Do we need this task?

Verdict (2026-04-21): No — largely redundant. The task addresses a presentational / cross-referencing gap, not a substantive evidence gap. The per-Bucket-C-feature evidence map the task proposes to create already exists in the CEP, in a different section. The fix is a one-sentence cross-reference edit, not a five-deliverable task folder. Recommended action: close the task as-scoped; apply the edits listed in the final section instead.

Context: the task was opened on 2026-04-20 during a CEP adequacy pass, with a finding attributed to bsi-clinical-auditor. The user later noted that the reviewer agents in use that day were slightly corrupted versions of the current definitions. The clean agents in place today (with the factual-claim-discipline clauses at bsi-clinical-auditor.md L267-269 and celine-clinical-consultant.md L164-166) would, on a re-run over the current CEP, most likely down-rank this finding to a low-severity cross-reference note.

The task's premise

From CLAUDE.md in this folder:

The CEP declares Phase 1 (exploratory / first-in-man / feasibility) "intentionally not populated," with exploratory evidence discharged via Phase 0 V&V, Phase 6 publications, and Phase 4 equivalence. This blanket discharge does not survive per-feature scrutiny for the Bucket-C novel features — they are, by definition, not covered by equivalence, and some are not covered by the Phase 6 severity-publication corpus either.

The premise is half-right. The per-feature evidence already exists in the CEP — it is just not in the Phase-1 paragraph.

What the CEP already contains

R-TF-015-001-Clinical-Evaluation-Plan.mdx lines 204-215, under the heading "Features not covered by equivalence: device-specific evidence requirement", contains a full Bucket-C table with four columns:

| Bucket-C feature | Pre-market evidence source | Acceptance criterion | Acceptable-gap / PMCF fallback |

Populated for all four Bucket-C items the task would enumerate:

• Six binary malignancy-surfacing safety indicators — risk-control verification in R-TF-013-002 for risk IDs R-BDR, R-HBD, R-SKK; integrator-requirement verification per the IFU; Pillar-3 clinician-decision-point observation of the six indicators in MC_EVCDAO_2019, IDEI_2023 and NMSC_2025. Acceptance criterion: observed miss-rate ≤ 0.01 % per use (mirrors R-75H safety threshold). No acceptable-gap — direct pre-market evidence; post-market confirmation via PMCF A.1 and A.2.
• P₂=1 architectural severity-prioritisation constraint — device-specific architectural-invariant test documented in R-TF-012-038; deterministic invariant with 100 % pass rate as a release gate. No acceptable-gap — direct pre-market evidence.
• Novel severity scales (IHS4, SALT, GAGS) — per-scale analytical V&V in R-TF-028-005; for IHS4, published Pillar 2 validation in AIHS4_2023 and pilot feasibility in AIHS4_2025; for SALT and GAGS, R-TF-028-005 only. SALT and GAGS Pillar-3 clinician-in-the-loop confirmation declared an acceptable gap per MDCG 2020-6 §6.5(e), discharged by PMCF Activity B.
• Novel ICD-11 categories (device-level delta scope) — enumerated in R-TF-015-003 §16.6 per-category evaluation table; pre-market analytical evidence via R-TF-028-005; pre-market Pillar-3 coverage via the MRMC studies (BI_2024, PH_2024, SAN_2024, MAN_2025). Categories without pre-market Pillar-3 evidence declared acceptable gaps per MDCG 2020-6 §6.5(e) and discharged by PMCF C.1, C.2 and D/E/F.

Line 215 explicitly points forward to the CER's "Equivalence gap analysis and device-specific evidence" section, so the narrative propagates into R-TF-015-003.

What is actually missing

Exactly one thing: the Phase-1 discharge paragraph at line 915 does not cross-reference the Bucket-C table at lines 204-215. A reviewer reading the Phased-progression section in isolation sees only the blanket "intentionally not populated" discharge and may not connect it to the Bucket-C evidence map documented earlier in the same document.

The fix is a single sentence appended to line 915:

"For the Bucket-C novel features enumerated in §[Features not covered by equivalence: device-specific evidence requirement] above, the exploratory evidence is discharged per feature in that table rather than by the blanket Phase-0/Phase-6/Phase-4 route described in this paragraph."

Content-complete; presentation-tightened.

Nick's lens — what he would actually push back on

Nick (MDR Annex XIV Part A §1(a), MDCG 2020-1 §4.3/§4.4, MDR Article 61(6)(b)): the existing discharge route for a software MDSW — Phase-0 analytical V&V + Phase-6 publications + equivalence — is a legitimate reading of the guidance for a pre-existing-software case. Nick's likely push-back surface is:

• Is the Pillar-3 real-patient evidence sufficient (Ranks 2-4 coverage of the indication space)?
• Is MRMC being over-claimed as primary Pillar-3 evidence rather than Rank-11 supporting evidence?
• Is the acceptable-gap / PMCF-confirmation structure under MDCG 2020-6 §6.5(e) explicit per-indication?

All three of those concerns are already being addressed elsewhere (tasks 3b2/3b3, 3b4, 3b5, 3b6, 3b7 and the CEP itself). None of them map onto "write a per-feature Phase-1 narrative." Nick would not draft a non-conformity around the absence of a Phase-1-taxonomy cross-link; he would ask about evidence sufficiency.

Erin's lens — where she would actually land

Erin (traceability and storyline): she would notice the disconnect between the Phase-1 paragraph and the Bucket-C table. She would ask for the connection to be made visible, not for a new body of work to be produced.

A reviewer-friendly cross-reference sentence in the Phase-1 paragraph, plus (optionally) a reverse cross-reference from the Bucket-C table back to the Phased-progression section, closes her concern entirely.

Signs this is agent overreach

Two tells in the task brief suggest the reviewer agent that spawned this over-generalised:

1. The task re-derives the Bucket-C inventory from scratch ("at minimum… safety indicators, P₂=1, novel severity scales, novel ICD-11 categories") without acknowledging that the CEP already has that exact inventory mapped to evidence at lines 204-215. Had the agent read those lines first, its own methodology (Check 1: pillar / evidence mapping) would have stopped it.
2. The Option-C recommendation ("add a Phase-1 row to the CDP mermaid diagram and to the Planned Evidence Classification table") is gold-plating. The Phase-1 "intentionally not populated" framing is an intentional CDP design choice documented in the paragraph itself; rewriting the CDP to "populated by feasibility-equivalent non-clinical pre-market evidence" is a narrative-shape change that would cascade into the CER and the PMCF Plan for no BSI-visible benefit.

This is consistent with the project's standing feedback_no_agent_file_fabrication.md rule: reviewer agents must verify a suspicious finding against the actual document before reporting it. The corrupted version of bsi-clinical-auditor.md in use on 2026-04-20 may have been missing the factual-claim-discipline clauses that the current version carries at L267-269.

Recommendation — the edits that replace this task

Close task-3b12 as-scoped. Replace it with a single short edit pass on two files:

1. R-TF-015-001 line 915 (Phase-1 discharge paragraph) — append one sentence cross-referencing the Bucket-C device-specific-evidence table earlier in the same document. Draft text above.
2. R-TF-015-001 lines 204-215 (Bucket-C evidence table) — optional one-cell addition to the "Six binary malignancy-surfacing safety indicators" row: cite R-TF-025-005/006/007 formative usability sub-studies as additional Phase-1-equivalent exploratory evidence for the UI-level novelty of the safety indicators. This is the only substantively new evidence-source surfaced by the task-3b12 brief.
3. R-TF-015-003 §16.6 ("Equivalence gap analysis and device-specific evidence") — mirror the cross-reference, if the corresponding CER section does not already link back to the CEP's Phased-progression section.

Total effort: one edit session, one reviewer-agent pass on the three changes. No bucket-c-inventory.md, per-feature-evidence-map.md or cep-rewrite.md needed — the content is already in the CEP.

De-risk assessment for the BSI response

• Substantive evidence de-risk: near-zero. The Bucket-C evidence is already documented.
• Presentational / reviewer-ergonomics de-risk: modest. The cross-reference edit closes the reviewer's parse path from Phased-progression to Bucket-C evidence in a single sentence.
• Opportunity cost of running the task as-scoped: non-trivial. Producing four new deliverables, re-running three reviewer agents, and potentially rewriting the CDP narrative introduces cross-document knock-on risk (CER §16.6, PMCF Plan, IFU cross-references) for a presentational gain that is better achieved with a one-sentence edit.

If a reviewer still asks for per-feature Phase-1 discharge

If, after the cross-reference edit, a BSI reviewer in Round 1 or Round 2 still explicitly asks for a per-feature Phase-1 discharge narrative separate from the Bucket-C device-specific-evidence table, task-3b12 can be re-opened with its current scoping intact. The brief is already drafted and can be picked up as-is. Until that happens, the task should stay closed — pre-empting the question by producing the task now trades certain effort for speculative regulatory benefit.

Cross-references

• Task brief (the thing being assessed): ./CLAUDE.md.
• CEP Phase-1 discharge paragraph (proposed edit target): R-TF-015-001 §Phased progression of the clinical evaluation, line 915.
• CEP Bucket-C device-specific-evidence table (content that makes this task redundant): R-TF-015-001 §Features not covered by equivalence: device-specific evidence requirement, lines 204-215.
• CER mirror section: R-TF-015-003 §16.6 Equivalence gap analysis and device-specific evidence.
• Reviewer agents (clean versions currently in place): .claude/agents/bsi-clinical-auditor.md, .claude/agents/celine-clinical-consultant.md.
• Standing rule on agent fabrication: feedback_no_agent_file_fabrication.md (auto-memory).