Evidence Anchoring in Practice: A Real Walkthrough

I have written before about why evidence anchoring is the structural fix for AI hallucination in legal work, the doctrine of it, the five-step structure, the reasoning behind each step. What I have not done is walk through what it actually looks like installed against one real task, start to finish, the way I would actually run it on a matter rather than describe it in the abstract. That is what this post is for.

The task

Take a demand letter for a personal injury matter, the kind of document I described in the personal injury lifecycle post as needing every damages figure traced to a specific medical record or bill. That is the concrete version of "evidence anchoring." Here is how it actually gets built, step by step, not as doctrine but as a sequence of decisions.

Step one: define the bounded pool before opening any drafting tool

Before a single word gets drafted, the source material gets assembled deliberately: the client's medical records from each treating provider, itemized bills, the police report, wage-loss documentation if applicable. Not a general instruction to "review the file." A specific, closed set of documents, and nothing outside that set is permitted to inform the draft. This step is unglamorous and it is the one people skip under time pressure, because it feels like preparation rather than progress. It is the single highest-leverage step in the whole walkthrough, because everything downstream depends on the pool being complete and closed.

Step two: pass one, extraction, not drafting

The first pass through the bounded pool is not drafting the letter. It is extraction: pulling out every discrete factual claim the letter will eventually need to make, a specific injury, a specific treatment date, a specific bill amount, each one tied at extraction time to the exact document, page and line it came from. This is the step that most people who skip evidence anchoring skip entirely, because it feels redundant with drafting itself. It is not redundant. Extraction with the citation attached, done before a single sentence of prose exists, is what makes anchoring structural rather than something bolted on afterward as a citation-adding exercise nobody actually wants to do.

Step three: pass two, drafting from the extracted, anchored claims

Only now does the actual letter get drafted, and it gets drafted from the extraction list, not from the raw source documents directly. Each paragraph in the demand pulls specific figures and facts from claims that already carry their citation, so the inline sourcing described in the anchoring doctrine is not an afterthought added during review, it is inherited automatically from step two because the claim never existed in the draft without its citation attached.

Step four: the no-source-found check, run deliberately

Before review, the draft gets checked specifically for any claim that does not trace back to an item on the extraction list. This sounds like it should never happen if steps two and three were followed correctly, and in practice it happens anyway, because drafting naturally wants to smooth a sentence, round a number, or generalize a claim slightly beyond what the underlying record actually supports. The check here is blunt: does this sentence, as written, say exactly what the cited document says, no more. Anything that fails gets flagged as unsupported and either gets corrected against the actual record or removed. This is the step that catches the fabrication risk that survives even careful drafting, because fluent smoothing is a different failure mode than outright invention, and it is just as dangerous in a document that will go to an insurance adjuster.

Step five: the human reviewer, spot-checking citations, not prose

The named reviewer for this matter, before the letter leaves the practice, does not read the demand letter for tone and flow first. They open the citation list and spend the review time checking whether the anchor actually supports the specific claim attached to it, three or four spot checks minimum, picked deliberately from the highest-dollar claims in the letter rather than at random. A citation existing is not the same as a citation being correct, and reviewers under time pressure default to trusting a citation that is present rather than verifying what it actually says. This step exists specifically to counter that habit.

What this looks like when it works

The finished demand letter reads like any well-drafted demand letter, nothing about the prose signals that it went through an anchoring process. What is different is what sits behind it: an extraction list, each claim tied to a page and line, that a paralegal, a reviewing partner, or the lawyer six months later trying to remember why a specific figure was included, can retrieve in seconds rather than reconstruct from memory. If an adjuster pushes back on a specific damages figure, the answer to "where does that number come from" is already on file, not something that has to be dug out of the underlying records under negotiation pressure.

What this looks like when it fails

The failure mode is not usually a wildly invented fact, that is the dramatic version people worry about and it does happen, but the more common failure is quieter: a figure that got rounded during drafting, a treatment date that got generalized from "March 2026" to "early 2026" in a way that no longer matches the record precisely, a claim that technically traces to a document but overstates slightly what that document actually says. None of these look like hallucination. All of them are exactly the kind of drift that step four exists to catch, and skipping that step because the draft "looks fine" is how a demand letter ends up making a claim the underlying record does not quite support.

Why the walkthrough matters more than the doctrine

The doctrine is easy to agree with in the abstract and easy to skip under real deadline pressure, because each individual step feels like it is slowing down a task that already has a deadline attached to it. Walking through it against one real document is what makes the actual time cost visible: extraction adds real minutes up front, and it is exactly those minutes that make the review pass fast and specific instead of a slow, general read hoping to catch problems. The full prompt scaffolds, review checklists and audit-trail template built to run this exact sequence are in the Evidence Anchoring Kit.