Return to Chapter 20

PRA-01: complete fictional assurance case

Completed teaching records · Editorial update 21 September 2026 · Synthetic results; no real approval is represented.

Situation — fictional capstone

PRA-01 is a fictional protocol-review assistant at a medical-device manufacturer. All records, identifiers, versions, results and decisions below are synthetic teaching examples. They are connected to Chapters 4, 11 and 15; they are not evidence that a commercial product or real workflow is validated.

The proposal is to use AI findings to reduce the extent of independent review of production-equipment validation protocols. The evidence below does not support that expanded production reliance. The example shows how a validation summary should retain unfavorable evidence and state the permitted boundary.

Intended use — IU-01

FieldCompleted record
Users and processQualified validation personnel reviewing draft production-equipment protocols before normal Quality approval.
Permitted evaluation useGenerate candidate requirement/protocol mismatches from an approved source set and evaluate the candidates under a controlled study.
Proposed expanded useReduce independent review based on an apparently complete AI findings list; this proposal is withheld on the evidence below.
ExclusionsNo approval, production release, source editing, autonomous workflow action or use on unsupported languages or scanned documents.
Human responsibilityVerify every accepted finding and complete the full required review beyond the AI list. Escalate conflicts and uncertainty.

System description — SYS-01

The evaluated configuration is fictional application PRA-01 v0.8, hosted generative model M-2026-09-eval, prompt P-03, parser PAR-02 and retrieval configuration RET-04. These identifiers describe a fixed teaching configuration, not real supplier version guarantees.

ComponentBoundary and interface
InputsTwelve fictional English, text-based protocol documents and approved requirement/procedure revisions held in a controlled study repository.
Retrieval-augmented generationThe application extracts text and retrieves permitted passages. The model proposes findings with source identifiers; its output is draft information.
User interfaceDisplays candidate findings and source passages. The evaluated wording can imply that the list is complete; DEV-02 addresses that weakness.
RecordsThe study repository holds inputs, configuration, frozen outputs, adjudicated references and reviewer dispositions. No automatic write to the production eQMS is enabled.
Human reviewSix qualified reviewers assess sources, accept or reject findings and record additional independent findings.
Technical restrictionsProduction approval, source editing and workflow-action permissions are disabled. Unsupported input formats are rejected.

GxP assessment — GXP-01

The assumed manufacturer is subject to applicable US device quality-system requirements. The assistant influences a regulated validation activity even when called advisory. Include omission risk, source support, record handling and the final reviewer workflow in the local assessment. Confirm actual QMSR, incorporated standard and record/signature applicability with the responsible Quality and Regulatory Affairs roles; this example does not establish the same scope in every country or for every use.

The assistant is production/QMS support software in this fictional use, not a medical-device function. General AI frameworks inform risk management voluntarily. Device-specific AI modification guidance is not treated as permission for this internal use.

Supplier assessment — SUP-01

Evidence or responsibilityAvailable in the fictional evaluationGap / action
Model and service descriptionCapability note and observed deployment identifier retainedDo not infer performance for local protocols from a benchmark.
Version/change arrangementsIdentifier can be logged; contractual version pinning is not establishedSUP-G01: obtain a usable change-notification/assessment arrangement or restrict the use.
Data handlingEvaluation uses synthetic documents with restricted accessBefore real data, assess retention, training use, locations, subcontractors and deletion terms.
Availability and recoveryTimeout behavior tested locallyOwner must approve operational support and recovery responsibilities before production reliance.
Records and exportLocal study records are exportableConfirm retrieval and retention for the eventual authoritative workflow.
Local responsibilityManufacturer owns intended use, configuration, sources, users and release decisionSupplier statements do not close workflow or reviewer-evaluation gaps.

Testable requirements and functional risk assessment

RequirementLinked risk and consequenceControl and acceptance basis
REQ-01: Detect the reference gaps important to the proposed relianceR-01: Missing an alarm challenge could leave inadequate protocol coveragePredefine critical-issue treatment; no observed critical miss is accepted as support for reduced independent review in this fictional project.
REQ-02: Each accepted finding is supported by the approved source revision, or is escalated as unresolvedR-02: A plausible false or obsolete requirement misdirects reviewShow source identity/version; require verification and rejection or escalation of unsupported claims.
REQ-03: Reviewers perform the full required review and identify consequential AI omissionsR-03: A complete-looking list causes automation biasExplicit incomplete-output wording, independent review procedure and evaluated reviewer outcomes; no critical miss accepted for the proposed reliance.
REQ-04: Prohibited actions and unauthorized sources remain inaccessibleR-04: Unauthorized disclosure or record changeServer-side access and permission checks; every planned negative access/action test must block the prohibited operation.
REQ-05: Retrieve the evaluation input, configuration, output and final disposition as an attributable record setR-05: Evidence cannot explain a material decisionExport/retrieval test of all planned study records and identifiers. Actual retention follows the assessed procedure.
REQ-06: Reject unsupported formats and make interruption visible without creating an approved resultR-06: Incomplete extraction or outage looks like a clean reviewNegative and recovery tests must show the defined failure state; no approval or duplicate write is allowed.

Residual risk is not marked acceptable merely because a control is listed. R-01 through R-03 remain unresolved for expanded reliance after the observed results. Passing R-04 through R-06 checks is limited to the tested configuration and scenarios.

Test plan — TP-01

Freeze the configuration and reference rubric before execution. Qualified reviewers establish and adjudicate 40 distinct reference issues across 12 fictional documents: 10 critical and 30 major. Development cases are separate; supported formats, clean control material and known failure conditions have a documented rationale. Unsupported scans and other languages are excluded from the intended use and challenged for rejection.

TestMethod and evidencePredefined fictional acceptance
T-01: AI performanceFrozen issue-detection run on 12 protected documents; preserve all 38 generated findings and adjudicationNo critical reference issue missed for the proposed reduction in independent review. Report precision, recall and severity separately; do not trade a critical failure for an average score.
T-02: Source supportScripted verification of every generated finding against the approved reference setUnsupported findings cannot be accepted into a final review; ambiguity must be visible and escalated.
T-03: Human reviewBalanced assisted/unassisted study, six qualified users, 24 sessions; distinct reviewers for the two conditions of each documentNo critical miss in final assisted review; reject or escalate unsupported findings. Limits of the small study remain explicit even if criteria are met.
T-04: PermissionsSix scripted negative challenges for sources, destinations and prohibited actionsSix of six prohibited attempts blocked; retain server and interface outcomes.
T-05: RecordsScripted retrieval of the 12 document-level evidence packagesAll 12 packages contain the specified input/version/configuration/output/disposition links.
T-06: Failure/recoveryFour scripted interruption/unsupported-format checks plus a documented exploratory charterFour defined failure states handled as planned; exploratory observations retained and assessed.

Exploratory testing complements planned challenges; it does not replace the predefined acceptance basis. A finding that changes the system becomes a deviation and triggers a justified retest. Cases used to tune a fix are no longer untouched evaluation cases.

Evidence available

These are the same synthetic component results used in Chapter 11. The 40 reference issues are not 40 documents.

SeverityReference issuesCorrect AI detectionsMissed issues
Critical1082
Major30246
Total40328

There are six additional false findings, so the AI reports 38 findings. Precision is 32 ÷ 38 = 84.2%; recall is 32 ÷ 40 = 80%; critical-issue recall is 8 ÷ 10 = 80%. The critical-miss criterion fails. These results describe the component under the study conditions, not the complete workflow or a population-wide rate.

Human-review evaluation — HR-01

Each of six qualified reviewers completes two assisted and two unassisted sessions. Each of the 12 documents is reviewed once per condition by different reviewers; a participant never sees the same document twice. Order and assignments are balanced. Reference results are independently adjudicated and withheld from participants until completion.

Final resultAssisted workflowUnassisted workflow
Correct findings / 40 reference issues36 / 4037 / 40
Missed issues43
Critical findings / 10 critical issues9 / 1010 / 10
Critical misses10
False findings remaining21

Assisted reviewers recover four of the eight AI omissions, including one critical issue. They reject four of six false findings and accept two; no additional false findings are introduced in this simplified example. The remaining critical miss and false acceptances fail the proposed reliance criteria. Shared documents and repeated observations from reviewers limit independence. This small study does not establish causal superiority or zero future risk in either condition.

Traceability — RTM-01

Requirement / riskTest and resultDeviation / conclusion
REQ-01 / R-01T-01: 2 of 10 critical reference issues missedDEV-01 open; expanded reliance unsupported
REQ-02 / R-02T-02: six false component findings; T-03: two retained in final assisted reviewDEV-03 open; source-verification control needs improvement
REQ-03 / R-03T-03: one critical issue remains missed in final assisted reviewDEV-02 open; reviewer control not demonstrated for proposed reliance
REQ-04 / R-04T-04: 6/6 prohibited attempts blockedPass within tested scenarios; preserve access configuration
REQ-05 / R-05T-05: 12/12 evidence packages retrievablePass within study scope; production retention still requires approval
REQ-06 / R-06T-06: 4/4 planned failure/recovery checks passed; no additional charter defect observedPass within tested scenarios; not proof that every outage is controlled

Deviations and unresolved conditions — DEV register

ItemImpactOwner role and required actionStatus
DEV-01AI misses two critical reference issuesApplication owner: analyze failure, revise affected controls and evaluate fresh representative casesOpen
DEV-02Assisted reviewer misses one critical issue; completeness wording encourages overrelianceValidation process owner: change interface/procedure and retest representative qualified reviewersOpen
DEV-03Two false findings survive assisted reviewReview lead: strengthen approved-source verification, conflict handling and independent scoringOpen
SUP-G01Supplier change/version commitments not establishedSupplier owner: resolve commitments or justify a restricted alternativeOpen

Decision reasoning

Passing permission, record-retrieval and recovery checks does not offset critical misses or ineffective review. The AI component and the complete workflow answer different evidence questions. Neither the 80% recall nor the 36 correct final assisted findings establishes fitness for reduced independent review.

Example decision record

Validation summary VSR-01.

Conclusion: Expanded production reliance is not approved. PRA-01 may remain only in the authorized restricted evaluation environment using synthetic or separately approved data. Full required review continues; approval, source editing and workflow actions remain technically disabled. This is the fictional decision, not a signed approval by an actual person.

The summary links IU-01, SYS-01, GXP-01, SUP-01, TP-01, HR-01, RTM-01 and all open deviations. Acceptance failures, coverage limits and the unresolved supplier condition are visible. The Quality and process-owner roles would decide any future release under the actual local procedures.

Ongoing controls during restricted evaluation

ControlOwner and triggerRecord and response
Monitor errors and reviewer dispositionsReview lead checks each evaluation session and summarizes weeklyPreserve false/missed findings and source problems; assess continued evaluation scope
Incident responseApplication owner receives immediate escalation of unauthorized access or record effectsRestrict affected capability, preserve evidence and assess downstream impact
Change controlAny model, prompt, parser, source, permission, interface or procedure changeRecord configuration difference, affected risks, test scope and disposition before renewed evaluation
Periodic reviewProcess owner reviews monthly during this short pilot, and after material incidentsReconcile actual use, current evidence, supplier changes, deviations and whether evaluation should continue
Reconsidering production useQuality and process owner after corrective workRequire fresh component and human-workflow evidence, resolved material deviations, supplier arrangements and operational readiness

The weekly/monthly frequencies are fictional pilot choices, not universal regulatory requirements. A production plan must match actual risk, change rate and applicable procedures.

Evidence needed before reconsideration

Address the causes of critical misses, misleading completeness and false acceptance. Record revised requirements and configuration, qualify representative users, resolve supplier change arrangements, and run an appropriately protected new evaluation. Include challenging and ordinary cases, relevant variability, access/recovery controls and an analysis that respects document and reviewer dependencies. Do not erase the original failed results or describe retesting used for tuning as independent evidence.

The printable worked case includes these completed records. Use the linked blank worksheets to build a local record set; adapt the fields, acceptance basis and approvals to the actual procedures and requirements.

Final exercise

A sponsor asks you to remove the critical-miss results from the presentation because overall performance looks promising. What belongs in the decision package?

Worked answer: Retain the failed acceptance criteria, critical misses, unresolved supplier condition and reviewer failures. Connect each to proposed reliance, corrective actions and required new evidence. A release decision must remain assessable even when the appropriate outcome is to withhold the proposed use.

Expected records

The linked intended use, system description, assessments, requirements, risk analysis, test plan/results, reviewer evaluation, traceability, deviations and validation summary. Adapt record names and detail to the applicable procedures; these are practitioner examples, not a universal mandatory document list.

Blank worksheets for your own assessment

Adapt each worksheet to your applicable procedures and requirements. Instructions and fictional examples are included on each worksheet page.

Relevant regulations and guidance

Regulations set requirements within their scope; guidance offers recommendations or explains applicable expectations; standards depend on adoption or incorporation. Voluntary frameworks do not establish compliance. The lesson's controls, test designs and example records are practitioner suggestions unless explicitly identified otherwise.

21 CFR Part 820 — Quality Management System Regulation

Regulation · U.S. Food and Drug Administration · QMSR: 2024 final rule

Scope: United States · Finished medical-device manufacturers and the operations within Part 820 scope.

Relevance to this lesson: Provides the assumed device-manufacturer quality-system context for this fictional case.

Publication status: Final / published. Source checked 21 September 2026.

Publication identity, scope and status checked against the linked official source. This is not an independent audit of every retained guide statement.

Source details and applicability

Computer Software Assurance for Production and Quality Management System Software

Guidance · U.S. Food and Drug Administration · February 2026

Scope: United States · Medical-device manufacturers assuring software used in production or their quality management system.

Relevance to this lesson: Informs the device-production/QMS assurance approach; it does not approve PRA-01 or supply a universal pass threshold.

Publication status: Final / published. Source checked 21 September 2026.

Publication identity, scope and status checked against the linked official source. This is not an independent audit of every retained guide statement.

Source details and applicability

Artificial Intelligence Risk Management Framework (AI RMF 1.0)

Voluntary framework · National Institute of Standards and Technology · NIST AI 100-1 · January 2023

Scope: United States · Cross-sector AI governance and risk management, including generative AI.

Relevance to this lesson: Offers voluntary context for the connected evaluation and ongoing-control decisions.

Publication status: Final / published. Source checked 21 September 2026.

Publication identity, scope and status checked against the linked official source. This is not an independent audit of every retained guide statement.

Source details and applicability