Build and maintain FMEA

Prerequisites

architectures.read; architectures.write for create, generate, edit, lifecycle, restore, and delete.
A reviewed architecture and current Memory for grounded AI generation; an active provider for generation.

Open /fmea or /fmea/{fmeaId}.

Review the workload architecture and refresh stale Memory.
Open /fmea and choose a buildable architecture/workload suggestion, create a blank document, or select Generate from Memory.
Monitor draft, verify/refine, and save phases; a parse failure in the refinement pass can fall back to the first draft.
Open the document and validate every table, scope reference, failure mode, effect, cause, current control, and recommendation.
Add/remove tables or rows as needed; use per-table regeneration only when that table can be safely replaced.

Expected result: A saved draft contains one or more reviewable failure-mode tables grounded in architecture Memory.

Verification: Trace each modeled component/failure to the architecture, Memory, evidence, or explicit workshop input.

In a cross-functional workshop, define consistent 1–10 criteria for Severity, Occurrence, and Detection.
Edit factor cells; the UI updates RPN immediately and the server recomputes authoritatively on save.
Use $\mathrm{RPN}=S\times O\times D$; a zero/blank factor leaves RPN blank.
Interpret bands: Critical $\ge 200$, High $120$–$199$, Medium $40$–$119$, Low $<40$.
Sort/review high RPN, but separately review catastrophic low-occurrence and common-cause failures.
Wait for the short auto-save before navigating or exporting.

Expected result: Every complete row has normalized factors, derived RPN, and risk band.

Verification: Reopen the document and spot-check multiplication and band; out-of-range values are normalized server-side.

Replace blank/generated owner and due-date placeholders with human-approved values.
Enter recommended action and row status; record post-mitigation factors only after controls are implemented and verified.
Review open/in-progress/completed summaries.
Move the document from Draft to In review and Published only after approval; use Archived for retained inactive records.
Do not treat lifecycle status as validation by itself.

Expected result: Risks have accountable action, due date/status, and reviewed document lifecycle.

Verification: Confirm owners accepted responsibility and completed controls have evidence before lowering post-mitigation scores.

Open revision history and preview the desired snapshot.
Restore only after comparing it with current content; restore creates a new current revision.
Select ⬇ CSV for flat tabular data or ⬇ Excel for Summary plus per-table sheets, conditional formatting, dates, and live RPN formulas.
Open the workbook and recalculate formulas in the target spreadsheet application.

Expected result: History remains intact and the export reflects current saved tables without unresolved token syntax.

Verification: Compare table/row counts, factors, RPN, status, owner, and dates between UI and export.

On /fmea, identify the amber workload deleted badge.
Open the retained FMEA and decide whether it must be exported, archived, relinked through a supported workflow, or deleted.
Do not expect new buildable suggestions for deleted workloads.
Move obsolete documents to Trash; restore if needed, or purge after retention review.

Expected result: Orphaned risk records remain visible for deliberate disposition rather than silent loss.

Verification: The deleted workload does not reappear as buildable and the FMEA’s chosen lifecycle/export/delete outcome is recorded.

AI can omit modes, invent controls, or score inconsistently; human facilitation is mandatory.
Regeneration can replace manual content. Preserve a revision/export first and prefer per-table regeneration.
RPN can under-rank catastrophic or systemic failures.
Owner/due date must never be inferred as fact.
Soft deletion is recoverable; purge/empty Trash is permanent. Deleted workloads do not cascade-delete FMEA.

Symptom	Resolution
No buildable suggestion	Confirm active workload, linked architecture, Memory, and permission.
Generation is partial	Check provider and Memory; review the first-pass fallback rather than assuming completeness.
RPN is blank	Enter non-zero S, O, and D values from 1 through 10.
Score changes after save	Inspect out-of-range/non-numeric values; server normalization is authoritative.
Edits are missing	Wait for auto-save, avoid concurrent tabs, and inspect revisions.
Excel differs	Recalculate formulas and verify spreadsheet support for formatting/formulas.