Can AI Draft a Police Report Without Making Things Up?

Vendors are already selling AI tools that turn an officer's notes into a draft report narrative. A police report is a legal document: prosecutors, defense attorneys and judges rely on it, and an officer signs it under penalty of perjury. So the question is not whether the draft reads well. It is whether every sentence in it is true.

This test gave two free AI models the same structured facts for five Las Vegas incidents (a DUI stop, a domestic disturbance, a burglary, a weapon found during a traffic stop, and a warrant arrest) and asked each for a first-person, past-tense, chronological narrative with no opinions. Then every draft was checked two ways: did it include each required fact, and did it add anything that was not in the input?

In a separate test of a small model running locally, raising the randomness setting (temperature) made it worse in ways that would matter in court:

Lowest randomness (0.1): the narrative said a drug field test was inconclusive, even though the input stated the field test was positive.

Medium (0.7): it invented a trip from Las Vegas to Los Angeles, and described the white powder found as containing chemicals used in explosives, in a simple possession arrest.

High (1.2): it introduced new locations and an entire second suspect, "James Smith," who did not exist.

A fluent draft is not a verified draft. If an officer signs a narrative containing an invented detail, the consequences reach the courtroom: impeached testimony, dismissed cases, or a fabricated witness that someone tries to subpoena.

Bigger helps, but does not solve it. The large model invented fewer facts, took five times as long, and still failed most reports. Any deployment needs a line-by-line check of the draft against the officer's own notes, and a policy that the officer, not the tool, is the author.

Measure before you buy. A five-scenario checklist like this one is cheap to build and catches the failure that matters. It is the kind of test an agency can ask a vendor to pass.

Both models were called through OpenRouter's free tier at temperature 0, with the same system prompt and the same five scenario prompts (fictional people, real Nevada statute numbers). Model IDs: liquid/lfm-2.5-2.6b:free and nvidia/nemotron-3-ultra-550b-a55b:free. Each scenario had a checklist of required facts (statutes, test results, booking location) and banned phrasing (opinions, conclusions). An invented fact is a specific factual claim in the narrative that is not in the input; a report passes strict review only if it includes every required fact and invents at most one.

This is a five-report pilot, not a validation study: results are from one run and will vary. The per-answer evidence is in the dataset linked below.