. It happens because nothing in the workflow required the AI's claim to trace back to a real source before it left the practice. Waiting for a better model to make hallucination go away is not a plan, because the underlying problem is not that models occasionally get things wrong, all of them do. The problem is a workflow with no checkpoint that catches it. Evidence anchoring is the structural fix, and it does not require a different model at all.

What evidence anchoring actually is

Evidence is one of the six MATTER primitives, and the doctrine behind it is simple to state: the connections are the asset, not the pile of documents. Every claim that ends up in a draft, a memo or a filing should be anchored to its source, page and paragraph, as structure rather than memory. Most practices have thousands of documents and almost none of the connections between a specific assertion and the specific page that supports it. That gap is exactly where hallucination gets in, because there is nowhere for a fabricated claim to visibly fail to attach.

Anchoring closes the gap by making the citation a required field, not an afterthought. A claim without an anchor is not a weaker claim, it is not a claim the system accepts at all.

The five-step structure

Constrain the retrieval pool. Do not let a general-purpose model reason freely across everything it was trained on. Point it only at the specific documents relevant to the matter, the actual filed contract, the actual precedent bundle, the actual case file. A smaller, bounded pool of real source material is the single biggest reduction in fabrication risk, because the model has less room to reach for something plausible-sounding instead of something real.

Require inline sourcing on every factual claim. Not a bibliography at the end of a memo, an inline citation at the point the claim is made, specific enough to check in seconds: this document, this page, this paragraph. A claim that cannot name its source at the point it is made does not get to exist in the draft.

Build an explicit "no source found" path. This is the step most systems skip, and it is the one that matters most. An AI system under pressure to produce an answer will often produce one even when the honest answer is "I could not find support for this in the available documents." The fix is to make that an acceptable, expected output rather than a failure state. A system with no path for "no source found" trains itself, and its users, to treat every fluent answer as a confident one.

Add a verification checkpoint before anything leaves the practice. A named human reviewer, at a named point in the sequence, checks the anchored claims before the document goes to a client or a court. This is not a suggestion for careful practices. It is the non-negotiable last line, because no upstream structure, however good, replaces a human taking responsibility for what leaves the building.

Spot-check the citations, not just the prose. Reviewers under time pressure tend to read a draft for tone and completeness and skim past the citations as though a citation existing is the same as a citation being correct. It is not. Spend the review minutes specifically on whether the anchor actually says what the draft claims it says, not on whether the paragraph reads well.

Why this works when model upgrades alone do not

A better model reduces the rate of fabrication. It does not eliminate the need for a checkpoint, because even a very good model will occasionally produce a fluent, confident, wrong answer, and a workflow with no structural check has no way to catch that regardless of how rare it has become. Evidence anchoring does not depend on model quality improving. It depends on the workflow requiring a source before a claim is allowed to exist, which works the same way whether the underlying model is good or excellent.

What this looks like installed, not theoretical

In practice this means: every drafting prompt includes the actual source documents, not a description of them. Every output template has a citation field that cannot be left blank. Every matter has a named reviewer for anchored claims before anything goes out. And every practice keeps an audit trail, so if a claim is ever questioned later, the anchor that supported it at the time is retrievable, not reconstructed from memory under pressure.

I built the full version of this, the prompt scaffolds, the review checklists and the audit-trail template, into the Evidence Anchoring Kit, because writing the doctrine down is only useful once it is installable rather than aspirational.