Detective Marcus Reyes had twelve hours of interview audio, a 340-page case file, and a charging deadline at the end of the week. The suspect interview alone ran ninety minutes. He opened the agency's AI workspace, pasted his question into the chat box, and typed: "Summarize the suspect's account of where he was on the night of the fourteenth and flag any inconsistencies with the victim's statement." Then he stopped, cursor blinking, and did the one thing that separates a detective who uses AI as a tool from a detective who uses AI as a liability. He scrolled back up and attached the actual interview transcript and the actual victim statement to the conversation first. Without those files, the model would have answered his question. It would have produced a fluent, confident, organized summary of a suspect interview that it had never read, built entirely from its statistical memory of what suspect interviews tend to contain. Reyes did not want what a typical interview contains. He wanted what this interview contained. That single discipline, feeding the model the record and constraining it to the record, is the subject of this lesson.

The Record Is the Source of Truth

Every prior lesson in this program has circled the same fixed point: the output is evidence, and the evidence has to match the record. This lesson is about the mechanism that makes that match possible. It is called grounding. Grounding means tying the model's answer to a specific, attached, verifiable source so that what the model reports is what the record contains, not what a typical incident of that kind contains.

The distinction sounds academic until you watch it fail. A large language model is, at its core, a prediction engine trained on an enormous body of text. When you ask it a question without giving it source material, it answers from that training: it generates the most statistically likely account of the kind of situation you described. For most uses, that is exactly what you want. For a sworn account that a defense attorney will cross-examine, it is a trap. The model does not know your suspect. It knows suspects in general. It does not know what your victim said. It knows what victims tend to say. Ask it to summarize an interview it cannot see, and it will hand you a summary of an imagined interview that reads like the real one.

The three pillars of the record in public safety are consistent across nearly every case. The body-worn camera (BWC) footage, the camera clipped to an officer's chest that records audio and video through a contact, captures what was said and what was visible. The computer-aided dispatch (CAD) entry, the system that receives and routes calls for service and logs the incident in real time, is authoritative for objective facts: received time, dispatch time, arrival time, the address, the call type. The case file in the records management system (RMS), the agency's central repository for offense reports, arrest records, statements, and evidence logs, holds the assembled body of the investigation. These three are the source of truth. The grounding discipline is the practice of making the model answer from them and only from them.

Grounding is the difference between a model that reports what your record contains and a model that reports what a typical record contains. Only one of those is evidence.

Why the Model Fills the Gap

To ground well you have to understand the behavior you are constraining. When a model encounters a hole in the material it has been given, its default is to fill the hole. It was built to produce coherent, complete text, and a coherent narrative has no gaps. So if the interview transcript is silent on whether the suspect owned a vehicle, and you ask a question that touches on transportation, an ungrounded model will not say "the transcript does not address this." It will generate a plausible answer, because in its training data, accounts of suspects usually mention how they got somewhere. The gap-fill is not a malfunction. It is the model doing exactly what it was designed to do, applied to a context where the design is dangerous.

This is why grounding is not a single instruction but a system. You feed the model the record. You constrain it to the record. You verify that its output traces back to the record. Each step closes a door the model would otherwise walk through. Skip any one of them and the gap-fill returns, wearing the confident, professional prose that makes it so hard to catch.

Feeding the Record: Attachment and Retrieval

You cannot ground a model on material it does not have. The first move is always to get the actual source documents into the model's working context. There are two ways this happens, and an L3 practitioner should understand both because they have different failure modes.

Direct Attachment

Direct attachment is what Detective Reyes did: he placed the interview transcript and the victim statement directly into the conversation so the model reads the full text. This is the cleanest form of grounding because nothing stands between the model and the source. The model sees every word. For a single interview, a single statement, a single BWC transcript, direct attachment is the right approach and you should default to it.

The discipline here is to attach the rawest available version of the record. Attach the BWC transcript, not your summary of the BWC transcript. The moment you summarize before attaching, you have filtered the material through your own memory and judgment, and every gap in your summary becomes a gap the model will fill. Give it the raw transcript and let the model do the synthesis against the complete text. Your job is verification, not pre-digestion. The same applies to the CAD entry: export it and attach the actual log, with its timestamps intact, rather than retyping the times you remember.

Retrieval Over Large Files

Reyes had a 340-page case file. No officer is going to paste 340 pages into a chat box, and many tools cannot hold that much text in a single context anyway. This is where retrieval comes in. Retrieval is a technique where the system stores the full case file, and when you ask a question, it pulls the specific passages relevant to your question and feeds those passages to the model along with your prompt. The model then answers from the retrieved passages rather than from its training. Agency AI platforms increasingly build this in: you load the case file once, and every question you ask is answered against the file.

Retrieval is powerful and it has a specific failure mode you must respect. The model only sees what the retrieval step pulled. If the relevant passage was on page 211 and the retrieval step missed it, the model answers as though page 211 does not exist, and it answers confidently. Retrieval can produce an answer that is perfectly faithful to the passages it received and still wrong, because the decisive passage was never retrieved. The practitioner's defense is to treat a retrieval answer as a lead, not a conclusion, and to verify the answer against the cited passages in the actual file before relying on it. When the stakes are a charging decision or a sworn statement, you read the source pages yourself.

Attachment shows the model everything. Retrieval shows the model what it found. With retrieval, the question is not only "is the answer right," it is "did the system even look at the page that matters."

What Belongs in the Context and What Does Not

Grounding is also about restraint. Loading more material is not always better. If you feed the model unrelated case files, prior reports about the same suspect, or your own theory of the case, you give it more raw material to blend, and blending is the enemy of grounding. A clean grounded prompt contains the specific record relevant to the specific question and nothing that invites the model to reach beyond it. There is also a Criminal Justice Information Services (CJIS) Security Policy dimension here, the FBI-maintained policy that governs how criminal justice information is collected, processed, and stored. What you load into an AI tool is governed by that policy and by your agency's data agreements, and the accountability for compliant handling stays with the agency, never the vendor. Ground on the record you are authorized to use, in a tool you are authorized to use it in.

Constraining the Model to the Record

Feeding the model the record is necessary but not sufficient. A model can have the full transcript in front of it and still reach past it, blending the attached text with its training patterns unless you explicitly forbid it. The constraint is the instruction that tells the model the attached material is a boundary it must not cross. This is where the prompting discipline from earlier lessons becomes a grounding discipline.

The Grounding Instruction

The core constraint is a sentence you attach to every evidentiary prompt. A workable formulation: "Answer using only the attached interview transcript and victim statement. Do not add, infer, or estimate any fact that is not stated in those documents. If the documents do not address something I ask about, say 'not addressed in the provided record' rather than answering from general knowledge." That instruction does real work. It shifts the model's reference point. Without it, the model treats your attached file as a starting point and fills outward from training. With it, the model treats the attached file as the authoritative boundary and is far more likely to flag a gap than to paper over it.

The constraint does not make the model infallible. It reduces gap-fills; it does not eliminate them. A model under a tight grounding constraint can still misread a timestamp, misattribute a statement to the wrong speaker in a transcript with overlapping voices, or misconstrue an ambiguous phrase. Grounding makes the failure modes rarer and more catchable. It does not retire the verification pass. Nothing does.

Forcing the Citation

The single most useful grounding technique an L3 practitioner can adopt is to require the model to cite its source for every claim. Add to the prompt: "For each statement in your answer, cite the specific source and location it came from, such as the transcript line, the timestamp, or the page of the case file." This produces two compounding benefits. First, the output becomes dramatically faster to verify, because you are not hunting through 340 pages to check a claim; the model has told you exactly where to look. Second, and less obvious, the act of producing a citation forces the model to process its relationship to the source more carefully, and a model required to cite is measurably less likely to fabricate, because it is hard to produce a believable citation for a fact that is not in the document.

When a model under a citation requirement produces a claim with a citation that does not check out, you have caught a hallucination at the cheapest possible moment, before it ever reached your report. When it produces a claim and marks it "no source in provided record," it has done your gap-flagging for you. Citation requirements turn the model into a partner in its own verification.

The Self-Audit Pass

After the model produces a grounded, cited draft, a second prompt sharpens it further. Ask the model to audit its own output: "Review your answer. For each statement, mark it as (1) directly stated in the provided record, (2) an inference you drew from the record, or (3) an addition not traceable to the record. List anything in category 2 or 3 separately." The self-audit is imperfect; a model that hallucinated a fact may also misclassify it. But it is not worthless. It surfaces the model's own inferences and additions before you start your verification, handing you a prioritized list of the claims most likely to be problems. You still verify everything. The self-audit just tells you where to look first.

Verifying the Output Traces to the Record

Grounding's whole purpose is to make verification possible and fast, not to make it unnecessary. The final pillar is the discipline of confirming that every claim in the output actually traces back to the record. A well-grounded, well-cited output makes this the most efficient it can be, but it is still a step you perform deliberately, not a box you assume is checked.

Tracing Claim by Claim

Work through the output one claim at a time and, for each claim, follow the citation back to the source. Open the transcript to the cited line and read it. Open the case file to the cited page and read it. Confirm three things: that the cited source exists, that it actually says what the model says it says, and that the model has not subtly shifted its meaning. The third check catches the most insidious failure, the softened or sharpened fact. A transcript line where the suspect says "I might have been near there" can become, in a model's summary, "the suspect admitted being near the scene." Both reference the same line. Only one is accurate. Tracing the claim to the actual words is the only way to catch that drift.

For attributed statements and direct quotes, the verification is verbatim, not approximate. If the output puts words in the suspect's mouth, those words must match the transcript exactly. A model summarizing audio sometimes cleans up grammar: "I didn't do nothing" becomes "I didn't do anything." In an interview summary feeding a charging decision, that is not a tidy edit. It is a changed statement, and the changed version may not match what a jury will hear when the audio plays. The same model can also misattribute a statement entirely, assigning a witness's words to the suspect when overlapping voices blur the speaker labels. Verbatim, attributed verification is the only thing that catches it.

The CAD Anchor for Objective Facts

For objective facts, the CAD entry is the anchor, and it overrides the model every time. If the model's narrative, drawn from interview audio, places the arrival at 11:42 p.m., but the CAD log shows 11:51 p.m., the CAD entry wins. Those times were logged in real time by the telecommunicator, the dispatcher who answers and codes calls for service. A nine-minute discrepancy between your account and the authoritative timeline is exactly the kind of inconsistency a defense investigator finds and uses to question the credibility of everything else in the report. Grounding the model on the actual CAD export, and verifying the times against it, closes that gap before it opens.

The Omission Check

The hardest thing to verify is what is absent. A grounded model reports what is in the record, which means it can faithfully omit something significant that the record contains but the model did not surface, or something that happened outside the recorded material entirely. After you have traced every claim the model made, ask the inverse question: is there anything in the record that matters and that the output did not address? For a suspect interview, did the model surface the one exculpatory statement buried at minute seventy-one? Brady v. Maryland (373 U.S. 83) requires the prosecution to disclose material exculpatory evidence to the defense, and Giglio v. United States (405 U.S. 150) extends that to evidence that could impeach a witness, including the testifying officer. An exculpatory statement the model glossed over is not just an omission. It is a Brady problem waiting in the file, and the grounding workflow only protects you if your verification includes the deliberate hunt for what the summary left out.

A grounded model tells you what the record contains. Your verification confirms that, and then asks the question the model cannot: what does the record contain that the answer left out?

The Systematized Grounding Workflow

At L3, individual techniques become a repeatable workflow you run the same way every time, so that grounding is not something you remember to do on important cases but something built into how you work on every case. Here is the workflow Detective Reyes followed, assembled into a sequence.

Step one, assemble the record. Identify the specific sources that answer your question: the interview transcript, the victim statement, the CAD export, the relevant case-file pages. Confirm you are authorized to load them in the tool you are using, under agency policy and CJIS obligations.

Step two, feed the record raw. Attach the actual documents, not summaries. For material too large to attach, use retrieval and remember that the model will only see what retrieval pulls.

Step three, constrain to the record. Add the grounding instruction (answer only from the attached material, flag gaps rather than filling them) and the citation requirement (cite the source and location for every claim).

Step four, ask the self-audit. Have the model classify its own statements as stated, inferred, or added, and list the inferences and additions separately.

Step five, trace and verify. Follow every citation back to the source. Confirm existence, accuracy, and unchanged meaning. Verify quotes verbatim and check every attribution. Anchor objective facts to the CAD.

Step six, run the omission check. Ask what the record contains that the output missed, with specific attention to exculpatory and impeachment material.

Step seven, document the grounding. Note what you attached, what constraint you used, and what you verified. That note is your record of the process and your answer when someone asks how the document was produced.

Why the Discipline Pays for Itself

This looks like more steps than typing a question into a chat box, and it is. It is also far faster than the alternative it replaces. A detective summarizing a ninety-minute interview by hand, cross-referencing it against a victim statement, might spend an hour. A grounded, cited model produces the same summary in under a minute, and the verification pass on a well-cited output, where every claim points to its source line, runs a fraction of the manual time. Axon's Draft One, the report-drafting tool that generates narratives from BWC audio, showed an 82% decrease in report-writing time in testing, against the backdrop that officers spend roughly 30 to 40% of every shift on paperwork. Those savings are real, and they should be treated as a benchmark to verify in your own context, not a guarantee. The grounding workflow is what keeps the saved time from being borrowed against accuracy. The detective who grounds is not slower than the one who does not. He is faster, and his work survives discovery.

Grounding as the Answer to the Objection

The grounding discipline is also the most direct answer to the legitimate objections this field faces. When the King County, Washington prosecutor's office barred AI-written police reports from its charging process absent specific documentation of the review, the concern was not that AI is inherently untrustworthy. It was that an ungrounded, unverified AI document is a sworn account no one can vouch for. A grounded, cited, verified, and documented output meets that objection on its own terms: every claim traces to the record, and there is a documented process showing it does.

The Electronic Frontier Foundation (EFF), the digital-rights organization that has raised transparency concerns about AI police reports, makes a related point. The public, and especially a criminal defendant, has a legitimate interest in knowing whether a government document that can deprive a person of liberty was generated by a machine pattern-matching from training data or grounded in the actual evidence of the case. That concern is legitimate and deserves a straight answer, not a dismissal. The grounding workflow lets you give one. The document was generated by a model constrained to the actual record, every claim was traced back to a source in that record, and here is the documentation of how. That is the concern taken seriously and answered.

The officer who has internalized grounding can sit across from a chief, a prosecutor, and an oversight board and tell all three the same true thing. The AI was fed the footage, the CAD, and the case file. It was constrained to report only what those contain. Every claim was traced back to its source before the document was adopted. And it is documented. That is the systematized practice this lesson asks you to build, and it is what makes AI a tool that earns its place in evidence rather than a liability that contaminates it. Detective Reyes scrolled up and attached the files before he asked his question. That two-second instinct is the whole discipline in miniature: the record first, the question second, and the verification before anyone relies on the answer.

Key Takeaways

• Grounding ties the model's output to specific attached sources so it reports what your record contains, not what a typical incident contains. An ungrounded model answers from training patterns and fills every gap with a plausible invention.
• The record is the BWC footage, the CAD entry, and the case file. Feed the model the rawest available version; the moment you summarize before attaching, your gaps become the model's gaps.
• Direct attachment shows the model the full text and is the default for single documents. Retrieval handles large case files but only shows the model what it pulled, so a faithful answer can still be wrong if the decisive passage was never retrieved.
• The grounding constraint ("answer only from the attached material, flag gaps rather than filling them") shifts the model from treating your file as a starting point to treating it as a boundary. It reduces gap-fills but never eliminates them, so verification stays mandatory.
• Requiring a source citation for every claim makes the output faster to verify and makes the model measurably less likely to fabricate, because a citation for an invented fact is hard to produce convincingly.
• Verification means tracing every claim to its cited source and confirming the source exists, says what the model says, and has not had its meaning softened or sharpened. Quotes are verified verbatim, attributions are checked, and objective facts like times are anchored to the CAD, which overrides the model.
• The omission check is the inverse question the model cannot ask itself: what does the record contain that the output left out? A glossed-over exculpatory statement is a Brady and Giglio problem the moment it stays buried.
• The seven-step grounding workflow (assemble, feed raw, constrain, self-audit, trace, omission check, document) is the systematized answer to the King County bar and the EFF transparency concern: every claim traces to the record, with documentation to prove it.