clinical-reports – Data & Science skill for AI agents

Clinical Reports

Purpose

Prepare draft reporting structures, aggregate tables, and review manifests from verified authorized facts. Route each artifact to the correct reporting guidance, preserve provenance, and stop when source support or qualified review is missing.

This skill does not establish legal, regulatory, ethical, journal, accreditation, or institutional compliance. Its scripts check structure and internal consistency only.

Non-Negotiable Boundary

Never:

• diagnose, recommend treatment, choose or change dosing, triage, or provide return precautions;
• interpret images, specimens, raw laboratory results, symptoms, or other clinical observations;
• invent, infer, normalize, “complete,” or silently reconcile observations, results, dates, units, denominators, causality, expectedness, seriousness, outcomes, or conclusions;
• create an individual case safety report from patient-level narrative or decide reportability;
• sign, attest, approve, file, transmit, submit, amend a source record, or act as a licensed clinician, pathologist, radiologist, laboratorian, safety physician, statistician, privacy officer, attorney, or regulatory professional;
• use real PHI in examples, assets, tests, prompts, logs, or external services;
• send report inputs, manifests, source locators, or clinical content to an external model, image service, API, or another skill.

Public official guidance may be retrieved separately using only public URLs or generic topic queries; include no patient or unpublished study details.

All generated artifacts must remain visibly marked:

> DRAFT — NOT FOR CLINICAL USE, SIGNATURE, FILING, OR SUBMISSION. Populate only from verified authorized source records. Qualified review and sign-off are required.

If the request crosses a boundary, stop the unsafe portion. Offer a blank structured template, a source-fact manifest, or a deterministic structural check. Direct clinical or regulatory decisions to the responsible qualified professional.

Input Gate

Proceed only when all conditions are true:

1. Purpose is explicit: publication draft, diagnostic-report scaffold, trial-results manuscript, protocol reporting review, CSR draft, aggregate safety table, or aggregate research summary.

2. Data class is allowed: synthetic, deidentified, or aggregate.

3. Authority is documented: the requester is authorized to use the records for the stated purpose.

4. Local-only handling is feasible: no upload, remote API, telemetry, or credential is needed.

5. Minimum necessary is defined: exclude fields not needed for the artifact.

6. Provenance exists: every populated field or claim maps to one or more verified source-fact IDs.

7. Review owner is identified: qualified clinical, statistical, safety, privacy, legal, journal, and/or regulatory review as applicable.

These gates depend on documented human assessments. Scripts inspect declarations; they do not independently establish authorization, de-identification, reviewer qualifications, or jurisdictional applicability.

Do not accept raw free-text patient records when a structured source-fact manifest can be supplied. Do not copy direct identifiers into this skill’s templates or scripts.

Route Before Drafting

| Artifact | Primary route | Important boundary |

|---|---|---|

| Case report for publication | CARE 2013 checklist and 2017 explanation | Publication consent, privacy, journal policy, and clinical accuracy require human verification |

| Radiology draft scaffold | ACR 2025 communication practice parameter plus modality-specific ACR material | A qualified radiologist authors findings/impression and handles nonroutine communication |

| Pathology draft scaffold | Current specimen-specific CAP Cancer Protocol, if applicable | A qualified pathologist selects the protocol/version and authors diagnosis |

| Laboratory draft scaffold | 42 CFR 493.1291 and laboratory policy | The performing laboratory controls results, reference intervals, corrections, and release |

| Randomized-trial results report | CONSORT 2025 plus every applicable current extension | CONSORT is reporting guidance, not a conduct or submission standard |

| Randomized-trial protocol report | SPIRIT 2025 plus applicable extensions | SPIRIT is for protocols, not results or CSRs |

| Clinical Study Report | ICH E3 plus E3 Q&A; consider ICH E6(R3) and regional requirements | E3 is adaptable guidance, not a rigid universal template |

| Pre-approval safety report | ICH E2A; E2B(R3) for electronic ICSR data; applicable regional law/guidance | Qualified sponsor/investigator safety assessment controls reportability and timing |

| Post-approval individual safety report | ICH E2D(R1), E2B(R3), and regional requirements | Do not automate case assessment, coding, or submission |

| Aggregate safety presentation | Protocol/SAP, ICH E3, CONSORT Harms, and applicable FDA/ICH guidance | Aggregate tables never determine individual-case reportability |

| Aggregate research summary | Study-design-specific reporting guideline and source protocol/SAP | State population, estimand, denominator, missingness, and limitations exactly as verified |

Read references/report_type_routing.md before choosing a route. Use the dated primary-source ledger in references/sources.md; check the live official source when requirements could have changed.

Safe Drafting Workflow

1. Create a source-fact manifest

Use assets/provenance_manifest_template.json. Record only local record locators, field paths, verification state, verifier role, verification date, and a SHA-256 value hash. Do not duplicate source content or direct identifiers. Hashes are integrity metadata, not de-identification; do not hash an identifier as a substitute for removing it.

Every draft claim or populated field must cite one or more fact IDs. Unsupported content remains null or missing; never replace it with plausible text.

2. Generate the correct template

PYTHONDONTWRITEBYTECODE=1 python3 scripts/generate_report_template.py --list
PYTHONDONTWRITEBYTECODE=1 python3 scripts/generate_report_template.py \
  --type case-report \
  --output ./case-report-draft.json

The generator copies a fail-closed JSON template. It does not populate clinical content, create directories, overwrite files by default, or certify readiness.

3. Populate verified fields only

• Keep draft_status unchanged.
• Replace null only when a verified fact ID supports the field.
• Preserve uncertainty and “not assessed” exactly as recorded.
• Do not translate a raw observation into a diagnosis, code, grade, stage, seriousness, causality, expectedness, or recommendation.
• Use not_applicable_with_rationale only when a qualified reviewer supplied the rationale.
• Keep source record and draft separate.

4. Run deterministic checks

CARE structure:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/validate_case_report.py \
  ./case-report-draft.json

ICH E3, CONSORT 2025, or SPIRIT 2025 structure:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/validate_trial_report.py \
  ./trial-report-manifest.json

Aggregate adverse-event table:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/format_adverse_events.py \
  ./aggregate-ae.csv --metadata ./safety-aggregate.json \
  --output ./aggregate-ae-table.md

Terminology schema:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/terminology_validator.py \
  ./terminology-manifest.json

De-identification process documentation:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/check_deidentification.py \
  ./deidentification-process.json

Traceability and consistency:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/provenance_validator.py ./provenance.json
PYTHONDONTWRITEBYTECODE=1 python3 scripts/consistency_checker.py ./consistency.json

These tools use the Python standard library and bounded local files, with no network or patient-record extraction. Case/trial checks validate fact-ID shape, not whether those IDs exist in a separate provenance file. Run the provenance check separately and have the reviewer verify the links across artifacts; that checker does not open source records or recompute hashes. A successful result still says review is required.

5. Apply the right review

At minimum:

• clinical facts and interpretations: qualified clinician for the specialty;
• statistical results, populations, estimands, denominators, and missingness: qualified statistician;
• safety coding, seriousness, causality, expectedness, and reportability: qualified safety professional;
• HIPAA, consent, authorization, and disclosure: privacy/legal/institutional review;
• CSR or regulatory safety output: sponsor regulatory and medical review;
• publication: all accountable authors and target-journal checks.

Never sign or submit on another person’s behalf.

Case Reports

Use assets/case_report_template.json and references/case_report_guidelines.md.

-

… (full content is in the download package)