It is 4:40 PM and Maria's colleague Dana, an LCSW with the infant and early childhood mental health endorsement, has just finished a Parent-Child Interaction Therapy session: forty-five minutes behind a one-way mirror, coaching a mother through an earbud while a three-year-old built a block tower, knocked it down, screamed, and slowly came back. Dana now owes the chart three different documents: a live-coaching observation log with DPICS coding tallies, a dyadic interaction summary, and an updated parent-coaching plan, all anchored to a DC:0-5 diagnosis that her billing system only understands after a crosswalk to DSM-5-TR and ICD-10. This is the densest documentation load in outpatient behavioral health, and it is exactly where clinicians are most tempted to let AI fill the gaps. This lesson teaches the early-childhood documentation stack end to end: PCIT live-coaching notes, dyadic parent-infant work, 0-5 ECMH consultation, DC:0-5 to DSM-5-TR crosswalking, and the state Medicaid billing pathways that fund it. It also draws the hardest line in the chapter: AI does not score the DPICS or any observational instrument, and AI does not invent attachment observations it never saw. By the end you will build a complete DC:0-5 Crosswalked Dyadic-Session Documentation Set you can adapt to your own 0-5 caseload.

Why Early-Childhood Documentation Is a Different Animal

Most psychotherapy documentation describes one patient. Early-childhood mental health documentation describes a relationship. In PCIT, in dyadic parent-infant psychotherapy, and in 0-5 ECMH consultation, the unit of treatment is the dyad: the caregiver and the child together, and the interaction patterns between them. The note has to capture what the child did, what the caregiver did in response, what the clinician coached in the moment, and how the interaction shifted, all without drifting into language that pathologizes the parent or adjudicates parenting quality. A note that reads like a custody evaluation is a clinical failure and a legal liability, because dyadic charts are subpoenaed in dependency and family-court matters more often than almost any other outpatient record.

The diagnostic frame is different too. The DSM-5-TR was not built for a 14-month-old. The field's answer is DC:0-5, the diagnostic classification of mental health and developmental disorders of infancy and early childhood, maintained under ZERO TO THREE and used across infant mental health practice, including AAIMHI-aligned training and endorsement systems. DC:0-5 gives you developmentally valid categories for the 0-5 population, a multiaxial structure that includes the caregiving relationship itself, and a vocabulary for relational disturbance that the DSM simply lacks. The catch: most EHRs, most payers, and most state Medicaid claims systems do not accept a DC:0-5 code on a claim. They want DSM-5-TR and ICD-10-CM. So every DC:0-5 diagnosis lives in the chart alongside a crosswalked DSM-5-TR equivalent, and the clinician documents both the DC:0-5 formulation that drives treatment and the crosswalk that drives billing.

Hold one picture in your head for this entire lesson: the clinician behind the one-way mirror is a flight instructor in the control tower, and the parent is the pilot in the cockpit. The instructor sees everything through the glass, radios real-time coaching through the headset, and logs every maneuver. AI, in this picture, is the transcriptionist sitting next to the instructor in the tower. The transcriptionist can hear the radio calls and type fast. The transcriptionist cannot see through the glass, cannot judge whether the landing was smooth, and absolutely cannot grade the pilot's checkride. Only the instructor saw the flight. Only the instructor scores it. That is the entire AI governance model for dyadic work, and we will return to the tower repeatedly.

The PCIT Session, DPICS Coding, and Who Holds the Tally Sheet

PCIT runs in two phases. In Child-Directed Interaction (CDI), the parent learns to follow the child's lead using the PRIDE skills: Praise, Reflection, Imitation, Description, and Enjoyment, while dropping commands, questions, and criticism. In Parent-Directed Interaction (PDI), the parent learns to give effective commands and follow through with consistent, calm consequences. Progress through the protocol is not vibes; it is measured. The measurement instrument is the DPICS, the Dyadic Parent-Child Interaction Coding System: a structured behavioral observation coding system in which the credentialed clinician watches a timed interaction segment and tallies discrete parent verbalizations into categories such as labeled praise, unlabeled praise, behavioral description, reflection, question, command, and negative talk. Phase mastery criteria are defined in DPICS terms, so the coding tallies are the spine of the clinical record.

Here is the hard limit, and it is not negotiable: AI does not score the DPICS, and AI does not score any observational instrument. A DPICS tally is a direct behavioral observation made in real time by a clinician trained on the coding system. The categories depend on tone, context, and the visible behavior of a child the language model never saw. An audio transcript cannot reliably distinguish a labeled praise delivered warmly from the same words delivered sarcastically, cannot see whether the child complied within the coding window, and cannot observe the nonverbal sequence that determines the code. An LLM offered a transcript will happily produce a confident-looking tally anyway, and that fabricated tally would then drive phase-advancement decisions in an evidence-based protocol. The transcriptionist in the tower never grades the checkride. The clinician holds the tally sheet, codes live, and the numbers in the note are the clinician's numbers, transcribed exactly.

What AI legitimately does in a PCIT workflow is structure and expansion after the observation. The clinician finishes the session holding a paper or tablet tally: labeled praise 11, behavioral descriptions 9, reflections 7, questions 3, commands 1, negative talk 0. AI takes those clinician-supplied numbers plus the clinician's shorthand and builds the formatted observation log: session number, phase, coaching segment length, the tallies in a clean table-like structure, mastery criteria status, and the homework assigned. The verification rule mirrors every risk lesson in this program: read the numbers in the draft against the tally sheet digit by digit before signing, because a transposed tally is a corrupted outcome measure in a manualized treatment.

DC:0-5, the DSM-5-TR Crosswalk, and Why You Document Both

DC:0-5 exists because infant and early-childhood presentations do not map cleanly onto adult-derived categories. A toddler's sleep disturbance, feeding difficulty, or disorder of relating reads differently at 18 months than at 18 years, and DC:0-5 captures relational context as a formal part of the diagnostic picture rather than a footnote. For the treating clinician, DC:0-5 is the formulation language: it tells you and the consultation team what is actually going on in this dyad and what the treatment targets are.

The billing system does not speak that language. Claims run on ICD-10-CM, and medical-necessity review runs on DSM-5-TR criteria, so the chart carries a crosswalk: the DC:0-5 diagnosis the clinician assigned, the nearest DSM-5-TR equivalent, and the ICD-10 code that goes on the claim, with a sentence of clinical reasoning connecting them. The crosswalk is a clinical judgment, not a lookup table exercise. Two children with the same DC:0-5 category can crosswalk to different DSM-5-TR codes depending on the dominant presentation, and some DC:0-5 categories have no tidy DSM twin, forcing the clinician to choose the least-distorting equivalent and document why. AI can format the crosswalk block and keep the three code systems aligned across documents, and it can flag an internal inconsistency, such as a treatment plan that names a diagnosis the crosswalk block does not contain. AI does not choose the codes. Diagnosis is the clinician's act in every jurisdiction's practice act, and a crosswalk is two diagnostic decisions stapled together.

The payer layer adds a third vocabulary. Several state Medicaid programs fund early-childhood mental health work through dedicated pathways: California supports ECMH services within Medi-Cal, Minnesota's ECMH provisions operate under its IIIBC framework, and Oregon funds Early Childhood Mental Health Consultation. Each pathway has its own documentation expectations, and dyadic services billed through them must show the dyadic frame in the note: who was present, that the caregiver-child interaction was the treatment target, and how the intervention addressed the relationship. A note written like individual child therapy can sink an ECMH claim even when the session itself was textbook dyadic work. This is the payer-content rule from earlier levels in a new costume: the verifiable details AI cannot supply are exactly the ones the claim lives on, which caregiver attended, the minutes of live coaching, the specific DPICS tallies, and the modality named.

The language model can hear the radio calls, but it cannot see through the glass. Anything in a dyadic note that depends on seeing the room, an attachment behavior, a DPICS code, a child's affect shift, must come from the clinician who watched it happen.

Dyadic Parent-Infant Work: Documenting Attachment Observations Without Inventing Them

Below PCIT's structured protocol sits the broader world of dyadic parent-infant psychotherapy: work with a parent and a baby or toddler where the clinical material is the moment-to-moment interaction, proximity-seeking, gaze, soothing, repair after rupture, the parent's reflective capacity. The session note for this work is observational in a way few other notes are. "Infant oriented to mother's face during feeding, mother responded with sustained eye contact and vocal matching; following a startle, infant sought proximity and was soothed within approximately one minute" is clinical data. It is also data only the person in the room possesses.

This is where ambient AI scribes fail structurally, not just occasionally. A scribe builds notes from audio. Most of what matters in a parent-infant session is silent: the gaze, the reach, the stiffening, the turn away. An AI drafting from the audio of a dyadic session will either produce a note that is mostly empty or, worse, will infer interactional content from the parent's narration and the clinician's spoken reflections, generating attachment-flavored observations that were never observed. An invented line like "secure attachment behaviors noted" is a fabricated clinical finding in a chart that may one day be read aloud in dependency court. The discipline: the clinician dictates or types the behavioral observations personally, in their own observational language, and AI's role is confined to organizing those clinician-authored observations into the note structure, expanding the clinician's shorthand without adding interactional content, and formatting the dyadic interaction summary. A standing instruction belongs in every scribe configuration used for 0-5 work: do not generate, infer, or characterize any caregiver-child interaction, attachment behavior, or child affect not explicitly stated by the clinician.

The same restraint applies to caregiver-facing language. Dyadic notes are read, eventually, by parents, by opposing counsel, by dependency social workers. The note describes behavior, not parental worth: "mother issued twelve commands in five minutes" is observation; "mother is controlling" is characterization that will be quoted back in a courtroom. AI is actually useful here as a tone auditor: after the clinician drafts, AI can flag evaluative adjectives applied to the caregiver and propose behavioral rewording. The clinician decides what stands.

0-5 ECMH Consultation: When Your Client Is a Classroom

Early-childhood mental health consultation is the third stream of this lesson, and it bends the documentation rules again. In ECMH consultation, the clinician consults to a childcare program, preschool, or home-visiting team about a child, a classroom, or the program's overall capacity, often without the child ever being the clinician's patient. Oregon's Early Childhood Mental Health Consultation funding, California's Medi-Cal ECMH supports, and Minnesota's IIIBC structure each fund versions of this. The documentation is consultation documentation: who requested the consult, the observation conducted, the recommendations made to the adults, and the follow-up plan. There may be no diagnosis at all, because consultation is often programmatic rather than clinical, and the consent architecture runs through the program and the family rather than a standard treatment consent.

AI fits ECMH consultation work better than almost any task in this lesson, precisely because the deliverables are adult-facing documents: the classroom observation write-up structured from the consultant's field notes, the teacher-strategy summary, the consultation-visit log the funder requires. The constraints still travel. The consultant's classroom observations are direct observations; AI formats them and does not embellish them. Names of non-client children observed incidentally do not belong in the consult record at all, and an AI drafting from raw field notes must be instructed to strip them. And if the consultation surfaces a concern rising to a mandated-report threshold, that determination is the clinician's alone, made under the state reporting statute, exactly as the mandatory-reporter lesson in this chapter taught. AI never makes the mandated-report call, in a clinic or a classroom.

One more boundary specific to consultation and coaching plans generally: AI does not author goals for the participant. Whether the participant is a parent in PCIT homework, a caregiver in dyadic treatment, or a teacher receiving consultation strategies, the goals in the plan are produced by the clinician with the participant, in the participant's words where possible, exactly as the Stanley-Brown lesson taught for safety plans. A coaching plan whose goals were composed by a language model is a plan the participant never agreed to, and adherence research is unambiguous that ownership drives follow-through. AI types the plan after the people in the room have built it.

Billing the Dyad: Codes, Minutes, and the Medicaid Pathways

Dyadic and PCIT sessions are commonly billed under the family psychotherapy codes, 90847 when the identified patient is present, which in PCIT they almost always are, or under individual psychotherapy time codes such as 90832, 90834, or 90837 where the payer's rules and the session structure support it, with the intake itself under 90791. Where measurement tools are administered and the payer covers it, CPT 96127 can apply to brief standardized instruments. The state ECMH pathways named above, CA Medi-Cal ECMH, MN IIIBC, OR ECMH consultation funding, layer state-specific service codes and documentation rules on top. The non-negotiable habit: confirm your specific payer's rules for dyadic billing before the first claim, because dyadic work sits in a coding gray zone where assumptions get recouped.

Whatever the code, the auditable elements are the ones this program has drilled since Level 1: who was present and for how long, the start and stop time or total minutes for time-based codes, the modality named (PCIT CDI coaching, child-parent dyadic intervention, ECMH consultation), and the medical-necessity thread connecting the diagnosis, the treatment plan, and what happened in this session. AI can assemble that skeleton; only the clinician can supply the true minutes, the actual tallies, and the modality that was really delivered. A 90847 note that does not establish the caregiver's presence and the relational target is a denial waiting in the queue, and a PCIT note whose DPICS numbers were "estimated" by a model is worse than a denial: it is a corrupted evidence-based treatment record. The clinician signs the note, and the signature is a legal attestation that every number and observation in it is the clinician's own. Read every word before signing.

The AI-Permitted Map for the 0-5 Stack

Pull the boundaries into one map you can hand a supervisee. AI may: expand clinician shorthand into the formatted PCIT observation log after the clinician supplies every tally; structure the dyadic interaction summary from clinician-authored observations; format the DC:0-5 to DSM-5-TR to ICD-10 crosswalk block the clinician populated; draft the parent-coaching plan document from goals the clinician and caregiver produced together; assemble the ECMH consultation write-up from the consultant's field notes; audit any draft for evaluative caregiver language, internal code inconsistencies, and missing billing elements; and translate parent-facing homework sheets into plain language at an appropriate reading level.

AI may not: score the DPICS or any observational coding system; generate or infer any attachment, interaction, or affect observation; choose or change a DC:0-5 or DSM-5-TR code; author goals for the parent, caregiver, or teacher; make a mandated-report determination from session or classroom content; characterize parenting quality in any document; or supply minutes, attendance, or tallies it was not given. Notice the pattern: every prohibited item is something that requires either eyes in the room or a licensed clinical judgment. The tower transcriptionist types what the instructor radios; the instructor flew the observation and owns every score. Write this map into your practice's AI policy as a 0-5 appendix, because a generic scribe policy written for adult individual therapy does not anticipate a single one of these failure modes.

Vendor diligence follows from the map. Before any scribe touches a dyadic caseload, the questions are: can ambient capture be disabled per-session for observation-heavy work; can a standing no-inference instruction be configured at the template level; does the vendor's model ever receive the audio of a minor and under what BAA terms; and does the template library actually contain a dyadic structure, or will the tool keep forcing your relational work into an individual SOAP shape? A tool that cannot answer those questions cleanly stays out of the 0-5 room.

The Applied Problem: The DC:0-5 Crosswalked Dyadic-Session Documentation Set

Your artifact is a three-document set for a single simulated PCIT case: a CDI-phase session with a parent and a 4-year-old, clinician-coded DPICS tallies in hand. Document one is the Live-Coaching Observation Log; document two is the Dyadic Interaction Summary; document three is the Parent-Coaching Plan. All three carry a shared header block containing the DC:0-5 diagnosis, the crosswalked DSM-5-TR equivalent, the ICD-10-CM claim code, and one sentence of crosswalk reasoning written by you.

Step one, create your clinician inputs on paper first, exactly as you would after a real session: the DPICS tallies (invent a plausible set, such as labeled praise 11, behavioral description 9, reflection 7, questions 3, commands 1, negative talk 0), the coaching-segment minutes, three behavioral observations in your own words, and two coaching goals phrased the way the parent agreed to them. This step matters because the artifact teaches the workflow's order of operations: clinician observations exist before AI touches anything. Step two, run the structuring prompt: "You are formatting early-childhood mental health documentation. Using only the clinician-supplied data below, produce three documents: (1) a PCIT CDI live-coaching observation log including the DPICS tallies exactly as given, session minutes, and mastery-criteria status as stated by the clinician; (2) a dyadic interaction summary using only the behavioral observations provided, with no inferred interaction, attachment, or affect content; (3) a parent-coaching plan containing only the two goals provided, in the caregiver's wording. Do not add, score, estimate, or characterize anything. Flag any field you were not given rather than filling it." Paste your inputs beneath it.

Step three is the verification pass, and it is the lesson in miniature. Check every DPICS number against your tally sheet digit by digit. Hunt the dyadic summary for any sentence describing an interaction you did not write; delete on sight. Confirm the goals are verbatim yours. Read the crosswalk header in all three documents for consistency. Then run the tone audit: ask the model to flag any evaluative language about the caregiver, and convert anything flagged into behavioral description. Done looks like this: three documents a PCIT supervisor would countersign, in which every observation, score, code, and goal can be traced to a clinician-authored source, and the only thing AI contributed was structure, formatting, and the audit. Save the prompt block; it becomes your standing 0-5 template.

Key Takeaways

• Early-childhood mental health documentation describes a relationship, not just a patient. PCIT, dyadic parent-infant work, and ECMH consultation all chart the caregiver-child interaction itself, which raises both the documentation load and the subpoena exposure of every note you write.
• AI never scores the DPICS or any observational instrument. DPICS tallies are real-time coded observations by a trained clinician; a model working from audio cannot see compliance, tone, or nonverbal sequence, and a fabricated tally corrupts the mastery criteria of an evidence-based protocol.
• DC:0-5 is the formulation language; DSM-5-TR and ICD-10-CM are the billing languages. The chart carries both, joined by a crosswalk that is a clinical judgment the clinician makes and explains. AI formats and consistency-checks the crosswalk block; it never selects a code.
• AI must never generate or infer attachment, interaction, or affect observations. Most dyadic clinical data is silent and visual; the clinician authors every behavioral observation personally, and the scribe configuration carries a standing no-inference instruction for all 0-5 work.
• AI does not author goals for the participant. PCIT homework, dyadic coaching goals, and ECMH teacher strategies are built by the clinician with the parent or teacher, in their words; AI types the plan after the people in the room have produced it.
• Bill the dyad deliberately. 90847 with the identified patient present is the common frame, state Medicaid ECMH pathways (CA Medi-Cal ECMH, MN IIIBC, OR ECMH consultation funding) add their own rules, and the audit-proof elements, attendance, minutes, modality, tallies, are exactly the details only the clinician can supply.
• The clinician signs, and the signature attests that every score and observation is the clinician's own. The tower transcriptionist types what the instructor radios; the instructor watched the flight, holds the tally sheet, and grades the checkride. Read every word before signing.