Reviewer-agent pass — three-agent review of the IFU/CEP/CER edits

Internal working document

This page captures the three reviewer-agent passes (audit-deliverable-reviewer, bsi-clinical-auditor, celine-clinical-consultant) run on the audit-visible edits produced by this task, together with the fixes applied in response. Internal only. The audit-visible artefacts are the updated IFU, CEP, and CER files.

Scope of the review

Three files in scope — all read-only to the reviewer agents, edited by this task prior to the review:

• apps/eu-ifu-mdr/versioned_docs/version-1.1.0.0/installation-manual/user-interface.mdx
• apps/qms/docs/legit-health-plus-version-1-1-0-0/product-verification-and-validation/clinical/Evaluation/R-TF-015-001-Clinical-Evaluation-Plan.mdx (integrator paragraph in § "Integrator integration requirements as risk controls")
• apps/qms/docs/legit-health-plus-version-1-1-0-0/product-verification-and-validation/clinical/Evaluation/R-TF-015-003-Clinical-Evaluation-Report.mdx (paragraphs in the "Distinction between API output and physician interface" block)

Review 1 — audit-deliverable-reviewer (leak / audit-readiness gate)

Verdict: clean, with one minor hygiene note.

Key findings:

• No critical leaks. No BSI ticket IDs, bsi-non-conformities/ path references, internal person names, dev-tool nouns, TypeScript / React / MDX / Docusaurus references, or task-folder paths in the rendered prose. JSX tags are used as render directives only; prose uses regulatory names ("Top-5 prioritised differential view", "malignancy-prioritisation gauge", "referral recommendation", "binary malignancy-surfacing safety indicators").
• One hygiene warning on R-SKK citation: the IFU Safety-Indicators sub-section originally cited R-SKK with a paraphrased description ("covering patient-visible erroneous-result pathways") rather than the canonical hazard field from R-TF-013-002.json.
• Cross-references in the five new IFU paragraphs and the CEP/CER paragraphs all resolve to real documents (R-TF-015-003, R-TF-013-002, Installation manual: User Interface, Installation manual: Endpoint specification).

Fix applied: R-SKK citation realigned to the literal hazard field, R-SKK "Incorrect results shown to patient", matching the style used for the other RC entries.

Review 2 — bsi-clinical-auditor (clinical substance gate)

Verdict: clean on IFU; one critical and several warnings on cross-document consistency between CEP, IFU, and CER.

Critical gap raised

• CER-IFU asymmetry on the malignancy surface. The CER's original bulleted list carried "The malignancy-surfacing output in the P₂ position when present (the architectural severity-prioritisation constraint described in section Novel features)" — describing a ranking rule, not the independent visually-distinguishable gauge the IFU mandates. The same surface described two different ways in two audit-visible documents. BSI-voice: "I cannot tell if I should be auditing the integrator's ranking logic or the integrator's widget rendering. That's not closeable until those are the same story."

Fix applied: the CER's four-bullet MUST list was refactored to match the IFU's surface-by-surface structure. The P₂ architectural property is now folded into the Top-5 bullet as a property of the device-produced order (not a separate integrator-MUST requirement — it is the device's algorithmic behaviour, not the integrator's). The malignancy bullet now describes the visually-distinguishable gauge per the IFU's language.

Warnings raised

• CER Top-5 bullet did not echo "cohesive block, clearly labelled" from the IFU. Fix: CER Top-5 bullet now states "rendered in descending-probability ranked order ... as a single visually cohesive block with each candidate clearly labelled", matching the IFU.
• CER referral bullet was bare ("The referral recommendation."). Fix: CER referral bullet now echoes "co-visible with the Top-5 prioritised differential view, not behind additional interaction".
• CER softer-than-CEP language: "requirements themselves are set, and clinically validated, by the device". Fix: promoted to "requirements themselves are MANDATED by the manufacturer and their conformance is a precondition of the clinical performance validated in this CER", with the non-delegation / co-controlled-risk-control-agent sentence appended for full CEP/CER parity.
• CER did not echo the per-surface RC-ID mapping present in the CEP (only the six binary indicators carried RC IDs in the CER). Fix: the refactored CER bullet list now carries per-surface RC IDs, matching the CEP's mapping.
• IFU Report-section preface missing the "outside intended use" consequence clause present in the CER. Fix: the Report-section preface now closes with "Integrators that do not meet these mandatory display requirements operate the device outside its intended use, and the clinical performance reported in R-TF-015-003 no longer applies to that deployment."
• IEC 62366-1 framing absent at the top of the IFU User Interface page. Fix: the IFU page's opening paragraph now identifies the integrator as an intended user under IEC 62366-1 and names the Installation Manual as the volume of the IFU directed at that user.

Positive observations kept as-is

• The IFU Report-section preface (Insert 1 of the plan) correctly distinguishes mandatory surfaces from recommended features.
• The six-binary-indicators sub-section is the cleanest of the four — correct independence rationale, correct cross-reference split (Endpoint spec for derivation; User Interface for UI mandate), correct anchoring to benefit 7GH sub-criterion (c).
• Non-delegation posture preserved in CEP and IFU; now also mirrored explicitly in the CER (previously only implicit).

Review 3 — celine-clinical-consultant (pillar-mapping + causal-chain gate)

Verdict: pass overall. One critical (already merged with BSI Review 2), several hygiene warnings, all addressed.

Critical gap raised (same as BSI #1 above)

• P₂-position architectural constraint referenced in the CER as an integrator-MUST, but not mirrored in the IFU's MUST clauses. Resolution: the P₂ rule is an algorithmic property of the device's ranked output, not an integrator requirement (the integrator preserves ranked order; the device does the P₂ promotion). Consistent resolution applied: P₂ now described in the CER Top-5 bullet as a property of the device-produced order, not as a separate integrator-MUST; the IFU Top-5 MUST clause ("preserve descending-probability ranked order") covers it implicitly.

Hygiene warnings and fixes

• R-75H hazard quote in IFU Referral sub-section (originally "The medical device outputs incorrect clinical information") did not literally match the JSON field. Fix: realigned to R-75H "Incorrect clinical information".
• R-SKK hazard quote in IFU Safety-Indicators sub-section (originally "covering patient-visible erroneous-result pathways") did not literally match the JSON field. Fix: realigned to R-SKK "Incorrect results shown to patient".
• R-75H missing from the Malignancy gauge RC mapping, even though malignancy misrendering is the paradigm case of "incorrect clinical information" reaching the care provider. Fix: R-75H added to the Malignancy gauge RC list in both the IFU (sub-section prose) and the CEP (per-surface RC enumeration). CER updated via the refactored bullet list.

Check-by-check verdict (Celine framework)

Check	Verdict
1 — Pillar mapping	Pass. Top-5, gauge, referral, six indicators all Pillar 3; 346-cat API claim remains Pillar 2.
2 — Causal-chain closure	Pass. VCA → Pillar 2 → Pillar 3 → patient benefit; UI mandates reinforce the Pillar-3 end.
3 — Integrator language (non-delegation)	Pass. MUST language consistent across all three documents; IEC 62366-1 intended-user framing now open-stated at the top of the IFU page.
4 — Surrogate-endpoint anchoring	Pass (not stressed by this edit).
5 — Cross-document consistency	Pass after fixes (P₂ conflation, referral-bullet bare-ness, RC-list asymmetry, and "MANDATED" wording all resolved).
6 — Class-IIb indirect-benefit readiness	Pass. Integrator-as-co-controlled-risk-agent formulation preserved.
7 — Load-bearing studies integrity	Pass. MAN_2025 correctly anchored as Pillar 3 Top-5-view evaluator; not mis-cited as a Pillar 2 anchor. R-TF-015-012 not invoked by this edit — no mis-attribution introduced.
7-bis — RC sanity	Pass after fixes. Hazard-quote alignment and R-75H-on-gauge addition resolved.

Net effect of the fixes

The three reviewer agents' combined report yielded one critical and ~8 warnings. All have been applied to the audit-visible files (IFU, CEP, CER). The cross-document MUST / surface enumeration / RC-ID mapping is now symmetric across the three documents:

Surface	CEP	IFU	CER
Top-5 prioritised differential view	MUST + R-BDR, R-A96	MUST + "cohesive block, clearly labelled" + R-BDR, R-A96	MUST + "cohesive block, clearly labelled" + R-BDR, R-A96
Malignancy-prioritisation gauge	MUST + R-HBD, R-BDR, R-DAG, R-75H	MUST + "visually distinguishable, always visible, re-surface regardless of rank" + R-HBD, R-BDR, R-DAG, R-75H	MUST + "visually distinguishable, re-surface regardless of rank" + R-HBD, R-BDR, R-DAG, R-75H
Referral recommendation	MUST + R-BDR, R-75H	MUST + "co-visible with Top-5, not behind additional interaction" + R-BDR, R-75H	MUST + "co-visible with Top-5, not behind additional interaction" + R-BDR, R-75H
Six binary malignancy-surfacing safety indicators	MUST + R-BDR, R-HBD, R-SKK	MUST + "always visible, not hidden behind expansion" + R-BDR, R-HBD, R-SKK	MUST + "always visible, not hidden behind expansion" + R-BDR, R-HBD, R-SKK

A BSI auditor following the paper trail from CEP → IFU → CER now finds three documents that agree on what the integrator MUST do, why they MUST do it (Pillar-3 clinical-performance anchoring in R-TF-015-003), and how failure is risk-controlled (R-TF-013-002 entries per surface).

Optional further hardening (outside this task's scope)

The reviewer agents flagged one item as "nice-to-have" that is not blocking and is not included in this task:

• A one-line pointer in the IFU to the IEC 62366-1 usability-engineering file (R-TF-012) explicitly identifying the integrator as an intended user under IEC 62366-1 and the User Interface section as integration-phase information-supplied-with-the-device content. If this chain exists in R-TF-012 today, a one-line cross-reference in the IFU would close the loop. Deferred.