Independent and not affiliated with the FDA, MHRA, ISPE, PDA, or any agency. Get the appgoutham@madhadi.com
madhadi.comData Integrity & GxP Quality
Browse all topics → Articles Templates & Procedures Learning paths GlossaryScenariosToolsRegulatory ReferencesLearning PathsTopics About Start here
Record Plug-and-play starting point AI & Automation

Record: AI Prompt, Model Version, and Retrieval Corpus Change Control

A plug-and-play per-change control record for a validated AI tool: what changed in the prompt, model version, output schema, or retrieval corpus, the classification that sets re-validation depth, the regression evidence, rollback plan, QA approval before production, and the post-change monitoring commitment, with a filled specimen and the regulatory expectations this record is intended to help you address, each of which you should verify against the current published source before relying on it.

Document type: Record

Read and copy the template below into your own quality system. It is a generic starting point for your own internal use, provided as is, with no warranty; see the Terms and License. Adopting it does not by itself create compliance.

This is a ready-to-use record for a single change to an already validated AI tool: an edit to the prompt, a move to a different model version, a change to the output schema, or a refresh of the retrieval corpus. It exists because of a common finding in this area, and one worth stating plainly: a team validates version 1, and then quietly edits the prompt or swaps the underlying model without assessing what that change did to the behavior they validated. This record is the executed artifact that the version-control and re-evaluation policy in the LLM evaluation and test plan requires, and it feeds the change control process in SOP: change control for validated systems. Replace every <<FILL: ...>> placeholder with your own specifics, set your document numbers and dates, and route it through your normal document control, review, and approval. A worked filled specimen follows the template so you can see how a completed version reads. Verify each cited regulation against the current source before you rely on it, and treat AI-specific guidance as evolving rather than settled.

Control header

FieldEntry
Document titleAI Prompt, Model Version, and Retrieval Corpus Change Control Record
Record number<<FILL: e.g. AICC-2026-014>>
Linked change request<<FILL: CR number from your change control system, or "none, see section 9">>
AI tool name and ID<<FILL: TOOL NAME / SYSTEM ID>>
Validated state reference<<FILL: validation summary report number and version this change acts on>>
Document owner<<FILL: role, e.g. AI System Owner>>
Record raised by<<FILL: name, role>>
Date raised<<FILL: date>>
StatusDraft / Under assessment / Approved / Implemented / Closed / Rejected / Rolled back

1. Purpose of this record

This record documents one change to a validated AI tool and the assessment that establishes whether the tool remains fit for its approved intended use after the change. It captures what changed, who asked for it and why, how far the change reaches into validated behavior, the regression evidence that the acceptance criteria still hold, how to get back to the previous state, the Quality Assurance decision made before the change reached production, and what will be watched afterwards.

The record does not restate the governing policy. Risk tiering and the standing governance rules sit in Policy: AI governance for GxP systems. The intended-use boundary sits in Form: AI intended-use and use-pattern classification. The full validation method sits in Protocol: AI/ML system validation. This record is the per-event artifact those documents call for.

2. Scope

Raise this record for any change to a validated AI tool in the following categories, whether the change originates internally or arrives from the model vendor:

  1. Prompt text, system instruction, rubric, or in-context examples.
  2. Model identifier, model version, vendor, hosting region, or serving configuration.
  3. Output schema, field names, enumerated values, or the parsing and validation logic applied to model output.
  4. Retrieval corpus content, corpus version, chunking, embedding model, or retrieval parameters.
  5. Model parameters that alter output distribution, including sampling temperature, top-p, maximum output tokens, and stop sequences.

Changes to the surrounding deterministic code, infrastructure, or hosting that do not touch the five categories above go through your normal change control route and do not need this record, though they may still need an impact statement against this tool.

3. Field definitions

FieldFormatRequiredWhoWhen
Record numberControlled sequenceYesDocument control or system ownerOn raise
AI tool name and IDText matching the system inventoryYesRequesterOn raise
Validated state referenceReport number and versionYesSystem ownerOn raise
Change categoryOne or more of the five scope categoriesYesRequesterOn raise
Prompt version from / toVersion identifier plus content hashYes when prompt changesRequesterOn raise
Model identifier from / toVendor, model family, exact version stringYes when model changesSystem owner or ITOn raise
Output schema version from / toSchema version identifier plus hashYes when schema changesRequesterOn raise
Retrieval corpus version from / toCorpus version identifier, document count, cut-off dateYes when corpus changesData ownerOn raise
Requester and business reasonName, role, free textYesRequesterOn raise
Change classificationTier 1, 2, or 3 per section 6YesSystem owner, confirmed by QABefore testing
Regression evidenceTest reference, pass or fail per criterionYes for Tier 2 and Tier 3Validation leadBefore approval
Intended-use impactYes or No plus justificationYesSystem owner and QABefore approval
Rollback planText plus named artifactsYesIT or system administratorBefore approval
QA assessment and approvalSignature and dateYesQuality AssuranceBefore production
Post-change monitoring commitmentMetric, threshold, duration, ownerYes for Tier 2 and Tier 3System ownerBefore production
Implementation confirmationDate, time zone, who implementedYesIT or system administratorOn implementation
ClosureSignature and dateYesQuality AssuranceAfter monitoring period

4. What changed

Complete every row that applies. Where a row does not apply, write “No change” rather than leaving it blank, because a blank row does not prove the item was considered.

ElementFromToContent hash or identifierEvidence location
Prompt version<<FILL: e.g. PROMPT-DEV-TRIAGE v2.3>><<FILL: e.g. v2.4>><<FILL: hash of both>><<FILL: repository path and commit>>
Prompt text diff<<FILL: attach or reference the exact diff>><<FILL>>
Model vendor<<FILL: vendor name>><<FILL: vendor name>><<FILL>>
Model identifier and exact version string<<FILL: full version string as returned by the API, not a marketing name>><<FILL>><<FILL: API response capture>>
Model parameters<<FILL: temperature, top-p, max output tokens, stop sequences>><<FILL>><<FILL: configuration file and commit>>
Output schema version<<FILL: e.g. SCHEMA-TRIAGE v1.1>><<FILL>><<FILL: hash>><<FILL>>
Schema field changes<<FILL: fields added, removed, renamed, or with changed enumerations>><<FILL>>
Retrieval corpus version<<FILL: corpus ID and version>><<FILL>><<FILL: manifest hash>><<FILL>>
Retrieval corpus scope<<FILL: document count, source systems, cut-off date>><<FILL>><<FILL>>
Embedding model and retrieval parameters<<FILL: embedding model version, chunk size, top-k>><<FILL>><<FILL>>
Deterministic code around the model<<FILL: version, or "No change">><<FILL>><<FILL>>

Record the exact version string the vendor API returns, not the marketing name. A marketing name that stays constant while the served weights change underneath it is a known trap and is the reason your log records the returned identifier rather than the name a person typed into a configuration file.

5. Who requested the change and why

FieldEntry
Requester name and role<<FILL>>
Date requested<<FILL>>
Trigger for the change<<FILL: e.g. vendor deprecation notice, monitoring signal, override-rate breach, new document type in scope, defect report, periodic review action>>
Business reason, stated plainly<<FILL: what problem this change is meant to solve>>
Was the change forced by the vendorYes / No. If Yes, state the vendor notice reference and the deprecation date: <<FILL>>
Related deviation, monitoring alert, or CAPA<<FILL: reference, or "none">>
Alternatives considered<<FILL: including "do nothing" and why it was rejected>>

A vendor-forced change is still a change. The fact that a version is being deprecated on a fixed date is a schedule constraint, not a reason to skip the assessment. If the deprecation date is close, that is a planning failure to log, not a justification to shorten the testing.

6. Change classification and the re-validation decision rule

Answer every question. The classification is the highest tier triggered by any Yes.

6.1 Screening questions

#QuestionAnswerIf Yes, tier
Q1Does the change alter the approved intended use, the decision the tool supports, or the population of records it processes?Yes / NoTier 3
Q2Does the change touch a behavior classified as high risk in the risk assessment?Yes / NoTier 3
Q3Does the change alter the output schema in a way a downstream consumer reads, including field names, types, or enumerated values?Yes / NoTier 3
Q4Does the change alter the human checkpoint, what the reviewer sees, the disposition options, or the reason-capture behavior?Yes / NoTier 3
Q5Does the change move to a different model family, a different vendor, or a different hosting arrangement?Yes / NoTier 3
Q6Does the change alter the retrieval corpus in scope, source, or cut-off date, such that the tool can now cite material it could not cite before?Yes / NoTier 3
Q7Does the change alter instructional content of the prompt, the rubric, the in-context examples, or a model parameter that affects the output distribution?Yes / NoTier 2
Q8Does the change move to a different version string within the same model family, including a vendor patch?Yes / NoTier 2
Q9Does the change add, remove, or reword a constraint the model is expected to obey?Yes / NoTier 2
Q10Is the change confined to non-instructional text, code comments, logging labels that are not record fields, or formatting with no effect on model input or output?Yes / NoTier 1

6.2 What each tier requires

TierNameWhat it meansRequired evidence before production
Tier 1Documentation onlyThe change cannot alter model input or output. Nothing the model sees has changed and nothing downstream reads a different value.This record completed through section 9, a stated basis for the “cannot alter behavior” conclusion, and QA acknowledgement. No test execution required. A Tier 1 claim that turns out to have changed behavior is a deviation.
Tier 2Targeted regressionBehavior may shift, but intended use, schema, high-risk behaviors, and the human checkpoint are untouched.Re-run of the reference case set, including the false-negative subset and the negative-path cases, against the approved acceptance criteria. Before and after results side by side. QA approval before production. Post-change monitoring per section 10.
Tier 3Full re-validationThe change reaches intended use, a high-risk behavior, the schema, the human checkpoint, the model family, or the retrieval corpus scope.Fresh or updated risk assessment per Risk assessment: AI/ML system where intended use changed, re-execution of the affected protocol sections per Protocol: AI/ML system validation, an updated model card, and a validation summary or addendum approved before production.

Two rules that stop this decision from being gamed. First, the tier is proposed by the system owner and confirmed by QA; the requester does not self-certify their own change as Tier 1. Second, a sequence of Tier 2 changes to the same prompt within one review period is assessed cumulatively, because ten small edits that each pass a targeted regression can still walk the tool away from its validated behavior. Set the cumulative trigger in section 10 of your own procedure and record it here.

FieldEntry
Proposed tier<<FILL: 1 / 2 / 3>>
Proposed by (name, role, date)<<FILL>>
Cumulative-change check: prompt edits since last full re-validation<<FILL: count and dates>>
Cumulative trigger reachedYes / No
QA confirmed tier<<FILL: 1 / 2 / 3>>
QA rationale where the tier differs from the proposal<<FILL: or "not applicable">>

7. Regression evidence

Complete for Tier 2 and Tier 3. Name the exact case set and its version, because “we tested it” without a versioned case set is not evidence.

FieldEntry
Reference case set identifier and version<<FILL: e.g. REF-TRIAGE-CASES v3, 120 cases>>
Case set held out from prompt developmentYes / No. If No, explain why the result is still meaningful: <<FILL>>
Cases re-run<<FILL: count and selection basis>>
Negative-path cases includedYes / No (malformed output, model unavailable, timeout, empty retrieval result)
False-negative subset includedYes / No (known-positive cases the tool must not miss)
Execution date and executor<<FILL>>
Raw results location<<FILL: path or record reference>>

7.1 Results against acceptance criteria

CriterionThresholdResult before changeResult after changeDeltaMetComment
<<FILL: e.g. recall on known-positive cases>><<FILL: e.g. not less than 0.95>><<FILL>><<FILL>><<FILL>>Yes / No<<FILL>>
<<FILL: e.g. schema validation pass rate>><<FILL: e.g. 100 percent, malformed output must route to a human>><<FILL>><<FILL>><<FILL>>Yes / No<<FILL>>
<<FILL: e.g. false-flag rate>><<FILL>><<FILL>><<FILL>><<FILL>>Yes / No<<FILL>>
<<FILL: e.g. negative-path routing to human>><<FILL: 100 percent of induced failures route to a human>><<FILL>><<FILL>><<FILL>>Yes / No<<FILL>>

7.2 Behavior differences worth noting even where criteria are met

Case referenceOutput beforeOutput afterAssessment
<<FILL>><<FILL>><<FILL>><<FILL: benign wording change / material change in reasoning / requires investigation>>

A criterion met on aggregate can still hide a changed behavior on a specific case class. Read the per-case differences, not only the summary number, and record what you found even when the answer is “no material difference.”

7.3 Regression outcome

FieldEntry
All acceptance criteria metYes / No
Where not met, deviation reference<<FILL: or "not applicable">>
RecommendationProceed / Proceed with additional controls / Do not proceed
Additional controls proposed<<FILL: e.g. temporary increase in review sampling, tightened monitoring threshold>>

8. Intended-use impact

QuestionAnswerJustification
Does the tool now do anything it was not approved to do?Yes / No<<FILL>>
Does the tool now process a record type, product, or process it was not approved for?Yes / No<<FILL>>
Does the output now carry more decision weight than before?Yes / No<<FILL>>
Has the human checkpoint changed in any way?Yes / No<<FILL>>
Does the intended-use statement need re-approval?Yes / No<<FILL>>
Is a fresh risk assessment required?Yes / No<<FILL: if Yes, reference the new assessment>>

If the answer to any of the first four questions is Yes, the intended-use statement is re-approved and a fresh risk assessment is completed before this record can be approved. A change to intended use is not a prompt edit with paperwork; it is a new validated use of the tool.

9. Rollback plan

FieldEntry
Previous state artifacts retained<<FILL: prompt file and hash, model version string, schema file and hash, corpus manifest and hash, configuration file and commit>>
Where the previous state is stored<<FILL: repository and tag, or controlled archive location>>
Rollback method, step by step<<FILL: exact steps, including who executes and what is verified afterwards>>
Rollback testedYes / No. Date tested: <<FILL>>
Estimated time to revert<<FILL>>
Rollback authority<<FILL: role authorized to order the revert without waiting for a governance meeting>>
Rollback trigger conditions<<FILL: e.g. schema failure rate above threshold, override rate breach, any missed known-positive case in the post-change window>>
Constraint on rollback<<FILL: e.g. previous model version is deprecated by the vendor on a stated date, after which rollback is unavailable>>
Records created under the changed version if rolled back<<FILL: how those records are identified, reviewed, and dispositioned>>

The last two rows are the ones teams skip. If the previous model version will be withdrawn by the vendor, your rollback has an expiry date and you should say so before you approve the change, not discover it on the day you need it. And if you do roll back, the records the tool produced while the changed version was live still exist and still need a disposition.

10. Post-change monitoring commitment

Complete for Tier 2 and Tier 3.

MetricBaseline before changeThreshold that triggers actionMonitoring frequencyDuration of heightened monitoringOwner
Override rate<<FILL>><<FILL>><<FILL: e.g. weekly>><<FILL: e.g. 4 weeks or 200 calls, whichever is later>><<FILL>>
Schema validation failure rate<<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>
Reviewer-reported false flags<<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>
Findings from false-negative sampling<<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>
<<FILL: tool-specific metric>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>
FieldEntry
Where the monitoring data comes from<<FILL: name your call-level log, the one that records each request, the output returned, and the human disposition taken on it>>
Routine monitoring procedure this hands back to<<FILL: e.g. /templates/sop-ml-model-monitoring-drift>>
Who reviews the heightened-monitoring result and closes it<<FILL>>

Heightened monitoring after a change is a defined period with an end, not an open-ended intention. State when it stops and who signs that it stopped cleanly.

11. QA assessment and approval

Approval is recorded before the change reaches production. A change implemented before this section is signed is an uncontrolled change to a validated system and is handled as a deviation under SOP: deviation management.

FieldEntry
QA assessment summary<<FILL: what QA reviewed and what it concluded>>
Tier confirmed<<FILL>>
Regression evidence adequateYes / No
Intended use unchanged or re-approved<<FILL>>
Rollback plan adequateYes / No
Monitoring commitment adequateYes / No
Conditions of approval<<FILL: or "none">>
QA decisionApproved / Approved with conditions / Rejected
RoleNameSignatureDate
Requester<<FILL>>
System owner<<FILL>>
Validation lead (Tier 2 and 3)<<FILL>>
Quality Assurance approver<<FILL>>

12. Implementation and closure

FieldEntry
Implemented by (name, role)<<FILL>>
Implementation date and time, with time zone<<FILL>>
First call executed under the new version (call ID)<<FILL: from the call log, establishes the boundary between versions>>
Post-implementation verification performed<<FILL: what was checked immediately after go-live and by whom>>
Training or communication to reviewers<<FILL: what reviewers were told, and when>>
Documents updated<<FILL: model card, intended-use statement, system inventory, SOP, training material>>
Heightened monitoring period ended<<FILL: date>>
Monitoring outcome<<FILL: within thresholds / action taken>>
Rolled backYes / No. If Yes, date and reason: <<FILL>>
Record closed by (QA name, signature, date)<<FILL>>

The “first call executed under the new version” row matters more than it looks. It is what lets an investigator, months later, sort the call log into before and after without guessing. It is also the join between this record and whatever call-level log your tool writes, the log that captures each request, the output returned, and the human decision taken on that output, so a reviewer can trace any individual result back to the exact prompt text and model version that produced it.

13. Instructions for use

  1. Raise the record before the change is made, not after. A record written to describe something already in production is a deviation record, not a change record, and should be labeled as one.
  2. Capture the exact artifacts and hashes in section 4 at the moment you raise it, while the previous state is still live and easy to capture.
  3. Get the tier confirmed by QA in section 6 before you spend effort on testing, so the test scope matches the tier.
  4. Execute the regression against a versioned case set. If you do not have one, build it before you make the change; a change is a bad time to discover you have no reference behavior to compare against.
  5. Do not sign section 11 on the same day the change went live unless the timestamps genuinely support that order of events. Inspectors read timestamps.
  6. Close the record only after the heightened monitoring period ends, so the record carries the outcome and not just the intent.

14. Retention

FieldEntry
Retention period<<FILL: not less than the retention of the GxP records the tool influenced, and not less than the life of the system plus your defined period>>
Retention basis<<FILL: reference your records retention schedule>>
Storage location<<FILL>>
Format<<FILL: electronic record under access control, with audit trail>>
Attachments retained with the recordPrompt files and hashes, schema files, corpus manifest, raw regression results, vendor notices, approval signatures

This record is itself a GxP record. It is subject to the same electronic record controls as anything else the quality system holds, including audit trail, access control, and protection from alteration. If it is held in a system that permits silent editing, it is not adequate.

15. References

21 CFR Part 11 (electronic records and signatures), for the record and signature controls applied to this record and to the tool’s own outputs. 21 CFR 211.68 (automatic, mechanical, and electronic equipment), 211.100 and 211.160 (written procedures and controls), and 211.180 and 211.194 (records), as applicable to the GMP activity the tool supports. EU GMP Annex 11 (computerised systems), current version effective June 2011, with a draft revision issued for public consultation in July 2025 and a companion draft Annex 22 addressing artificial intelligence; both are drafts and are not final, so treat them as direction of travel rather than requirement. EU GMP Chapter 4 (documentation) and the change control expectations in EU GMP Part I Chapter 1. FDA guidance, “Computer Software Assurance for Production and Quality Management System Software” (current version issued 3 February 2026). The 2022 draft, dated 13 September 2022, and the final issued 24 September 2025 were both titled “Computer Software Assurance for Production and Quality System Software”; the current version added “Management” to the title and supersedes the 2025 version. ICH Q9(R1), Quality Risk Management, for sizing assessment and test effort to risk. ICH Q10, Pharmaceutical Quality System, for the change management system this record sits inside. GAMP 5 Second Edition (ISPE, 2022), for the software category logic and the change control expectations applied to custom applications; described here in original wording, not quoted. EU Artificial Intelligence Act, Regulation (EU) 2024/1689, which applies in phases with obligations staged over time. Confirm the current application dates and whether your system falls in scope before relying on them.

This is general guidance to adapt to your own quality system, not legal or regulatory advice. Confirm the current version and clause numbers of each reference, and the current status of any draft instrument, before issue.


Filled specimen

The following shows the same record completed for an illustrative internal deviation-triage assistant that suggests a preliminary criticality tier for a QA reviewer to confirm or override. The company, tool, version strings, and results are invented for illustration. The vendor is described generically; where a real vendor is named anywhere in this template it is a neutral example only.

Control header (filled)

FieldEntry
Record numberAICC-2026-014
Linked change requestCR-2026-0873
AI tool name and IDDeviation Triage Assistant, DTA-01
Validated state referenceVSR-AI-004 v2.0, approved 14 February 2026
Record raised byJ. Okafor, AI System Owner
Date raised08 June 2026
StatusClosed

What changed (filled)

ElementFromToHashEvidence
Prompt versionPROMPT-DTA v2.3PROMPT-DTA v2.49f1c…a072 to 4b8e…c119Repo dta-tool, commit 7ad20c9
Prompt text diffAdded two worked examples covering aseptic-fill environmental excursions; no change to the criticality rubric or the output instructionDiff attached, Attachment 1
Model vendorCommercial LLM API provider, unchangedUnchangedSupplier assessment SA-2025-11
Model identifier and exact version stringvendor-lm-3.5-20260112vendor-lm-3.5-20260112, no changeAPI response capture, Attachment 2
Model parameterstemperature 0.0, top-p 1.0, max output tokens 800Unchangedconfig/dta.yaml, commit 7ad20c9
Output schema versionSCHEMA-DTA v1.1Unchanged2c40…ff81Repo dta-tool
Retrieval corpus versionCORPUS-DTA-2026Q1, 412 documents, cut-off 31 March 2026UnchangedCorpus manifest
Deterministic code around the modeldta-core v1.6UnchangedRepo dta-tool

Requester and reason (filled)

FieldEntry
RequesterM. Ibrahim, QA Deviation Coordinator
TriggerMonitoring signal: override rate on environmental monitoring excursions from the aseptic fill area reached 31 percent over eight weeks against a threshold of 25 percent, while the site-wide rate stayed at 9 percent
Business reasonThe tool consistently under-rates environmental excursions in the fill area because the prompt examples do not cover that record shape. Reviewers are correcting the same category of miss repeatedly
Vendor-forcedNo
Related alertMON-2026-031, override-rate threshold breach
Alternatives consideredDo nothing was rejected because the override pattern is systematic, not random. Retraining reviewers was rejected because the reviewers are already correcting it, which is the control working; the tool is the thing that is wrong. Removing fill-area deviations from the tool’s scope was considered and held in reserve if this change does not resolve the pattern

Classification (filled)

#QuestionAnswer
Q1Intended use, decision, or record population alteredNo
Q2High-risk behavior touchedNo. Criticality suggestion is advisory; QA confirms every tier with a reason
Q3Output schema alteredNo
Q4Human checkpoint alteredNo
Q5Different model family, vendor, or hostingNo
Q6Retrieval corpus scope alteredNo
Q7Instructional content, rubric, or examples alteredYes
Q8Different version string within the same familyNo
Q9Constraint added, removed, or rewordedNo
Q10Confined to non-instructional textNo
FieldEntry
Proposed tierTier 2, targeted regression
Proposed byJ. Okafor, 08 June 2026
Prompt edits since last full re-validation2 (v2.2 on 03 April 2026, v2.3 on 11 May 2026)
Cumulative trigger reachedNo. Site trigger is 4 prompt edits or 6 months since full re-validation, whichever comes first
QA confirmed tierTier 2
QA rationaleAgreed. Examples added, rubric untouched, schema and checkpoint untouched. QA added a condition that the false-negative subset be expanded to include ten fill-area cases before execution

Regression evidence (filled)

FieldEntry
Reference case setREF-DTA-CASES v4, 140 cases, expanded from v3 by 10 fill-area cases per the QA condition
Held out from prompt developmentYes for 130 cases. The 10 new fill-area cases were drawn from closed 2025 deviations not used to write the new examples; the two examples added to the prompt came from a separate 2024 set
Cases re-run140, full set
Negative-path cases includedYes, 8 cases: malformed JSON, truncated output, API timeout, API 500, empty retrieval, oversized input, non-English free text, null required field
False-negative subset includedYes, 24 known-critical cases the tool must not under-rate
Execution date and executor15 to 16 June 2026, R. Delacroix, Validation Lead
Raw resultsVAL-EXEC-2026-119, attached
CriterionThresholdBeforeAfterDeltaMet
Recall on known-critical cases (must not under-rate)Not less than 0.950.92 (22 of 24)1.00 (24 of 24)+0.08Yes
Agreement with QA-confirmed tier, all casesNot less than 0.800.810.86+0.05Yes
Agreement, fill-area subsetNot less than 0.800.600.85+0.25Yes
Schema validation pass rate on valid input100 percent100 percent100 percent0Yes
Negative-path routing to a human100 percent of 8 induced failures8 of 88 of 80Yes
Over-rating rate (false escalation)Not more than 0.150.090.11+0.02Yes
Case referenceBeforeAfterAssessment
REF-DTA-071Minor, rationale cites “single excursion within alert limit”Major, rationale cites “Grade A excursion, action limit, aseptic fill in progress”Material change, and the intended one. QA-confirmed tier for this closed deviation was Major
REF-DTA-112MajorMajor, rationale rewordedBenign wording change, no change in output tier
REF-DTA-129MinorMajorMaterial change, and incorrect. QA-confirmed tier was Minor. Over-rating, absorbed by the reviewer, counted in the over-rating rate above and flagged for monitoring
FieldEntry
All acceptance criteria metYes
Deviation raisedNone
RecommendationProceed with additional controls
Additional controlsOver-rating rate rose from 0.09 to 0.11. Monitor weekly rather than monthly for the first four weeks and report to the AI governance forum

Intended-use impact (filled)

QuestionAnswerJustification
Does anything newNoSame output, same records, same reviewer step
New record type, product, or processNoFill-area deviations were already in scope; the tool was performing badly on them, which is not the same as them being out of scope
More decision weightNoSuggestion remains advisory, QA confirms every tier with a reason
Human checkpoint changedNo
Intended-use statement re-approval neededNo
Fresh risk assessment requiredNo. RA-AI-004 v1.2 reviewed on 12 June 2026 and confirmed still applicable, noted in the assessment file

Rollback (filled)

FieldEntry
Previous state retainedPROMPT-DTA v2.3 (hash 9f1c…a072), config/dta.yaml at commit 3e91b40, repository tag dta-v2.3-validated
StorageRepository dta-tool, protected tag, plus controlled archive QA-ARC/AI/DTA
Rollback methodIT reverts to tag dta-v2.3-validated, restarts the service, runs the 12-case smoke set, and confirms the prompt hash in the first live call record matches 9f1c…a072. Executed by IT, verified by the system owner
Rollback testedYes, 17 June 2026, in the qualified test environment, 6 minutes end to end
Estimated time to revertUnder 30 minutes including verification
Rollback authorityQA Manager or AI System Owner, either acting alone, without waiting for the governance forum
Rollback triggersAny known-critical case under-rated in the post-change window; over-rating rate above 0.20 in any monitored week; schema failure rate above 1 percent; override rate above 35 percent in any week
Constraint on rollbackNone. The model version is unchanged and the prompt artifact is held internally, so rollback does not depend on the vendor
Records created under the changed version if rolled backIdentified by prompt hash in the call log. All affected deviations would be re-triaged by a QA reviewer within 5 working days and the re-triage recorded on each deviation record

Monitoring commitment (filled)

MetricBaselineAction thresholdFrequencyDurationOwner
Override rate, fill-area subset31 percentAbove 25 percent, or below 3 percentWeekly6 weeksM. Ibrahim
Override rate, all deviations9 percentAbove 25 percent, or below 3 percentWeekly6 weeksM. Ibrahim
Over-rating (false escalation) rate0.11 in regressionAbove 0.20 in any weekWeekly6 weeksR. Delacroix
Schema validation failure rate0 percentAbove 1 percentWeekly6 weeksIT system administrator
False-negative sampling: 20 unflagged-as-Major deviations per month reviewed by QA0 misses in the last cycleAny single miss of a case QA rates MajorMonthly3 cyclesJ. Okafor

Approval and closure (filled)

FieldEntry
QA assessment summaryTier 2 confirmed. Regression executed on an expanded case set with the fill-area gap closed and the false-negative subset clean at 24 of 24. Over-rating rose slightly and is controlled by weekly monitoring and the reviewer step. Rollback tested and independent of the vendor. Approved with the monitoring condition
QA decisionApproved with conditions
RequesterM. Ibrahim, signed 19 June 2026
System ownerJ. Okafor, signed 19 June 2026
Validation leadR. Delacroix, signed 19 June 2026
QA approverS. Whitfield, QA Manager, signed 22 June 2026
Implemented byIT system administrator, T. Lindqvist
Implementation date and time23 June 2026, 07:12 Central European Time
First call under the new versionCall ID DTA-2026-06-23-0007
Post-implementation verificationPrompt hash in the first three live call records confirmed as 4b8e…c119 by the system owner on 23 June 2026
Reviewer communicationBriefing note issued to the 6 QA deviation reviewers on 22 June 2026 explaining that fill-area suggestions should improve and that they must continue to confirm every tier with a reason
Documents updatedModel card REC-AI-004 v1.3, system inventory entry DTA-01, reviewer job aid JA-QA-018
Monitoring period ended04 August 2026
Monitoring outcomeFill-area override rate settled at 11 percent, site-wide at 8 percent, over-rating 0.12, no schema failures, false-negative sampling clean across 3 cycles. Returned to routine monitoring
Rolled backNo
Record closed byS. Whitfield, QA Manager, signed 07 August 2026

Reading it: the change was triggered by a monitored metric rather than by someone deciding the prompt could be better, the tier was confirmed by QA rather than self-certified, the regression was run against a versioned case set that included the cases the tool must never miss, a small degradation was found and honestly recorded rather than smoothed over, the rollback was tested before approval, and the record was not closed until the monitoring period ended with an outcome. That sequence, signal to assessment to evidence to approval to monitored outcome, is what an inspector is looking for.

Common inspection findings this record prevents

  • A validated AI tool is running a prompt that does not match the prompt in the validation package, and nobody can say when or why it changed.
  • The model version in production differs from the version in the validation report because the vendor updated it and nobody assessed the update.
  • Prompt changes are made under a general IT change ticket that contains no assessment of validated behavior and no regression evidence.
  • Every change is classified as low impact by the person who made it, with no independent quality confirmation of the classification.
  • Regression testing is claimed but there is no versioned reference case set, so “it still works” cannot be demonstrated.
  • The tool’s output schema changed and a downstream consumer silently broke, discovered later during a record review.
  • The retrieval corpus was refreshed with new source documents, changing what the tool can cite, with no assessment.
  • No rollback path exists, or the rollback depends on a model version the vendor has already withdrawn.
  • Change records are approved after the change was already live, with signature dates that prove it.
  • A series of small prompt edits accumulates until the production behavior no longer resembles the validated behavior, with each individual edit defensible and the aggregate never assessed.
  • The change record is closed on the day of implementation, so it records an intention to monitor rather than a monitored outcome.

How to adapt this record

  1. Set your record numbering, owner, and the change control procedure this record feeds in the control header and section 11.
  2. Rewrite the five scope categories in section 2 to match how your tool is actually built. If you do not use retrieval, delete those rows rather than leaving them permanently marked “not applicable.”
  3. Set your own cumulative-change trigger in section 6. Four prompt edits or six months is a starting point, not a rule; base it on how much your tool’s behavior moves per edit.
  4. Replace the acceptance criteria in section 7.1 with the criteria from your own validation package, including the exact thresholds. The criteria here are illustrative.
  5. Name your reference case set and its version in section 7. If you do not have one, build it first; this record depends on it.
  6. Point section 10 at your routine monitoring procedure so heightened monitoring hands back cleanly rather than trailing off.
  7. Set the retention period in section 14 against your own records retention schedule and the retention of the GxP records the tool influenced.
  8. Confirm every regulation in section 15 against the current published version before issue, and check the status of the draft Annex 11 revision and draft Annex 22, which were not final at the time of writing.
Use madhadi.com as an app Full screen, works offline, one tap from your home screen.