Acme Neurovascular

A Class II stroke triage algorithm, mid 510(k) preparation. Everything below is sample data for one fictional company.

62%

Submission readiness

4

Open CAPAs

3

Traceability gaps

128

Controlled documents

What needs attention

RAi ranks by regulatory consequence, not by date.

Item	Why it matters	Status
`UN-004` Implant confirmation	User need with no requirement addressing it. Validation under 820.30(g) is against user needs.	Gap
`CAPA-017` Field complaint trend	Effectiveness check is overdue by 11 days.	Overdue
`SOP-QMS-004` Document control	Approved, signature valid, no action.	Current
2 AI proposals	Waiting on a human to authorize or reject. Oldest is 3 days.	Awaiting you

Design control traceability

21 CFR 820.30. The matrix is anchored on user needs, so a need that nothing addresses is visible rather than hidden.

4

User needs

9

Requirements

75%

Needs addressed

78%

Reqs verified

Open traceability gaps (3) Every link an inspector would expect and cannot find.

Ref	Type	Gap	Clause
`UN-004`	User Need	No requirement addresses this user need	`820.30(c)`
`SWR-014`	Requirement	No verification test	`820.30(f)`
`DS-006`	Design Spec	Design output with no design input	`820.30(d)`

User needs to requirements to validation

Ref	User need	Requirements	Validation	Status
`UN-001`	View 3D bone anatomy overlaid on the surgical field	`SR-001`, `HW-003`	`VAL-001`	Addressed
`UN-002`	Import pre-operative CT from the hospital PACS	`SWR-014`	`VAL-002`	Addressed
`UN-003`	Alignment metrics shown during resection	`SR-007`	—	No validation
`UN-004`	Confirm the correct implant is loaded	— missing	—	Orphan

Requirements arrive by spreadsheet in most companies. RAi imports one sheet holding user needs, system, software and hardware requirements together, reads the level from a Level column or the ID prefix, skips section banner rows, and keeps your original IDs so existing cross references still resolve.

Predicate finder

Searches the FDA 510(k) database and ranks candidates on what that dataset actually contains.

Device name

Stroke Lesion Detection Algorithm

Classification

892.2050

Intended use

Radiological CAD for triage of stroke lesions on head CT in adult patients

Top predicate matches

K213567 — Viz.ai LVO Stroke Algorithm

Viz.ai Inc. · cleared 2021 · Class II, QAS, 892.2050

80

match score

Classification

Class II, QAS, 892.2050

HIGH MATCH

Device & indication

Viz.ai LVO Stroke Algorithm

HIGH MATCH

Recency

2021-06-01

HIGH MATCH

Intended use wording, technological characteristics and performance data are not in the FDA dataset. They are in the 510(k) summary.

K051234 — Bone Screw Fixation System

Orthopaedic supplier · cleared 2005 · Class II, HWC, 888.3030

10

match score

Different product code, different regulation, no overlap with the indication, and eighteen years old. Ranked low so it does not waste your time.

Scoring uses product code and regulation number, device name overlap with your indication, and how recent the clearance is. RAi does not invent sensitivity, specificity or intended use wording for a predicate, because the FDA dataset does not publish them.

The AI cannot write to your quality system

AI reads and proposes. It cannot create, change, approve or sign a controlled record. This is enforced in the API, so no interface path bypasses it.

A connected AI client was asked to open a CAPA. This is what came back.

A proposal is awaiting authorization. I have not created a record. A person in your organization needs to review the fields, edit anything that is wrong, and authorize it under their own identity with a reason.

Proposal PRO-0042 Field by field. Edit anything before it becomes a record.

Field	Proposed
Title	Recurring false positives in low-dose CT series	unchanged
Severity	High	edited by reviewer
Root cause	Training set under-represented low-dose acquisitions	unchanged
Owner	Quality Manager	unchanged

Authorized by a named person, with the reason recorded. The ledger marks this one accepted after edits, not accepted as written.

21 CFR Part 11 audit trail

Field level before and after, a compelled reason for change, attributed and timestamped, and not editable by anyone.

When	Who	Field	Before → after	Reason
2026-09-24 14:02	j.rivera	Severity	Medium → High	Second field report with the same failure mode
2026-09-24 14:02	j.rivera	Owner	— → Quality Manager	Assigned at intake review
2026-09-22 09:15	a.chen	Status	Draft → Open	Opened following complaint trend review

Signature bound to content

A signature covers a hash of the document, so editing it afterwards invalidates the approval instead of inheriting it.

Signed by	a.chen, Quality Manager
Meaning	Approved
Content hash	`sha256:9f2c41a8…`
Current state	Valid, content unchanged

AI ledger

Every generation is recorded: which model, which prompt version, what grounded it, and what the reviewing person then did with it.

When	Surface	Model	Prompt	Disposition
2026-09-24 14:01	CAPA draft	Opus	`capa_draft` sha256:3b7e…	Accepted after edits
2026-09-23 11:40	Predicate analysis	Opus	`predicate_analysis` sha256:c14f…	Accepted as written
2026-09-23 10:02	Document draft	Opus	`draft_doc` sha256:8a05…	Rejected

The prompt text is never stored or shown. The hash proves which prompt version ran, so two records sharing a hash provably ran the same prompt and a changed hash provably marks a change, without disclosing the prompt itself.

Protected health information

Records can be marked as containing PHI, and submitted content is screened for patient identifiers regardless, so the control does not depend on someone remembering to mark the record. PHI is not transmitted to the AI sub-processor.