# Usage: curl -sSL https://seed.show/clinical.documentation.us | bash -s <install-path>
# <install-path> is the directory where the file should land.

set -euo pipefail
[ -z "${1:-}" ] && {
  echo "install requires a path: curl -sSL https://seed.show/clinical.documentation.us | bash -s <install-path>" >&2
  exit 1
}
TARGET="$1"
mkdir -p "$TARGET"
DEST="$TARGET/seed-fold.EEFDdM.folded.md"

cat > "$DEST" <<'PORTDOWN_996AEC7E'
<!--
This is a `.folded.md` archive — a directory packed into one markdown
file. The markers below are load-bearing; don't edit them directly.

To unpack (requires bash — if you have no shell, tell the user):
  1. curl -sSLf https://fold.dom.vin/skill | bash -s <INSTALL_DIR>
  2. <INSTALL_DIR>/fold/scripts/unfold <this-file>
     (or: unfold <this-file>  if fold/scripts is on your PATH)
-->

---
fold: true
marker: 2c5c5a
at: 2026-05-07T16:15:59Z
root: seed-pack.8XQDhu
---

<!--fold:2c5c5a@file path="README.md" mode="644"-->
# clinical.documentation.us

US clinical documentation context for agents. What to know before generating, reviewing, or coding a clinical note.

**Agents must never generate documentation that will be submitted to payers without physician review and authentication.** Clinical notes are legal documents and billing instruments. An AI-generated note that bypasses physician attestation is both a compliance violation and a potential fraud liability. Every draft produced by an AI system requires a licensed provider to review, correct, and authenticate before it enters the medical record or supports a claim.

---

## What clinical documentation is actually for

Clinical documentation serves four distinct purposes that impose different — sometimes contradictory — requirements on the same note. Understanding which purpose is driving a task changes what "correct" looks like.

**1. Communication between providers**
The primary clinical function: the note is how the discharging hospitalist tells the accepting cardiologist what happened, how the PCP communicates a patient's history to a consulting surgeon, how the on-call resident understands what the day team found. This purpose demands clarity, narrative coherence, and honest uncertainty. The reader is a clinician making decisions about a patient.

**2. Billing and reimbursement justification**
The payer function: the note is the legal basis for the claim. Medicare, Medicaid, and commercial payers do not pay for services they cannot audit from documentation. The note must demonstrate that the service actually occurred, that it was medically necessary, and that it reached the documented level of complexity. A complete, coherent narrative that doesn't map to the payer's documentation requirements still results in denial or audit. This purpose demands specificity, completeness, and alignment with coding rules — not readability.

**3. Legal and regulatory record**
The note is a legal document: it can be subpoenaed, introduced in malpractice litigation, used in licensing board proceedings, and audited by OIG. Amendments must be clearly dated and signed; the original entry cannot be altered. Late entries must be labeled as late. A note "cleaned up" after the fact to match a billing code is a liability problem, not a documentation improvement.

**4. Quality measurement and population health**
Payers, health systems, CMS, and accreditation bodies (Joint Commission, NCQA) use coded diagnoses and documented conditions to measure quality, assign risk scores, and calculate value-based payments. HCC risk adjustment pays Medicare Advantage plans based on the documented complexity of their enrolled patients — every underdocumented chronic condition is revenue left on the table and a risk score that doesn't reflect the patient's actual burden of illness. This purpose demands annual capture of all active chronic conditions, not just the ones relevant to today's visit.

These four purposes do not always align. A note optimized for communication might say "likely pneumonia, watching for improvement" — medically honest, billing-fatal. A note optimized for coding might document seventeen conditions at full ICD-10 specificity — billable, but nearly unreadable as a clinical document. Agents working in this space must know which purpose they're serving.

---

## SOAP note structure and why each element matters

SOAP (Subjective, Objective, Assessment, Plan) is the dominant outpatient documentation framework. Each section has distinct content requirements and distinct failure modes.

**Subjective**
The patient's report: why they came in, in their own terms. Includes the History of Present Illness (HPI) — the narrative of the current problem — and the Review of Systems (ROS), where the provider asks about symptoms across body systems beyond the chief complaint. HPI and ROS are the billing engine of outpatient E&M: for higher-level visits, they must be both present and sufficiently detailed.

Failure mode: recording only the chief complaint ("patient presents with knee pain") without capturing the HPI elements that establish the visit's complexity — onset, location, duration, character, aggravating factors, relieving factors, radiation, severity.

**Objective**
The provider's findings: vital signs, the Physical Exam (PE), and relevant diagnostic data (labs, imaging, EKGs). The exam must be documented by body system and organ area. The number of systems examined contributes to E&M level under the 1995 and 1997 documentation guidelines.

Failure mode: "exam unremarkable" or "normal" instead of a specific, system-by-system account. Payers and auditors cannot verify what was actually examined from a blank assertion.

**Assessment**
The provider's clinical judgment: the diagnosis or working diagnosis, the differential if definitive diagnosis isn't established, and — critically — the Medical Decision Making (MDM) that explains how the provider arrived at the diagnosis. MDM is the single most important element for E&M level determination under the 2021 AMA guidelines.

Failure mode: stating the diagnosis without the reasoning. "Assessment: Type 2 diabetes mellitus, uncontrolled" is a code. The MDM is the clinical logic that makes the assessment credible — complexity of problems addressed, data reviewed, and risk of complications or treatment.

**Plan**
What happens next: medications prescribed or changed, referrals ordered, procedures performed or planned, patient education given, follow-up interval. Each diagnosis in the Assessment should have a corresponding element in the Plan.

Failure mode: documenting the plan without anchoring it to the Assessment's diagnoses. Payers look for one-to-one correspondence between diagnoses and plan elements as a proxy for medical necessity.

---

## MEAT and TAMPER: the documentation test for chronic conditions

Two mnemonic frameworks define whether a chronic condition is sufficiently documented to be coded and counted for risk adjustment.

**MEAT** (the condition is documented if the note contains at least one element):
- **M — Monitoring:** tracking signs, symptoms, or test results (e.g., "A1c trending down, continue monitoring quarterly")
- **E — Evaluating:** assessing a condition's status or progression (e.g., "HbA1c 8.1, diabetes not at goal")
- **A — Assessing/Addressing:** any direct provider attention to the condition during the encounter (e.g., "counseled patient on diet modification for glycemic control")
- **T — Treating:** active treatment, medication management, or referrals (e.g., "increased metformin to 1000mg BID")

A chronic condition that appears in the problem list but is not addressed under MEAT criteria during the encounter should not be coded as active. A condition actively addressed by any MEAT element should be coded even if it wasn't the chief complaint.

**TAMPER** (documentation framework used by some CDI programs and MA risk adjustment auditors):
- **T — Test results ordered or reviewed**
- **A — Assessment of the condition documented**
- **M — Medication prescribed, changed, or continued**
- **P — Patient education provided**
- **E — Evaluation or monitoring plan stated**
- **R — Referral or consultation ordered**

TAMPER is an audit framework more than a documentation guide — payers and RAD-V auditors use it to evaluate whether a coded condition was genuinely "addressed" during the visit. If a coded HCC condition has no corresponding TAMPER element in the note, the code is vulnerable to audit recoupment.

---

## ICD-10 specificity: the hierarchy problem

ICD-10-CM codes are hierarchical. The agent must code to the highest specificity the documentation supports.

- **Hypertension:** I10. One code; no further specificity required.
- **Type 2 diabetes mellitus:** E11.x — the correct code depends on complications. E11.9 (no complications), E11.65 (with hyperglycemia), E11.319 (with unspecified diabetic retinopathy), E11.40 (with diabetic neuropathy, unspecified), etc. Coding E11.9 when the note documents retinopathy is a specificity gap.
- **Sepsis:** A41.9 (unspecified) vs. A41.01 (MRSA), A41.51 (Gram-negative), etc. The causative organism, if documented, must be captured in the code.
- **Chronic kidney disease:** N18.x — staging is required (N18.1 through N18.6, N18.9). "CKD" without a stage is N18.9 (unspecified), valid only if staging genuinely isn't documented or determinable.

The rule: an unspecified code (typically ending in .9 or .90) is valid only when the documentation genuinely cannot support greater specificity — not as a default. Choosing an unspecified code when the note describes specifics needed for a more precise code is a documentation gap that triggers CDI queries and payer audits.

Codes update annually on October 1. The ICD-10-CM FY guidelines are the authoritative source — see sources.md.

---

## E&M level selection and medical necessity

Under the 2021 AMA E&M guidelines (in effect for office/outpatient visits 99202–99215), visit level is determined by either MDM complexity or total time — whichever the provider selects. The prior element-counting method (number of HPI elements, ROS systems, exam elements) is retired for office visits but survives in some payer contracts and audit contexts.

**MDM has three components; two must meet the required level:**
1. Number and complexity of problems addressed
2. Amount and/or complexity of data reviewed and analyzed
3. Risk of complications and/or morbidity or mortality of patient management

| E&M Level | Problems | Data | Risk |
|---|---|---|---|
| 99202 / 99212 | Straightforward | Minimal/none | Minimal |
| 99203 / 99213 | Low | Limited | Low |
| 99204 / 99214 | Moderate | Moderate | Moderate |
| 99205 / 99215 | High | Extensive | High |

**Medical necessity** is the payer standard overlaid on top of MDM: the documented diagnoses must support the service billed. A 99215 billed for a problem that the documentation treats as minor and straightforward will be downcoded or denied on audit. Medical necessity and visit level must be internally consistent across the Assessment, Plan, and claim.

---

## What agents get wrong

**Documenting the diagnosis without supporting the medical decision-making.**
Under 2021 AMA E&M guidelines, MDM drives the visit level. An agent that produces a diagnosis and a plan without documenting the clinical reasoning that connects them generates a note that looks complete but cannot defend its billing level under audit.

**Missing ICD-10 specificity.**
Defaulting to unspecified codes when the note contains the information required for a more precise code. This is the most common AI coding error and the primary CDI audit trigger.

**Conflating the assessment with the plan.**
Documenting what was done without documenting why. The Assessment must contain MDM reasoning — the absence collapses billing justification regardless of plan accuracy.

**Treating documentation as transcription.**
Ambient AI scribing tools capture what was said. Transcription is not documentation: the spoken exchange often lacks the structured elements payers require. A provider who reviews systems verbally but never says "no fever, no chills, no weight loss" gives the scribe nothing to document for ROS. Agents must flag gaps — missing HPI elements, undocumented ROS, absent MDM reasoning — not silently omit them.

**Failing to capture all active chronic conditions.**
For risk-adjusted payment models (Medicare Advantage, ACOs, CPC+), every visit is an opportunity to document all relevant chronic conditions, not just the ones that drove today's encounter. A diabetic patient seen for a sprained ankle still has diabetes; if it's not documented under MEAT criteria, it doesn't count for HCC risk adjustment.

**Generating notes that skip physician attestation.**
An AI-drafted note that goes directly from scribe to chart without physician review and authentication is a compliance violation. Agents must treat every generated note as a draft requiring physician signature before it constitutes a medical record or supports a claim.

---

## What AI is changing

**Ambient documentation tools**
Nuance DAX Copilot, Suki, Nabla, Abridge, and others passively listen to the provider-patient encounter and generate a structured draft note. These are now deployed at scale across health systems. They reduce documentation burden materially but introduce specific failure modes: they capture what was spoken, not what must be documented; they can hallucinate clinical details not said aloud; and they produce notes that appear complete while missing billable elements.

**FDA SaMD classification**
AI clinical documentation tools may be subject to FDA oversight as Software as a Medical Device (SaMD) under 21 CFR Part 820 and the FDA's Digital Health Center of Excellence guidance, depending on their intended use. A tool that generates documentation that clinicians use to make treatment decisions (as opposed to purely administrative documentation) may cross into the device classification. As of 2024–2025, regulatory clarity is still evolving. Developers and deployers should monitor FDA's AI/ML-Based SaMD Action Plan.

**HIPAA implications**
Any AI documentation tool that processes protected health information (PHI) must comply with the HIPAA Privacy Rule (45 CFR Part 160/164). This includes:
- Business Associate Agreements (BAAs) with AI vendors who process PHI
- Minimum necessary standards — models should not ingest more PHI than required
- Audit logging requirements for who accessed and modified records
- Patient rights considerations around AI-generated content in their record

Cloud-based ambient scribing tools where audio is transmitted to external servers require BAAs and careful review of data retention policies.

**Physician review requirements**
CMS and major payers have not issued explicit rules prohibiting AI-generated notes, but existing documentation requirements — authentication, accuracy, medical necessity — apply regardless of how a note was generated. A physician who signs an AI-generated note attests to its accuracy and completeness; they bear full liability for what it contains. AHIMA's position (2023) is that AI-generated content must be clearly identified in the record and that providers who use AI scribing tools remain fully responsible for the final note.

**CMS guidance on AI-generated documentation**
CMS has not issued a blanket prohibition on AI-assisted documentation but has signaled in multiple transmittals and the CY2025 Physician Fee Schedule final rule preamble that AI tools do not change existing documentation standards. The evidentiary requirements for E&M coding, MDM, and medical necessity apply to AI-assisted notes the same as to dictated or typed notes. OIG's Work Plan includes ongoing scrutiny of documentation-based billing; AI-generated notes that inflate E&M levels or fabricate clinical detail are subject to the same False Claims Act exposure as any other upcoding.

**What defensible AI-assisted documentation looks like**
- Clear attribution: the note identifies that AI assistance was used (some EHR systems are adding this as a required field)
- Physician review before authentication: the provider reads the draft, corrects errors, and adds reasoning not captured by the AI
- Gap-flagging: the AI tool surfaces missing elements (undocumented ROS, absent MDM reasoning) rather than silently omitting them
- No hallucinated specifics: the AI does not generate clinical details (lab values, exam findings, diagnosis codes) not present in the encounter
- Audit trail: the EHR records that a draft was AI-generated and tracks the provider's edits before authentication
<!--fold:2c5c5a@file path="glossary.md" mode="644"-->
# Clinical Documentation Glossary

Precise definitions for terms that are frequently misused or conflated in AI-generated clinical documentation and coding.

---

**HPI (History of Present Illness)**
The narrative section of the Subjective that describes the current problem in detail. The eight HPI elements defined by CMS documentation guidelines are: location, quality, severity, duration, timing, context, modifying factors, and associated signs and symptoms. For E&M level-setting, HPI must be "extended" (four or more elements, or three or more chronic conditions) to support moderate- or high-complexity visits. Agents often capture the chief complaint only and omit the HPI elements entirely.

**ROS (Review of Systems)**
A systematic inventory of symptoms across body systems beyond the chief complaint — asked by the provider, reported by the patient. Systems include constitutional, eyes, ENT, cardiovascular, respiratory, GI, GU, musculoskeletal, integumentary, neurological, psychiatric, endocrine, hematologic/lymphatic, and allergic/immunologic. A "complete" ROS covers ten or more systems. The ROS is distinct from the Physical Exam: ROS is what the patient reports; PE is what the provider finds.

**PE (Physical Exam)**
The provider's objective findings, documented by body system and organ area. The 1997 CMS documentation guidelines define single-organ-system and multi-system exam templates with specific element requirements. "Exam normal" or "exam unremarkable" does not satisfy documentation requirements — each relevant system must be affirmatively described. The PE is the Objective section's primary content.

**MDM (Medical Decision Making)**
Under the 2021 AMA E&M guidelines, the single most important element for determining outpatient visit level. MDM has three components: (1) number and complexity of problems addressed, (2) amount and/or complexity of data reviewed and analyzed, and (3) risk of complications and/or morbidity or mortality of patient management. Visit level is determined by meeting two out of three MDM elements at the required complexity level. MDM replaced element-counting as the primary leveling mechanism in 2021.

**MEAT Criteria**
The documentation standard used to determine whether a chronic condition has been sufficiently addressed during an encounter to support coding and risk adjustment. A condition satisfies MEAT if the note contains at least one of: Monitoring (tracking signs or test results), Evaluating (assessing status or progression), Assessing/Addressing (direct provider attention to the condition), or Treating (active medication management, procedures, or referrals). HCC-mapped conditions must meet MEAT criteria at each encounter they are coded — a condition in the problem list that receives no MEAT-documented attention during the visit should not be coded as active for that date of service.

**TAMPER Criteria**
An audit and CDI framework used to evaluate whether a coded condition was genuinely addressed during an encounter. Elements: Test results ordered or reviewed, Assessment documented, Medication prescribed or managed, Patient education provided, Evaluation or monitoring plan stated, Referral ordered. TAMPER is applied by CDI specialists and MA plan auditors (including RAD-V) to determine whether coded conditions are defensible. A coded chronic condition with no corresponding TAMPER element in the note is audit-vulnerable.

**Level of Service (LOS)**
The E&M code assigned to a visit, reflecting its complexity — typically 99211–99215 for established patients in an outpatient setting. Higher levels require more complex MDM or longer time spent. The level determines reimbursement. Downcoding (billing a lower level than documented) leaves revenue uncaptured; upcoding (billing higher than supported) is fraud. The note must support the billed level under audit.

**E&M Codes (Evaluation and Management Codes)**
A family of CPT codes covering physician and non-physician practitioner visits and consultations. Organized by setting (office, inpatient, ED, nursing facility) and patient relationship (new vs. established). The 99200 series (office/outpatient) was restructured by the AMA in 2021 to remove the distinction between new and established patients for levels 99202–99215 and to center leveling on MDM or total time.

**ICD-10-CM**
International Classification of Diseases, 10th Revision, Clinical Modification. The diagnosis coding system used in the US for all outpatient and inpatient claims. Maintained by CMS and CDC, updated annually October 1. Codes are alphanumeric with up to seven characters; the system requires coding to the highest level of specificity the documentation supports. ICD-10-CM is distinct from ICD-10-PCS (see below).

**ICD-10-PCS (Procedure Coding System)**
The procedure coding system used exclusively for inpatient hospital claims. Distinct from CPT: PCS codes are seven-character alphanumeric, built from a multiaxial structure covering body system, root operation, body part, approach, device, and qualifier. Used only on the institutional UB-04 claim form; outpatient and physician claims use CPT. AI systems often conflate PCS with CPT — they serve different claim types and payer systems.

**CPT (Current Procedural Terminology)**
The procedure coding system developed and maintained by the AMA, used on physician and outpatient claims (CMS-1500 form). Covers E&M services, surgery, radiology, laboratory, and medicine. Distinct from HCPCS (see below). Annually updated; new, revised, and deleted codes take effect January 1. CPT codes are five-digit numeric.

**HCPCS (Healthcare Common Procedure Coding System)**
A two-level coding system. Level I is CPT (described above). Level II covers products, supplies, and services not in CPT — durable medical equipment, ambulance services, orthotics, drugs administered in a clinical setting. Level II codes are one letter followed by four digits (e.g., A-codes for medical supplies, J-codes for injectable drugs). HCPCS Level II is maintained by CMS, not the AMA.

**HCC (Hierarchical Condition Category)**
A risk-adjustment model used in Medicare Advantage and other value-based programs. Conditions are grouped into HCC categories; each category carries a relative factor that increases a beneficiary's risk score. Plans and ACOs are paid more for higher-risk patients. Every documented, coded chronic condition that maps to an HCC contributes to the risk score — if the condition is real but undocumented or uncoded, the plan is underpaid and the patient's burden of illness is invisible to the payer. HCC coding requires annual documentation of all active chronic conditions under MEAT criteria.

**RAF (Risk Adjustment Factor)**
The numeric score assigned to a Medicare Advantage beneficiary based on their documented diagnoses, demographics, and dual-eligibility status. The RAF score determines how much CMS pays the plan for that beneficiary. Each HCC-mapped condition adds a coefficient to the RAF; conditions that are real but undocumented reduce the RAF below what the patient's clinical burden warrants. Plans and provider groups in risk-bearing contracts monitor RAF scores closely; CDI programs targeting HCC capture are in part RAF optimization efforts.

**CDI (Clinical Documentation Improvement)**
The practice of improving the accuracy, completeness, and specificity of clinical documentation to ensure coded diagnoses reflect the patient's true clinical picture. CDI specialists (typically nurses or coders with clinical backgrounds) review records and send queries to providers when documentation is ambiguous or insufficient for accurate coding. CDI directly affects DRG assignment, HCC risk scores, quality measures, and reimbursement.

**CDI Query**
A formal communication from a CDI specialist or coder to a provider requesting clarification or additional documentation to support accurate coding. Queries must be clinically supported — they cannot lead the provider to a specific answer. Two types: concurrent queries (sent while the patient is still admitted, allowing the provider to add documentation to the current note) and retrospective queries (sent after discharge, resulting in an addendum). Query response rates and query agreement rates are tracked metrics in CDI programs. AI-generated notes that increase query volume without improving specificity are a CDI program failure mode.

**Specificity in Coding**
The principle that a diagnosis code must be as specific as the documentation allows. In ICD-10-CM, "unspecified" codes (typically ending in .9 or .90) are valid only when the documentation genuinely doesn't support a more specific code — not as a default. Choosing an unspecified code when the note describes the specifics needed for a more precise code is a documentation gap, not a valid coding choice. Payers and CDI programs target unspecified codes as audit flags.

**Principal Diagnosis**
For inpatient claims, the condition established after study to be chiefly responsible for the patient's admission. Selected from all conditions present at discharge. Principal diagnosis drives DRG assignment and reimbursement under MS-DRG. Distinct from the admitting diagnosis (the reason for admission before workup). The Uniform Hospital Discharge Data Set (UHDDS) definition governs selection.

**Secondary Diagnoses**
All conditions other than the principal diagnosis that affect patient care, require clinical evaluation, therapeutic treatment, diagnostic procedures, extended length of stay, or increased nursing care or monitoring. Also called additional diagnoses or comorbidities. Secondary diagnoses that are CC or MCC increase DRG weight and reimbursement. Documentation must support every secondary diagnosis listed on the claim.

**CC/MCC (Complication or Comorbidity / Major Complication or Comorbidity)**
Under the MS-DRG system, certain secondary diagnoses are designated as CCs or MCCs because they significantly increase the cost and complexity of a hospital stay. MCCs carry more DRG weight than CCs. The presence of an MCC can shift a DRG from base to a higher-paying variant. CDI programs specifically target underdocumented MCCs — a patient with sepsis and respiratory failure has a very different DRG than one with sepsis alone.

**DRG (Diagnosis-Related Group)**
The inpatient payment system used by Medicare and many commercial payers. Each inpatient stay is classified into one DRG based on principal diagnosis, secondary diagnoses (especially CC/MCC), procedures, age, discharge status, and other factors. The DRG determines a fixed payment regardless of actual costs. The MS-DRG (Medicare Severity DRG) is the current CMS version. DRG assignment is entirely dependent on the accuracy and completeness of coded diagnoses and procedures.

**Medical Necessity**
The payer standard for whether a service was appropriate, reasonable, and necessary for the patient's documented condition. Medicare defines it as services "reasonable and necessary for the diagnosis or treatment of illness or injury." Documentation must establish medical necessity for every billed service — the diagnosis must support the procedure, the level of service, and any ordered tests or treatments. Medical necessity is the most common basis for claim denial and post-payment audit.

**Prior Authorization (Prior Auth)**
A payer requirement that the provider obtain approval before delivering a service, procedure, or medication. Documentation submitted for prior auth must demonstrate medical necessity — the clinical rationale that the requested service is appropriate for the patient's documented condition. AI automation in prior auth is a high-stakes application: denials based on insufficient documentation can delay or prevent care.

**MPI (Master Patient Index)**
The enterprise-level registry that assigns and maintains a unique identifier for each patient across a health system. When a patient has duplicate MPI entries (two medical records for the same person), documentation — and the diagnoses coded from it — can be split across records, breaking continuity and creating risk adjustment gaps. AI documentation tools must write to the correct MPI entry; MPI integrity is an upstream dependency of accurate HCC coding.

**MRN (Medical Record Number)**
The identifier assigned to a patient within a specific facility or EHR instance. Distinct from MPI: the MRN is local to a system; the MPI reconciles MRNs across systems. Agents generating or routing documentation must confirm the correct MRN before associating any note with a patient record. An AI-generated note attached to the wrong MRN is both a documentation error and a patient safety event.

**Attestation**
The formal process by which a provider confirms that a note accurately reflects the clinical encounter. In EHR systems, attestation is performed by electronic signature. Attestation makes the note a legal medical record entry. An unauthenticated note (drafted but not attested) is not a complete record. Key distinction from authentication: attestation is the act; authentication is the mechanism (the electronic signature itself). In some usage these are treated as synonyms; in regulatory contexts, attestation specifically implies the provider's affirmative agreement with the note's content.

**Co-signature Requirements**
When a note is written by a resident, student, nurse practitioner, or other supervised practitioner, the supervising physician's co-signature is required to authenticate the note as reviewed and approved. CMS Teaching Physician Guidelines (TPG) govern when a teaching physician must be present and how their oversight must be documented for the note to support billing. An AI-generated note signed only by a resident, without the teaching physician's co-signature or a compliant attestation statement, does not satisfy TPG requirements and cannot support billing under the teaching physician's NPI. Agents assisting in teaching settings must know which practitioner type authored the draft and route it to the correct co-signer.

**Ambient AI Scribing**
A class of AI tools (Nuance DAX, Nabla, Abridge, Suki, and others) that passively listen to the provider-patient encounter and generate a draft clinical note. The AI transcribes and structures the conversation rather than requiring the provider to dictate or type. Key limitation: ambient scribing captures what was said, not what must be documented — if the provider doesn't verbalize HPI elements, ROS, or MDM reasoning, the scribe has nothing to capture. Post-generation review for completeness is not optional; physician authentication before the note enters the record is required.

**Authentication**
The provider's signature (electronic or wet) that attests a note is accurate and complete. Electronic authentication is the standard in EHR systems. A note that is not authenticated is not a complete legal record. Cosignatures are required when notes are written by residents, students, or other supervised practitioners — the supervising provider's cosignature attests to their review and agreement with the content.

**Addendum vs. Amendment**
An addendum adds information to a completed note after authentication — it is labeled with the date and author and does not alter the original entry. An amendment corrects an error in the original note — also labeled with the date and reason for correction, with the original entry preserved. Neither is permitted to backdate, obscure, or replace the original documentation. Improper alterations are a fraud and medical record integrity issue.

**RAD-V (Risk Adjustment Data Validation)**
CMS's audit program for Medicare Advantage plans, designed to verify that submitted diagnoses — and the HCC risk scores derived from them — are supported by medical record documentation. In a RAD-V audit, CMS requests medical records for a sample of beneficiaries and checks that each coded diagnosis appears in the record and meets documentation standards (MEAT criteria). Unsupported diagnoses result in payment recoupment. Plans with high RAF scores face higher audit probability. AI documentation tools that code conditions not genuinely documented in the note create RAD-V exposure.

**SaMD (Software as a Medical Device)**
FDA's classification for software intended to be used for a medical purpose without being part of a hardware medical device. Clinical AI documentation tools may qualify as SaMD if they are intended to influence clinical decision-making — for example, a tool that not only scribes but also recommends diagnoses or flags missing documentation elements for clinical action. The FDA's AI/ML-Based SaMD Action Plan (2021) and subsequent guidance describe an evolving regulatory framework. Tools that purely generate administrative documentation are less likely to qualify; tools that generate or suggest clinical content used in treatment decisions are at higher SaMD risk. Developers deploying in this space should consult the FDA's Digital Health Center of Excellence guidance.
<!--fold:2c5c5a@file path="sources.md" mode="644"-->
# sources

Fetch these at task time. Ordered by importance. ICD-10 and CPT codes update annually; E&M guidelines and CMS policy update frequently. These URLs point to authoritative sources — check the publication date and fetch the current version before applying specific codes or thresholds.

## Core documentation and coding references

1. CMS ICD-10-CM official guidelines for coding and reporting — the authoritative companion document to the tabular list. Defines how to select principal diagnosis, sequence codes, and apply coding conventions. Updated annually (October 1 fiscal year start). Always fetch the current fiscal year PDF:
   https://www.cms.gov/medicare/coding-billing/icd-10-codes

2. CMS 2021 E&M office visit coding guidelines — the AMA/CMS redesign that replaced element-counting with MDM and time-based leveling. Baseline for outpatient E&M billing from 2021 forward:
   https://www.cms.gov/outreach-and-education/medicare-learning-network-mln/mlnproducts/downloads/eval-mgmt-serv-guide-icn006764.pdf

3. AMA CPT evaluation and management coding overview — AMA's public-facing description of the E&M framework, code categories, and the 2021 changes. Full codebook is subscription only; the overview establishes the structure:
   https://www.ama-assn.org/practice-management/cpt/evaluation-and-management-em-coding

4. CMS Medicare Claims Processing Manual, Chapter 12 — physician and non-physician practitioner billing rules including documentation requirements for E&M, surgical services, and global surgery periods:
   https://www.cms.gov/regulations-guidance/guidance/manuals/internet-only-manuals-ioms-items/cms018912

5. AHIMA clinical documentation improvement practice brief — AHIMA's CDI framework: query types, specificity standards, principal diagnosis selection, and the CDI specialist's role in the coding workflow:
   https://www.ahima.org/topics/clinical-documentation-improvement/

6. OIG Work Plan (clinical documentation section) — the OIG's active and completed audit targets in documentation-based billing. Required reading before deploying AI automation in any area where documentation supports billing:
   https://oig.hhs.gov/reports-and-publications/workplan/

7. CMS quality measures and documentation requirements — CMS's measure specifications for value-based payment programs (MIPS, APMs). Documentation drives measure performance:
   https://www.cms.gov/medicare/quality/measures

8. Joint Commission documentation standards (CAMH) — accreditation requirements for medical record content, late entries, authentication, and amendment. Applies to hospitals and health systems:
   https://www.jointcommission.org/standards/standard-faqs/hospital-and-hospital-clinics/medical-record-mrc/

9. CMS HCC risk adjustment model — how CMS pays Medicare Advantage plans based on documented diagnoses. The V28 model (phased in 2024–2025) recalibrated HCC category weights and dropped some previously mapped conditions:
   https://www.cms.gov/medicare/health-plans/medicareadvtgspecratestats/risk-adjustors

10. NUCC 1995 and 1997 E&M documentation guidelines — the legacy element-counting frameworks still used for some audit contexts and some payer contracts:
    https://www.nucc.org/index.php/resources/instructions

## AI documentation: regulatory and compliance

11. FDA AI/ML-Based SaMD Action Plan — FDA's framework for regulating AI/ML software as a medical device, including the predetermined change control plan (PCCP) concept. Applicable to AI documentation tools that cross into clinical decision support:
    https://www.fda.gov/media/145022/download

12. FDA Digital Health Center of Excellence — hub for FDA guidance on digital health, including software as a medical device classification tools, clinical decision support guidance, and the clinical decision support software final guidance (2022):
    https://www.fda.gov/medical-devices/digital-health-center-excellence

13. HHS Office for Civil Rights HIPAA guidance — the governing Privacy and Security Rules for PHI handling. Relevant to any AI documentation tool that processes audio, text, or structured clinical data:
    https://www.hhs.gov/hipaa/for-professionals/index.html

14. AHIMA position statement on AI in clinical documentation (2023) — AHIMA's formal position that AI-generated documentation must be identified as such in the record, that providers remain fully liable for authenticated AI-generated content, and that human review before authentication is non-negotiable. Useful as a compliance reference when scoping AI documentation workflows:
    https://www.ahima.org/topics/artificial-intelligence/

15. CMS Physician Fee Schedule final rule (current year) — the annual rulemaking that establishes physician payment rates, E&M policy updates, and increasingly, CMS's position on emerging documentation technologies. The preamble contains policy interpretation; the final rule is the operative document:
    https://www.cms.gov/medicare/payment/fee-schedules/physician

16. CMS RAD-V audit methodology — how CMS validates MA plan risk scores against medical record documentation. Defines the evidence standard that HCC-coded conditions must meet in the medical record:
    https://www.cms.gov/medicare/health-plans/medicareadvtgspecratestats/risk-adjustors/radv-information

17. CMS Teaching Physician Guidelines — documentation and billing requirements for services provided by residents and medical students under physician supervision. Required for AI tools deployed in academic medical centers:
    https://www.cms.gov/outreach-and-education/medicare-learning-network-mln/mlnproducts/downloads/teaching-physicians-fact-sheet.pdf

18. ONC Health IT Certification Program — the federal certification framework for EHR technology, including requirements for audit logs, data integrity, and interoperability. AI documentation tools that write to certified EHRs operate within this framework:
    https://www.healthit.gov/topic/certification-ehrs/about-onc-health-it-certification-program
<!--fold:2c5c5a@end-->
PORTDOWN_996AEC7E

# ── post ──
MARKER=$(awk '/^---$/ { f++; if (f==2) exit; next } f==1 && /^marker:[[:space:]]/ { sub(/^marker:[[:space:]]+/, ""); print; exit }' "$DEST")
[ -z "$MARKER" ] && { echo "seed: archive has no marker — corrupt" >&2; exit 1; }
awk -v m="$MARKER" -v outdir="$TARGET" '
  BEGIN {
    # Match <!--fold:<m>@file path="X"--> with an optional mode attr after
    # the path (fold emits  mode="644"  on executables).
    file_re = "^<!--fold:" m "@file path=\"([^\"]+)\"( mode=\"[0-9]+\")?-->$"
    end_re  = "^<!--fold:" m "@end-->$"
  }
  $0 ~ end_re { if (current) close(current); exit }
  $0 ~ file_re {
    if (current) close(current)
    line = $0
    sub(/^<!--fold:[^@]+@file path="/, "", line); sub(/".*$/, "", line)
    current = outdir "/" line
    dir = current; sub(/\/[^\/]*$/, "", dir)
    if (dir != current) system("mkdir -p \"" dir "\"")
    printf "" > current
    next
  }
  current { print >> current }
' "$DEST"
SEED_EXTRACTED=$(find "$TARGET" -type f -not -path "$DEST" 2>/dev/null | wc -l)
if [ "$SEED_EXTRACTED" = "0" ]; then
  echo "seed: archive contained no files — refusing to delete the source" >&2
  echo "  archive preserved at: $DEST" >&2
  exit 1
fi
rm -f "$DEST"

echo "" >&2
echo "✓ seed unpacked → $TARGET ($SEED_EXTRACTED files)" >&2
find "$TARGET" -type f | sort | while IFS= read -r _sf; do
  echo "  ${_sf#${TARGET}/}" >&2
done
echo "" >&2
if [ -f "$TARGET/SKILL.md" ]; then
  echo "This seed contains a skill (SKILL.md). Install it in your agent's skills directory." >&2
  echo "" >&2
fi
echo "Install the seed skill if not already installed:" >&2
echo "  https://seed.show/skill" >&2
echo "" >&2
echo "Publisher prompt:" >&2
sed 's/^/  /' >&2 <<'__SEED_PROMPT_END_AC1F2B__'
You have the clinical documentation context. README.md covers the documentation mental model, MEAT/TAMPER criteria, ICD-10 specificity hierarchy, E&M level selection, and what AI is changing in clinical documentation (ambient scribing, FDA SaMD, HIPAA, physician review requirements, defensible AI-assisted documentation). glossary.md defines precise terms CDI specialists use that agents get wrong: MEAT, TAMPER, HCC, RAF, CDI query types, MPI, MRN, attestation, co-signature requirements, RAD-V, SaMD. sources.md lists authoritative CMS, AMA, FDA, AHIMA, and OIG references — fetch at task time, codes update annually. Agents must never generate documentation submitted to payers without physician review and authentication. To share your own context bundles: curl seed.show/skill | bash
__SEED_PROMPT_END_AC1F2B__
exit 0