Detective Maria Pena spent three hours the night before the preliminary hearing re-reading every AI-assisted report in the case file. She had no standardized checklist. She was working from memory and instinct, comparing sentences to her mental image of the footage, hoping she had not missed anything. She had not. But the defense attorney noticed something she did not: a timestamp in the second report placed a key event 11 minutes earlier than the body-worn camera (BWC, the officer-mounted video recorder) showed it occurring. The AI had not flagged the discrepancy. Detective Pena's review, honest and thorough as it was, had not caught it either. Not because she was careless, but because the verification method she used was a mood, not a process.

Why Verification Must Be a Process, Not a Mood

There is a significant difference between "I reviewed the report" and "I ran a documented verification pass against this checklist." The first statement is a claim about effort. The second is a claim about a specific, repeatable process applied in a specific way. When a defense attorney challenges an AI-assisted report, the question that matters is not whether the officer believes they reviewed it carefully. The question is whether the review was conducted in a way that any qualified person could replicate, producing the same result from the same source materials.

A verification checklist converts the review process from a personal judgment into a documented, reproducible procedure. When every reviewer follows the same checklist against the same source materials, two things become true. First, the review is more likely to catch errors, because each step forces attention to a specific category of potential failure. Second, the review is auditable: the completed checklist shows exactly which checks were performed, who performed them, when, against which source materials, and what the result was. That audit trail is the difference between a reviewer who says "I looked at it" and a reviewer who says "I ran step 7, which requires verifying every timestamp in the narrative against the BWC footage timeline, and here is the result."

The court standard is the right target for verification, even for reports that will never go to trial. Not because every incident becomes a case, but because the verification standard that satisfies a defense attorney in cross-examination is the same standard that produces accurate, defensible records for all purposes: internal review, civil litigation, oversight investigation, and public-records requests. Building toward the court standard once produces a process that is robust enough for all uses.

Verification is not the act of reading the report. It is the act of checking each claim in the report against the specific source material that supports it, and documenting the result of each check.

The Structure of a Court-Ready Verification Checklist

A court-ready verification checklist for an AI-assisted report has five sections, each targeting a specific category of failure that AI-drafted documents produce. The sections are not guidelines to apply generally: they are specific checks with specific source-material citations and pass or fail outcomes for each item.

Section 1: Timestamp and sequence verification. Every time reference in the report, whether absolute (14:32:07) or relative ("upon arrival," "approximately five minutes later"), must be traced to a specific source. For absolute timestamps, the source is the BWC footage timestamp, the CAD (computer-aided dispatch, the platform that logs every call and timestamps every action) timestamp, or the dispatch audio recording. For relative time references, the source is the elapsed time visible in the footage or the CAD entry sequence. Each timestamp check produces one of three outcomes: confirmed (the narrative matches the source within the acceptable tolerance defined by agency policy, typically 30 to 60 seconds for narrative references), discrepant (the narrative states a different time than the source), or not sourced (no source material supports the timestamp and the report must be corrected or the unsupported reference flagged). A report with any discrepant or not-sourced timestamps should not be adopted until the discrepancy is resolved or the unsupported reference is removed and corrected.

Section 2: Quoted speech verification. Every quoted statement attributed to any person in the report, whether subject, witness, or officer, must be verified verbatim against the source audio. This is the highest-stakes verification check. An invented quote or a materially altered quote is a Brady (Brady v. Maryland, the 1963 ruling requiring disclosure of exculpatory evidence) violation if the accurate version of that statement is exculpatory. It is a Giglio (Giglio v. United States, the 1972 ruling requiring disclosure of evidence affecting witness credibility) problem if the inaccurate quote affects the witness's or officer's credibility. Each quoted statement check requires locating the specific point in the audio where the statement occurred, confirming the words in the report match the words in the audio, and noting any material differences. If the statement was not recorded, the checklist records that fact and flags the quote for officer review and correction.

Section 3: Factual claim verification. Every factual claim in the report that is not a quoted statement must be verified against the source materials. This includes physical descriptions of people, vehicles, and objects; descriptions of observed actions and behaviors; location and geography references; and any other statement of fact that could be tested against the BWC footage, the CAD entry, or documentary evidence. The standard is traceability: each claim must be traceable to a specific point in the footage, a specific CAD field, or a specific document in the file. Claims that cannot be traced are either gap-fill inferences the AI added without support, or facts the officer knows to be true from personal observation but that need to be explicitly attributed to that source. The checklist distinguishes between the two: "confirmed by footage at [timestamp]," "confirmed by CAD entry field [field name]," or "officer personal observation, not captured on footage, attributed to officer."

Section 4: Legal characterization review. The checklist's fourth section addresses a failure mode specific to AI output in law-enforcement contexts: the AI's tendency to include legal characterizations or conclusions that require officer judgment and that may be inaccurate or prejudicial. Phrases like "the subject was clearly attempting to flee," "the circumstances were consistent with drug trafficking," or "the suspect's behavior indicated probable cause" are not factual descriptions: they are legal or analytical characterizations that belong in the officer's own assessment, not in the AI draft. The checklist requires a review of every sentence for legal or analytical language that the officer did not place there. Any such language must be either removed, rewritten as a factual observation, or explicitly adopted by the officer as their own assessment with the supporting facts also stated. This section protects the officer from having a characterization they did not make attributed to them in a sworn report.

Section 5: Completeness and disclosure check. The final section verifies that the report is complete and properly documented. The completeness check covers required fields in the RMS (records management system, the agency's central case file repository): all subject fields populated with verified data, all evidence items accounted for, all required case numbers and cross-references present. The disclosure check confirms that the AI assistance notation is appended in the required format, that the tool name and review date are recorded, and that the officer's adoption statement is present. This section also includes a check for over-disclosure: confirming that the report does not contain protected information, such as juvenile identifying information or confidential source details, that should not be in the narrative.

Making the Checklist Reusable and Defensible

A checklist that exists in a personal notebook is not a system. A checklist that is standardized across the agency, version-controlled, and attached to every verified report as a completed record is a system. The difference matters legally and operationally.

The reusable checklist has three properties that distinguish it from a personal verification habit. First, it is standardized: every officer in the agency uses the same checklist for the same type of report, so the verification that happened in a 2 a.m. patrol report and the verification that happened in a detective's case summary can be compared against the same standard. Second, it is attached to the report: the completed checklist travels with the report through the case file and the discovery package, not as a separate document to be provided on request, but as a built-in exhibit showing exactly what was checked, against which source, and with what result. Third, it is version-controlled: the checklist has a version number and effective date, so when a defense attorney asks which version was used for a specific report, the answer is documented and specific.

Defensibility has a practical dimension that goes beyond content. A completed checklist in the format the lesson describes is what allows the officer to answer the following deposition question clearly: "Officer, please walk me through the specific steps you took to verify that the AI-generated draft of this report was accurate before you adopted it as your sworn account." The answer, with a completed checklist in hand, is not a general claim about care and thoroughness. It is a specific account: "I ran the five-section verification checklist, version 3.2, on this date, against these source materials. Section 1 confirmed all twelve timestamps within the 30-second tolerance. Section 2 confirmed four quoted statements verbatim and identified one discrepancy, which I corrected before adoption. Section 3 confirmed all factual claims except two, which I also corrected. Sections 4 and 5 found no issues." That is what a court-standard verification looks like in practice.

The checklist also protects supervisors. When a supervisor signs off on a report that was produced with AI assistance, they are attesting that the verification was adequate. A completed checklist tells the supervisor specifically what was checked and what was found, rather than requiring the supervisor to trust that the officer's general review caught everything. The supervisor who approves a report with a completed checklist attached is making a documented, specific judgment about a documented, specific verification process. The supervisor who approves a report without one is trusting a process they cannot see.

Calibrating the Checklist to Incident Type and Stakes

Not every incident requires the same checklist depth. The resource commitment of a full five-section court-standard verification is appropriate for reports that are likely to be challenged: use-of-force incidents, in-custody events, officer-involved shootings, and any case that will be prosecuted. For routine incident reports on matters unlikely to result in prosecution or civil litigation, a streamlined version of the checklist may be appropriate, covering the highest-risk failure modes (timestamps, quoted speech, and legal characterizations) without the full factual-claim audit for every sentence.

Agencies should define at least two checklist tiers in policy: a standard tier for routine AI-assisted reports and a full tier for high-stakes incident types. The tier designation should be made by the supervisor at the time of review, not left to officer discretion. An officer who defaults to the standard tier for a use-of-force report because they are tired at the end of a shift has made a risk decision the agency may not support. Supervisor-assigned tiering ensures the stakes are considered by someone with oversight authority, not just the individual completing the report.

Within each tier, the checklist should specify the tolerance for each check. Timestamp tolerance: how many seconds of discrepancy between the narrative and the footage constitutes a discrepancy requiring correction? Quote accuracy: does a paraphrase that preserves the substance of a statement pass, or does verbatim accuracy require matching the exact words? These tolerances should be consistent across the agency, established by policy and legal counsel review, and documented in the checklist header so that every reviewer applies the same standard.

The tolerance question is not merely technical. A timestamp discrepancy of 4 seconds is probably measurement imprecision. A discrepancy of 11 minutes, like the one that surfaced in the opening story, is a substantive factual error that affects the reconstruction of the incident timeline. A quote that renders "I think I might have" as "I said I would" crosses from paraphrase into material alteration. The checklist's job is to make these distinctions explicit and consistent, so that different reviewers applying the same checklist to the same report would reach the same conclusions about which discrepancies require correction and which are within tolerance.

Integrating the Checklist Into the Agency Workflow

A checklist that officers complete after filing the report is less useful than a checklist that is integrated into the review and adoption workflow. The ideal integration point is before the report is submitted to the RMS: the officer completes the checklist, resolves any issues it surfaces, and then adopts the corrected draft. The completed checklist is attached to the report in the RMS as a supporting document, not stored separately.

The agency's AI-use policy should specify that no AI-assisted report may be adopted without a completed verification checklist. This policy creates an enforcement mechanism without requiring supervisors to audit every report: the RMS itself can require the checklist to be attached before the report is filed. In agencies where the RMS has this capability, the requirement becomes part of the workflow rather than a manual compliance obligation.

Training on the checklist should be scenario-based. Officers should practice completing the checklist against a report that contains one or more planted errors: a timestamp discrepancy, an invented quote, and a legal characterization the officer did not provide. The training exercise should be scored: did the officer catch all three errors? Did they document the catches in the checklist format? Did they resolve the errors before adoption? Training against planted errors builds the pattern-recognition habit that makes the checklist effective in the field, rather than a form-filling exercise that officers complete without engaging with its purpose.

The checklist should be reviewed annually, or any time a new AI tool is deployed or a significant failure is identified. As AI tools improve and their specific failure modes evolve, the checklist should evolve with them. The timestamp discrepancy may become less common as AI tools improve their integration with BWC metadata. New failure modes may emerge that the current checklist does not cover. Annual review by legal counsel, supervisors, and line officers who use the tool is the mechanism for keeping the checklist current. The version number changes with each substantive revision, and the effective date is recorded, so there is always a documented answer to "which checklist was in use on this date?"

Key Takeaways

• Verification is not the act of reading a report. It is the act of checking each specific claim against the specific source material that supports it and documenting the result. "I reviewed it" is a claim about effort; a completed checklist is evidence of a process.
• A court-ready verification checklist has five sections: timestamp and sequence verification, quoted speech verification, factual claim verification, legal characterization review, and completeness and disclosure check. Each section produces specific, documented outcomes rather than a general impression.
• Quoted speech verification is the highest-stakes check. An invented or materially altered quote attributed to any person in the report is a Brady problem if the accurate version is exculpatory, and a Giglio problem if the inaccuracy affects the witness's or officer's credibility.
• A reusable, defensible checklist has three properties: it is standardized across the agency, it is attached to every verified report as a completed record, and it is version-controlled so the specific version in use for any report can be documented and produced on demand.
• Agencies should define at least two checklist tiers in policy: a standard tier for routine reports and a full tier for high-stakes incidents. Tier assignment should be made by supervisors, not left to officer discretion, because an officer at the end of a demanding shift may under-apply the stakes assessment.
• Checklist tolerances (how many seconds of timestamp discrepancy constitute an error, whether verbatim or paraphrase accuracy is required for quotes) must be defined consistently by agency policy and legal counsel review, not left to individual reviewer judgment on each report.
• Training on the checklist should be scenario-based with planted errors, scored for detection and documentation, not just procedural compliance. Officers who understand why each checklist step exists are more effective at applying it than officers who treat it as a form-filling requirement.
• The checklist should be integrated into the pre-submission workflow so that no AI-assisted report can be adopted without a completed checklist attached. Annual review and version control ensure the checklist evolves as AI tools and their failure modes evolve.