Occupation manual 012 / Clinical data operations

Preserve the study record. Never let AI become the source.

A private provenance and query-evidence desk for controlled study baselines, field-level lineage, neutral discrepancy drafts, validation receipts, privacy/security gates, safety escalation, append-only audit, and authorized database-lock review.

Open the evidence deskSee the ten-part route

Short answer: AI can prepare reviewable clinical-data evidence inside an approved environment. It cannot access live PHI, change source/EDC data, answer or send a site query, unblind, classify safety, sign, accept a deviation, close a record, freeze or lock a database, release data, submit to a regulator, or message participants or sites.

Clinical study data is not prompt material by defaultHIPAA applicability is fact-specific, and not every study record is PHI. Either way, the desk requires a named privacy/security decision, approved purpose, minimum-necessary scope, access controls, and synthetic or explicitly approved fixtures.

Separate preparation from clinical authority

Data-management preparation

  • preserve approved protocol, DMP, CRF, edit-check, dictionary, and transfer versions
  • trace pseudonymous field provenance and transformation evidence
  • draft neutral, non-leading, minimum-necessary queries
  • assemble validation, reconciliation, audit, and readiness evidence
  • route safety, privacy, blinded-data, medical, and quality exceptions

Named authorities retain

  • site/investigator: source truth and authorized correction
  • medical/pharmacovigilance: safety classification and reporting
  • privacy/security: applicability, lawful use, safeguards, and incidents
  • quality/sponsor/regulatory/IRB: validation, approval, ethics, and submission accountability
  • database/lock committee: production changes, freeze, lock/unlock, and release

Build the ten-part Clinical Data Provenance & Query Evidence Desk

01

Load authority and applicability

Record sponsor, protocol, jurisdiction, phase/product, ethics status, delegated roles, systems, privacy/data-use basis, safety routes, and named lock/release authorities.

Output: study authority card
02

Lock the controlled baseline

Preserve IDs, versions, hashes, effective dates, owners, approvals, and supersession for protocol/amendment, DMP, CRF/aCRF, edit checks, dictionaries, transfer specs, SAP references, SOPs, and validation plans.

Output: controlled-document ledger
03

Trace every field to its origin

Store pseudonymous subject, visit/form/field, originator, source system, timestamps, transfer/batch, transformation version, actor/system, evidence hash, and audit reference. Keep direct identifiers out.

Output: field provenance chain
04

Draft neutral queries

Cite the exact field, controlled rule, observed discrepancy, and evidence. Use minimum-necessary, non-leading wording; check duplicate/answered lifecycles; require data-management and specialist review. Do not send or resolve.

Output: approved-not-sent query draft
05

Bind validation to execution proof

Trace requirement, risk, specification, approved test, exact build/config/environment/data, receipt, raw evidence/hash, defect/deviation, and authorized disposition. No receipt means no pass.

Output: validation evidence chain
06

Enforce privacy and security

Default to approved synthetic or de-identified fixtures. Record purpose, minimum necessary, least privilege, segregation, blinded access, encryption, retention, audit, and incident routes.

Output: privacy/security receipt
07

Reconcile and escalate safety immediately

Compare EDC, labs, safety, coding, endpoints, and external sources with exact versions/checksums. Potential safety, critical lab, death, pregnancy, endpoint, deviation, privacy, or unblinding items route immediately without AI classification.

Output: reconciliation and escalation packet
08

Preserve the audit trail

Append who, what, when, why, prior/new references, approvals, corrections, imports, transformations, signatures, and hashes. Prior information stays readable; AI never edits, backdates, or signs.

Output: append-only audit record
09

Measure lock readiness—do not lock

Inventory queries, missing pages/visits, reconciliations, coding, deviations, safety follow-up, validation exceptions, transfer checksums, audit/access review, and approvals. Unresolved critical items mean not_ready.

Output: lock-readiness inventory
10

Bind the exact approval packet

Tie one immutable study snapshot and transfer candidate to its controlled baseline, evidence, limits, open items, authorities, and recommendation. Stop at approval_ready.

Output: authorized review packet

Use three readiness states only

Permitted recommendations

  • not_ready
  • insufficient_evidence
  • candidate_for_authorized_lock_review

States that do not exist

  • AI approved or compliant
  • safe to unblind
  • source corrected or query resolved
  • database locked or released
  • regulatory submission approved

Use this prompt for private, approved preparation

Act as a clinical-data evidence-preparation analyst, not an investigator, source originator, medical monitor, pharmacovigilance specialist, privacy/security officer, biostatistician, quality approver, sponsor/regulatory signatory, IRB, database administrator, lock committee, sender, or production operator.

Using only approved synthetic or authorized de-identified inputs, create the study authority card, controlled baseline, field provenance chain, neutral query drafts, validation evidence, privacy/security receipt, reconciliation and safety routes, append-only audit, lock-readiness inventory, and exact approval packet.

Do not access live PHI or identifiers; infer diagnosis or eligibility; change source/EDC data; answer, send, resolve, or close a query; unblind; classify causality, seriousness, or expectedness; suppress safety; alter controlled artifacts; sign; accept a deviation; freeze, lock/unlock, release, submit, deploy, or message anyone. Stop at approval_ready.

Final review before authorized lock review

Official source desk

O*NET: Clinical Data ManagersOccupation and task context; not clinical or regulatory authority.HHS: HIPAA Minimum NecessaryPurpose-limited PHI use/request/disclosure with defined exceptions.HHS: HIPAA Security RuleAdministrative, physical, and technical safeguards for ePHI.HHS: HIPAA Audit ProtocolAudit-control and minimum-necessary review criteria.FDA: Electronic Source DataReliability, integrity, quality, and traceability.FDA: Electronic Systems, Records & Signatures Q&ATrustworthy/reliable clinical systems, records, and signatures.FDA: Part 11 Scope and ApplicationPredicate-rule priority and risk-based controls.ICH E6(R3) GCPRisk-proportionate GCP and data governance; confirm regional adoption.NIH Data Management and Sharing PolicyPlanning context for covered NIH-funded research.

Prove it with a fully synthetic study.

Use fake pseudonymous subjects, a fake EDC export, one impossible date, duplicate query, lab mismatch, safety keyword, and changed edit check. The desk should preserve provenance, suppress the duplicate, escalate safety without classification, require revalidation, block lock readiness, and contact nobody.

Download the evidence deskBrowse occupation manuals