The AI-native forensic psychologist: a field guide to collateral records, chronology and AI

The evaluee in front of you has a reason to shade the story. A defendant facing a competency hearing may want to look sicker or better than he is. A plaintiff claiming emotional damages has a lawsuit riding on the answer to "how were you before the crash". A parent in a custody evaluation is describing the other parent. None of that makes them liars. It makes self-report 1 source, and the profession's own rules say 1 source is not enough.

"Forensic practitioners ordinarily avoid relying solely on one source of data, and corroborate important data whenever feasible. When relying upon data that have not been corroborated, forensic practitioners seek to make known the uncorroborated status of the data, any associated strengths and limitations, and the reasons for relying upon the data."

The AAPL guideline for psychiatrists says the same thing from the other side of the hall: collateral sources "are usually an important element of the forensic assessment", the expert should personally review the records rather than rely on an attorney's summary, and collateral information "may be compared with the evaluee's account to help detect malingering and assess reliability". It also asks for something most reports skip: a list of everything requested but not received.

Then there is the courtroom. Since the 2023 amendment, Federal Rule of Evidence 702 requires the proponent of an expert opinion to show it is "more likely than not" that the testimony "is based on sufficient facts or data" and "reflects a reliable application of the principles and methods to the facts of the case". A psychologist who opined on competency without reading the jail mental health chart has a facts-or-data problem before the first cross-examination question. State courts that follow Daubert or their own version of Frye ask the same question in different words.

So the field guide version of the rule: every important fact in a forensic report should trace to a page, every source should be listed, and every source you asked for and did not get should be listed too. The rest of this guide is the procedure for doing that on a 2,000-page file without spending a week on it.

A records request that does not start from the legal question produces a box of paper. The question tells you which custodian holds the weight-bearing facts and which time window to request. The table below is the map this guide uses throughout.

Competency has a clock, and the clock is in the records

In federal court, a defendant committed for a competency examination can be held "not to exceed thirty days", with 1 extension "not to exceed fifteen days" (18 USC 4247(b)). If found incompetent, the first restoration commitment runs "a reasonable period of time, not to exceed four months" (18 USC 4241(d)). Jackson v. Indiana (1972) set the constitutional outer bound: a defendant held only because he cannot stand trial can be held no longer than "the reasonable period of time necessary to determine whether there is a substantial probability that he will attain that capacity in the foreseeable future". States have their own clocks. All of them run from dates that live in the court file and the hospital admission record, and the evaluator who cannot find those dates cannot answer the restoration question.

The federal report statute is also a checklist. Under 18 USC 4247(c), a report must include "the person's history and present symptoms", "a description of the psychiatric, psychological, and medical tests that were employed and their results", "the examiner's findings", and the examiner's opinions on diagnosis and prognosis plus the specific legal question. The first of those 4 is where the collateral records go, and the statute expects history, not a sentence.

Psychiatric records are narrative. Unlike an orthopedic file, there is no lab value that settles the question; there are progress notes, screening forms, discharge summaries and a sick call slip in pencil. Each custodian produces a different kind of document, and each has a characteristic way of misleading a reader who is in a hurry.

Jail medical and mental health

Intake screening (often a checklist completed by a booking officer, not a clinician), sick call slips, nurse and psychiatric provider notes, the medication administration record (MAR), suicide watch and segregation logs, refusal forms. The trap: intake screens record what the defendant said on the worst night of his life. "Denies psychiatric history" at 2 a.m. is a data point, not a finding. The MAR is where you learn whether the medication was actually taken.

State hospital

Admission evaluation, treatment plans, competency restoration group attendance, nursing notes, incident reports, the forensic report to the court, discharge summary. The trap: copy-forward. A treatment plan reviewed every 30 days can carry a sentence for a year after it stopped being true. Read the dates on the content as well as the date on the page.

Prior psychological testing

Earlier reports, raw test data, protocols, scoring printouts. The trap: a report that names a test and an interpretation without the data. You cannot judge a prior examiner's reading of a validity scale without the scores; ask for test data, and expect a test-security conversation about test materials.

School records

Cumulative file, attendance, discipline, IEPs and 504 plans, psychoeducational evaluations, nurse visits. The trap: the diagnosis label in an IEP is an educational classification, not a clinical diagnosis. "Emotional disturbance" is a category under IDEA; it tells you services were given, not what the child had.

Child protective services

Intake reports, investigation narratives, findings (substantiated, unsubstantiated, unfounded, with meanings that differ by state), safety plans, case closures. The trap: an unsubstantiated finding is not a finding that nothing happened. It is a finding that the agency did not meet its own evidentiary threshold. Report it as what it is.

Police and court

Incident reports, arrest reports, CAD or 911 logs, booking records, body-camera summaries, use-of-force reports, crisis intervention team notes, protective orders, the docket. The trap: the narrative is written by the person who made the arrest. Behavior descriptions ("rambling", "paranoid", "did not respond to commands") are useful; diagnostic words from an officer are not.

Pharmacy and PDMP

Fill histories from pharmacies and, where you can lawfully obtain it, the state prescription drug monitoring program. The trap: a fill is not a dose taken, and a gap in fills can be a hospitalization, a lapse, or a switch to a 90-day supply. Match gaps against the hospital dates before you call them non-adherence.

Prior IMEs and evaluations

Earlier competency or sanity reports, Social Security consultative examinations, workers' compensation psychiatric IMEs, VA compensation and pension examinations, disability insurer reviews. The trap: the prior examiner's opinion is not data about the evaluee; the history and test results inside it are. Separate them when you cite.

Which records to chase first

Every custodian takes weeks and some take a court order. If the evaluation date is fixed, order by weight and by delay, and say so in the report if something arrives after you wrote it.

Usually retained counsel or the court does the requesting, and you tell them what to ask for. Even so, the evaluator who knows the rules gets better productions, because the request names the right instrument for each custodian. 4 regimes come up on nearly every file.

What the HIPAA pieces actually say

Under 45 CFR 164.508(a)(2), "a covered entity must obtain an authorization for any use or disclosure of psychotherapy notes", with narrow exceptions. Psychotherapy notes in the HIPAA sense are the therapist's separately kept process notes; the ordinary chart (diagnoses, medications, session dates, treatment plan, test results) is not psychotherapy notes and comes with the regular authorization. In practice, most forensic evaluations get what they need from the ordinary chart, and the separate authorization is worth requesting only when the process notes are the point.

Under 164.512(e)(1)(i), a covered entity may release records "in response to an order of a court or administrative tribunal, provided that the covered entity discloses only the protected health information expressly authorized by such order". So the order should name the record types and the date range. A vague order produces a vague production, and a jail health contractor reading a vague order will send the least it can.

COLLATERAL RECORDS REQUEST
Matter: [CASE CAPTION AND NUMBER]      Evaluee: [NAME, DOB]
Referral question: [COMPETENCY / RESPONSIBILITY / DAMAGES / FITNESS / RISK / CUSTODY]
Evaluation date: [DATE]     Records needed by: [DATE, 14 DAYS BEFORE]
Instrument: [AUTHORIZATION / COURT ORDER / SUBPOENA], copy attached

#  Custodian                 Record types                                   Date range        Instrument   Received  Pages  Gaps
1  [County jail / contractor] Intake screening, sick call, MH provider notes, [ARREST DATE   HIPAA auth    [ ]      [ ]    [ ]
                             MAR, suicide watch logs, refusals, transfers    to present]     or order
2  [State hospital]           Admission eval, treatment plans, restoration   [ADMIT to       Order         [ ]      [ ]    [ ]
                             notes, nursing notes, incidents, discharge,     DISCHARGE]
                             forensic report to court
3  [Prior evaluator(s)]       Reports, test data, protocols, raw scores      [DATES]         Order or      [ ]      [ ]    [ ]
                                                                                             auth
4  [Outpatient MH / PCP]      Full chart: notes, problem list, med list,     [5 YEARS        HIPAA auth    [ ]      [ ]    [ ]
                             labs, referrals (not psychotherapy notes        before event
                             unless separately authorized)                  to present]
5  [Pharmacy / PDMP]          Fill history, all pharmacies                   [5 YEARS]       Auth; PDMP    [ ]      [ ]    [ ]
                                                                                             per state law
6  [SUD program]              Admission, treatment, discharge                [DATES]         Part 2 consent [ ]     [ ]    [ ]
                                                                                             or 2.64 order
7  [School district]          Cumulative file, attendance, discipline, IEP,  [GRADES K-12]   FERPA consent [ ]      [ ]    [ ]
                             504, psychoeducational evaluations                              or order
8  [CPS agency]               Intakes, investigations, findings, case plans  [LIFETIME]      Court order   [ ]      [ ]    [ ]
9  [Police / sheriff]         Incident and arrest reports, CAD/911 logs,     [EVENT DATE     Discovery     [ ]      [ ]    [ ]
                             booking, body-cam summary, CIT, use of force    plus prior calls]
10 [Court]                    Docket, prior competency orders and findings,  [ALL]           Public / counsel [ ]   [ ]    [ ]
                             protective orders
11 [Employer] (fitness/damages) Referral letter, incident report, discipline, [2 YEARS]      Employer      [ ]      [ ]    [ ]
                             attendance, prior fitness evaluations                           authorization
12 [Insurer / SSA / VA / WC]  Prior IMEs, consultative exams, C and P exams, [ALL]           Auth or order [ ]      [ ]    [ ]
                             claim decisions

Requested but not received as of [DATE]: [LIST BY NUMBER]. These limits are stated in the report.

Before anyone reads a page, someone should know what pages exist. The index is a 1-page table of every document in the file: who produced it, what it is, the date range it covers, how many pages, and whether it reconciles with what was requested. It is the first thing an opposing expert will ask for on cross ("Doctor, what records did you review?") and the first thing a good retaining attorney will want before the evaluation. The published collateral records index sample shows the finished form on a fictional competency matter with 22 documents grouped by court, evaluation, custody and interview source.

Reconciling the index against the request

1. Match every request line to a production. Each row in the request list from chapter 4 should have a received date and a page count, or a "not received" entry with the date you last asked.
2. Check the inside dates. Open each document to its first and last dated entry. Write those on the index. Cover letters lie by omission.
3. Find the duplicates. Productions overlap: the state hospital summary appears again inside the jail chart, the prior competency report quotes the 2024 discharge. Duplicates inflate page counts and hide the single page that differs. Collapse them and note the collapse.
4. Find the wrong-patient and co-mingled pages. Family files, jail scanners and shared custodians produce pages that belong to someone else. 1 wrong-patient page cited in a report is a cross-examination that lasts an hour.
5. Freeze the numbering. Once the index is set, do not renumber. If a supplemental production arrives, add it as a new letter group (E, F) and compare it against the file rather than merging it.

A psychiatric chronology is a dated list of what the records show, each line cited to its page, with the source named so the reader knows whether it came from a clinician, an officer or a form. It is not a narrative and it contains no opinion. The published forensic case chronology sample merges a court file, jail records and state hospital notes into 1 sequence for a fictional competency matter; the psychiatric and medical records summary sample shows the history summary that sits beside it. The hypothetical below walks a competency referral from arrest to the evaluation interview.

The 2 documents that disagree

Inconsistencies between sources are the most useful output of a chronology and the easiest to lose in a 2,000-page file. Show them side by side, cited, and let the evaluator decide what they mean. The pair below is from the hypothetical above.

What the pair shows is not that the defendant lied. A disorganized man at 3 a.m. answering a booking officer's checklist is a known source of false negatives. What it shows is that the intake screen cannot carry weight on history, and that the question for the evaluator's own symptom-validity work is why the same denials reappeared, in a calm interview, 10 weeks later.

The medication history table

Every psychiatric chronology should produce a medication table as a by-product. Dose, start, stop, source, and whether it was prescribed, filled or administered, which are 3 different facts from 3 different custodians.

The table is also where copy-forward shows itself. If the state hospital treatment plan lists lithium in 2024 but the pharmacy shows no fills after 2022 and the discharge summary does not list it, the plan is carrying a dead line. Cite both pages and let the reader see it.

Evaluators underbid the records. A retainer that budgets 6 hours for "record review" on a 2,000-page competency file either produces a skimmed review or a bill the attorney did not expect. The ledger below is a planning illustration for 1 hypothetical file; your own numbers depend on the records, the scanner and how often the jail changed health contractors.

The point of the ledger is where the time sits. Reading and transcription are 5 out of every 6 hours. Judgment, the cross-checking and the writing, is the rest. An AI tool that drafts the index, the chronology and the medication table moves the reading hours into verification hours, which are faster because you are checking a cited line against a page rather than hunting for the line.

2 cautions on the right-hand column. First, verification is not optional and it is not fast if you skip the index: you need stable page addresses to check against. Second, the hours you save are reading hours, not thinking hours. If the AI-native evaluator finishes in 15 hours instead of 41, the 15 should contain more cross-checking than the 41 did, because the inconsistencies are now in a list instead of in your memory.

The workflow below assumes a tool that returns page-level citations. Without those, the steps collapse into "ask a chatbot and hope", which chapter 10 explains is not a method you can defend. Each step names the output and what you do with it.

What each output looks like

1. Index by source. The chapter 5 table, generated: custodian, document, inside date range, pages, duplicates removed. You add the "requested, not received" rows yourself, because the tool only knows what you uploaded.
2. Cited chronology by source. Each entry: date, source type (clinician note, officer narrative, screening form, fill record), the fact in the record's own words where possible, and the page. Sorted by date, filterable by custodian so you can read "jail only" or "everything from 2022".
3. Medication history table. Drug, dose, first and last seen, evidence type (prescribed, filled, administered, refused), page. Gaps shown as gaps, not filled in.
4. Prior diagnoses and testing list. Every diagnosis as written, by whom, when, on what page; every named test with its date, examiner and whether scores are in the file. Educational classifications kept separate from clinical diagnoses.
5. Inconsistencies, cited to pages. Statements that conflict across sources (history denied at intake, documented at discharge; "never hospitalized" against 3 admissions). These are signals for your symptom-validity work, which you do with your own instruments and judgment. A tool that labels an evaluee a malingerer has left its lane.
6. Missing-records list. Providers named in the file with no chart produced, date gaps in continuous records like the MAR, tests referenced with no data. Each flag cited to the page that implies the gap.
7. Questions over the record. "When was lithium last filled?" "Which pages mention a head injury?" Answered with citations, so a cross-examination question at 4 p.m. gets a page number, not a guess.

What the workflow never does

"I verified the AI output" is a sentence you will be asked to unpack under oath. Have an answer that is a method. The tiers below sort every line of the drafted chronology by what you are going to do with it, and set a verification rule for each. Write the rule into the report's procedures section, the same way you describe test administration.

3 rules sit on top of the tiers. First, anything that came from a low-confidence OCR page, a handwritten entry or a fax goes to tier 1 regardless of use; those are where the transcription errors live. Second, anything in the inconsistencies list is tier 1 by definition, because you are about to build symptom-validity reasoning on it. Third, the index, the missing-records list and the verification worksheet are all "data considered" under Guideline 10.06, which asks for documentation detailed enough "to allow for reasonable judicial scrutiny and adequate discovery by all parties". Keep them.

Disclosing the method

Guideline 11.03 encourages forensic practitioners "to identify the source of each piece of information that was considered and relied upon in formulating a particular conclusion". A page-cited chronology satisfies that line by design. Guideline 2.05 adds that opinions based on "novel or emerging principles and methods" should come with their status and limitations made known. The honest sentence for a procedures section reads something like: "Records were indexed and a draft chronology with page citations was prepared with [tool], a document review tool. I verified every cited fact relied on in this report against the source page and reviewed the remaining records for relevance. The tool did not administer, score or interpret any test and offered no opinion." Adapt it, but do not leave it out; an undisclosed method is worse on cross than a disclosed one.

• Every tier 1 line checked against its page by meand the check is initialed on the worksheet
• Every handwritten, faxed or low-confidence page read by me in fullOCR errors concentrate here
• Every inconsistency in the list traced to both pagesbefore it enters the symptom-validity analysis
• Missing-records list reconciled with counseland the unresolved items stated in the report as limitations
• Index, worksheet and draft tables retainedas data considered, available in discovery
• Tool use disclosed in the procedures sectionwith what it did and did not do
• No quoted record text in the report that I did not read on the pagethe tool's transcription is not the record
• Wrong-patient and co-mingled flags resolvedand the quarantined pages listed on the index

0 of 8 checked

The tools in this category are large language models (LLMs) wrapped in a document pipeline: OCR and handwriting recognition turn scanned pages into text, clinical NLP finds dates, drugs, diagnoses and providers, and retrieval-augmented generation (RAG) makes the model answer from the uploaded pages and cite them rather than from whatever it absorbed in training. That architecture is what makes AI medical record review usable in a forensic context. A chatbot without it is a very fluent colleague with no file in front of them.

What an LLM does well on these records

• Extraction at scale. Pulling every dated event, medication mention and diagnosis label out of 2,000 pages of narrative notes, and attaching the page. This is the 20-hour task in chapter 7, and it is the task the technology is good at.
• Cross-source listing. Putting the intake screen and the discharge summary next to each other because both answer "psychiatric history". The model does not know which to believe; it knows they disagree.
• Cited answers. "Which pages mention a head injury?" answered with 4 page numbers is a question you can act on. The same question answered with a paragraph and no pages is not.
• Dedupe and triage. Finding the 11 pages that appear twice and the 1 page that belongs to someone else, before a human reads anything.

Where it fails, and what the failure looks like

• Hallucination. Asked for a conclusion ("was he competent in March?") or a fact the file does not contain, a general-purpose model fills the gap with something plausible. The failure looks like a confident sentence with no page behind it. A grounded tool should answer "not found in the file" and a human should treat any uncited sentence as unverified.
• OCR errors on faxes and handwriting. Jail sick call slips, MAR initials, 1990s school records. "10 mg" becomes "16 mg"; a date loses a digit. The fix is a confidence flag on the page and a rule that those pages are read by a person.
• Copy-forward dates. A treatment plan reviewed monthly carries the original text with a new date. The model reads 12 statements where there was 1. A tool that groups near-identical text and shows the first and last date helps; so does your own eye on the content date.
• Family and shared files. Custody productions mix 2 parents and 3 children. Officer names and party names collide. Wrong-patient detection catches some of this; the rest is why every line is cited and verified.
• Abbreviation and context. "SI" is suicidal ideation in a psychiatric note and sacroiliac in an orthopedic one; "PRN" entries on a MAR are not doses given. Clinical NLP handles most of this and fails on some; chapter 9's tier 1 rule exists for the some.

The court side, stated plainly

In Mata v. Avianca (S.D.N.Y., June 22, 2023), 2 attorneys filed a brief citing 6 judicial opinions that ChatGPT had invented, complete with quotes, and then stood by them after the court asked. Judge Castel's sanctions order opens with the sentence this chapter is built on: "Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings." The penalty was $5,000 under Rule 11, plus letters to the client and to each real judge whose name had been attached to a fake opinion. Federal Rule of Civil Procedure 11(b)(3) requires that "the factual contentions have evidentiary support"; for an expert, Rule 702(b) and (d) do the same work. A hallucinated page cite in a forensic report is the expert's version of a fake case, and the opposing expert will find it by opening the page. Check your forum's local rules and the assigned judge's standing orders for any certification requirement on generative AI use in filings before counsel files anything built on your report.

The vendor checklist: HIPAA-compliant AI for forensic records

Jail and hospital records are protected health information in the custodian's hands and often in yours. A consumer chatbot is not a place for them. The checklist below is what to ask before a single page is uploaded.

• Business associate agreement signeda HIPAA-compliant AI vendor signs a BAA before you upload; no BAA, no upload
• SOC 2 report available on requestand the controls cover the AI pipeline as well as the website
• Written commitment that your files do not train the modelincluding the vendor's subprocessors
• A page citation on every line of outputchronology, tables, answers; uncited output is unusable in a report
• Confidence flags on OCR and handwritingso you know which pages to read yourself
• An AI use log of who uploaded, viewed, edited and exportedfor the chain-of-custody question on cross
• Retention you control, deletion on requestthe file leaves when the case does
• No opinion, no scoringthe tool does not rate competency, malingering, risk or parenting, and the contract says so

0 of 8 checked

The right posture is human-in-the-loop, with the human being you. Agentic AI that runs a multi-step review on its own is fine for the index and the first draft of the chronology; it is not fine for the opinion, and a vendor that blurs the line is selling you a deposition problem. For a longer treatment of the same questions see AI versus human medical record review and HIPAA-compliant AI medical record review.

2 more copyable pieces. The worksheet is the paper trail for chapter 9; keep it with the file as data considered. The prompt is written for a tool that cites pages; if yours does not, the prompt will tell you quickly, because the output will have no page numbers to check.

CHRONOLOGY REVIEW WORKSHEET
Matter: [CAPTION]   Evaluee: [NAME]   Evaluator: [NAME]   Tool and version: [TOOL]
Index version: [DATE]   Productions covered: [A to D]   Supplemental: [E, date]

Line  Date       Fact as drafted                          Source type     Cite        Tier  Page checked  Result        Initials/date
1     03/02/2026 Intake: denies psych hx, denies meds     Screening form  B1 p. 11    1     [x]           Verbatim OK   [AJ 05/14]
2     03/04/2026 Sick call: requests olanzapine 10        Handwritten     B2 p. 41    1     [x]           Read on page; "10" legible  [ ]
3     03/06/2026 NP restarts olanzapine 10 mg qhs         Clinician note  B2 p. 48    1     [ ]           [ ]           [ ]
4     03/07-03/29 MAR not produced                        Gap             B3          1     n/a           Raised with counsel [DATE]  [ ]
5     [DATE]     [FACT]                                   [TYPE]          [CITE]      [1/2/3] [ ]         [OK / corrected to: / struck]  [ ]

Tier 2 sampling: custodian [B] lines [n], sampled [n], failures [n]. If failures > 0, custodian moved to tier 1: [Y/N]
Low-confidence pages read in full: [LIST PAGES]
Inconsistencies traced to both pages: [LIST LINE PAIRS]
Missing records reconciled with counsel on [DATE]: resolved [LIST], unresolved and stated as limitations [LIST]
Procedures-section disclosure drafted: [Y/N]

You are drafting a records chronology for a forensic psychological evaluation. Use only the uploaded documents. Do not use outside knowledge about the evaluee, the case or the diagnoses.

Referral question: [COMPETENCY / RESPONSIBILITY / DAMAGES / FITNESS / RISK / CUSTODY]
Index groups: [A Court, B Jail, C Treatment history, D Prior evaluations, E Supplemental]

Produce 5 outputs.

1. CHRONOLOGY. One line per dated event: date, source type (clinician note, officer narrative, screening form, fill record, court order, test report), the fact in the record's own words where possible, and the citation as [GROUP p. N]. Sort by date. If a document's content date differs from its page or cover date, use the content date and note the other.

2. MEDICATION TABLE. Drug, dose, first seen, last seen, evidence type (prescribed, filled, administered, refused), citation. Show gaps as gaps. Do not infer adherence.

3. DIAGNOSES AND TESTING. Every diagnosis as written, by whom, date, citation. Keep educational classifications (IEP, 504) separate from clinical diagnoses. Every named psychological test: date, examiner, whether scores or protocols are in the file, citation.

4. INCONSISTENCIES. Pairs of statements from different sources that conflict, each with both citations. Describe the conflict only. Do not say which is true and do not characterize the evaluee's honesty, effort or symptom validity.

5. MISSING RECORDS. Providers, facilities, tests or date ranges referenced in the file for which no record is in the upload, each cited to the page that references it.

Rules. Every line carries a citation. If a fact is not in the file, write "not found in file" rather than guessing. Do not diagnose, do not offer an opinion on competency, responsibility, risk, damages or parenting, and do not summarize the evaluee's credibility. Flag any page you could not read with confidence (handwriting, fax, low OCR confidence) in a separate list with page numbers.

A note on the prompt's last paragraph. It is there because a model will answer the question you ask. If you ask "was the defendant malingering", a general model will tell you, and that sentence has no place in anything you sign. Ask for pages, and do the clinical work yourself.

Medrecords AI is medical record review software. You upload the productions, and it returns the index, a cited chronology with a citation on every line, the medication, diagnosis and testing tables, the inconsistencies list and the missing-records list described in chapter 8. It is built for the narrative files this guide is about: jail charts, state hospital records, prior evaluations, school and CPS productions, and the police and court file beside them. The solution pages for forensic psychological evaluators, competency to stand trial record review and child custody evaluation record review describe each referral type; the samples linked in chapters 5 and 6 show the finished documents on a fictional matter.

Its flags are signals, not verdicts. A listed inconsistency tells you where 2 pages disagree; what that means for response style is your clinical judgment with your own instruments. Throughput is about 100 pages a minute, so a 2,000-page file comes back in about 20 minutes. Security is SOC 2 and HIPAA with a BAA (security, HIPAA). Self-Service starts at 10 cents a deduplicated page, down to 5 cents at volume, duplicates free; Enterprise On-Prem is an annual license with no per-seat fees (pricing). Self-Service access starts within 1 business day of the demo. Retention is configurable, with deletion on request.

What are collateral records in a forensic psychological evaluation?

Documents about the evaluee that did not come from the evaluee's own account in the interview: jail and hospital charts, prior evaluations and test data, school records, child protective services files, police and court records, pharmacy fill histories and prior IMEs. The APA Specialty Guidelines ask forensic practitioners to avoid relying on 1 source and to corroborate important data whenever feasible.

Why can a forensic evaluator not rely on the evaluee's self-report?

Because the evaluee has a stake in the outcome and because memory, illness and the setting all distort the account. Guideline 9.02 says to corroborate and, where you cannot, to disclose that the data are uncorroborated. Rule 702(b) requires an opinion to rest on sufficient facts or data, and the records are most of the data.

What is the Dusky standard for competency to stand trial?

The 1960 Supreme Court test: whether the defendant has "sufficient present ability to consult with his lawyer with a reasonable degree of rational understanding" and "a rational as well as factual understanding of the proceedings against him". Federal statute restates it at 18 USC 4241(a); states have their own versions built on it.

Can a forensic psychologist obtain psychotherapy notes?

Only with a separate authorization or a court order. Under 45 CFR 164.508(a)(2) a covered entity must obtain an authorization for any use or disclosure of psychotherapy notes, apart from narrow exceptions. The ordinary chart (diagnoses, medications, session dates, treatment plans, test results) is not psychotherapy notes and comes with the standard authorization.

How long does a federal competency evaluation take?

A defendant committed for the examination can be held "not to exceed thirty days", with 1 extension "not to exceed fifteen days", under 18 USC 4247(b). If found incompetent, the first restoration commitment is capped at 4 months by 18 USC 4241(d). State timelines differ and the forum's statute controls.

Can AI summarize psychiatric records for a forensic evaluation?

Yes, for the extraction part. AI medical record review tools build an index, a cited chronology, a medication table and a list of inconsistencies from narrative records, each line linked to its page. The evaluator verifies every fact used in the report, does the symptom-validity work and writes the opinion. A tool that offers an opinion on competency, risk or malingering has gone outside what the method can support.

Is it HIPAA compliant to upload jail or hospital records to an AI tool?

Only to a vendor that signs a business associate agreement and protects the records under the HIPAA Security Rule, with a written commitment that your files do not train its models. Consumer chatbots without a BAA are not a place for an evaluee's records. Ask for the SOC 2 report and for a citation on every line of output.

Can ChatGPT build a psychiatric chronology for court?

A general chatbot can produce a chronology-shaped document, but without page-level citations its output cannot be checked, and it will fill gaps with plausible text the records do not contain. In Mata v. Avianca (2023) a court sanctioned attorneys $5,000 under Rule 11 for filing 6 fabricated AI-generated case citations. The expert's equivalent is a fabricated page cite, and the opposing expert will open the page.

Will using AI get my report excluded under Daubert or Rule 702?

Using a document review tool for extraction is not itself a reliability problem; relying on unverified output is. Rule 702 asks whether the opinion rests on sufficient facts or data and reflects a reliable application of the method to the facts. Disclose the tool in your procedures section, describe the verification you did, and keep the index and worksheet as data considered under Guideline 10.06.

Does Medrecords AI score symptom validity or detect malingering?

No. It lists inconsistencies between sources with both page citations, as signals for the evaluator's own symptom-validity work with the evaluator's own instruments. It never renders the opinion, scores effort or response style, or characterizes the evaluee's credibility. You review, you revise, you sign.

Guidelines, statutes, rules and opinions were checked against the primary sources listed below in October 2026, and quoted text is verbatim. The competency hypothetical, the index, the medication table, the diff and the hours ledger are illustrations and are labeled as such. Product facts come from this site's product pages. Nothing here is legal, medical or clinical advice; competency, responsibility, damages, fitness, risk and custody law vary by state and the forum's statute and case law control.

• APA Specialty Guidelines for Forensic Psychology, American Psychologist 68(1), 7 to 19 (2013): Guidelines 2.05, 9.01, 9.02, 9.03, 10.06, 11.01, 11.02, 11.03 and 11.04.
• APA Guidelines for Child Custody Evaluations in Family Law Proceedings, American Psychologist 65(9), 863 to 867 (2010): Guideline 10 on multiple methods of data gathering.
• AAPL Practice Guideline for the Forensic Assessment, Glancy et al., J Am Acad Psychiatry Law 43(2 Supplement), S3 to S53 (2015): collateral sources, personal review of records, listing sources requested but not received. The guideline states it should not be construed as dictating the standard for forensic evaluations.
• Dusky v. United States, 362 U.S. 402 (1960), per curiam; Jackson v. Indiana, 406 U.S. 715 (1972); Daubert v. Merrell Dow Pharmaceuticals, 509 U.S. 579 (1993).
• 18 USC 4241, competency standard and restoration period; 18 USC 4247, examination period (b) and report contents (c).
• Federal Rule of Evidence 702, as amended December 1, 2023; current Federal Rules of Evidence at the United States Courts; Federal Rule of Civil Procedure 11.
• 45 CFR 164.508, authorizations and psychotherapy notes; 45 CFR 164.512(e), judicial and administrative proceedings; 42 CFR 2.64, court orders for substance use disorder records; 34 CFR 99.31(a)(9), FERPA disclosure under a judicial order or subpoena.
• Mata v. Avianca, Inc., 22-cv-1461 (PKC), Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023), as posted by the United States District Court for the District of Connecticut.
• Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Magesh, Surani, Dahl, Suzgun, Manning and Ho, Stanford RegLab (2024).

Related guides: collateral records in forensic evaluation, the AI-native IME physician, AI versus human medical record review, a verified medical chronology from scanned records, how to get medical records for a lawsuit, medical abbreviations for legal professionals and child abuse medical records.