Inspection to CAPA

Chain 2. ncr-writeup → root-cause-capa.

The path from "we found a bad part" to "we fixed the reason, and we can prove it." It ends on the 8D and its redacted customer pack.

Quick start

→ Write the NCR against a CITED requirement. Cannot find one? Block. Never invent
→ Search history AT CREATION, not on request. Did the earlier fix hold?
→ Gate 1: the quality owner confirms the NCR and its containment
→ Verify the clocks BEFORE correlating. Misaligned? Refuse, and say what survives
→ Ranked candidate chains with evidence FOR and AGAINST. Never one verdict
→ 8D: D1-D5 and D7 populated. D6 and D8 stay empty placeholders
→ Gate 2: nothing customer-facing leaves without the redaction diff reviewed

Inputs

The seven steps

1. An inspection result or a suspect lot starts it.

2. Write the nonconformance against the requirement, with quantity and location. All five containment locations are addressed, including with a zero quantity, because an omitted location reads as "none there" rather than "not checked."

Gate 1 — the quality owner confirms the NCR and the containment plan. This is a pause, and the chain says so: "Paused at Gate 1, after step 2. The NCR is a draft. A quality owner confirms four things before step 4 — the finding, the cited requirement, the suspect scope, and the containment plan." A missing citation is one reason the NCR is not ready; it is not the gate. Fixing the citation does not unblock the chain — the owner's confirmation does. The history search in step 3 still runs at write time, because it belongs to the NCR. Per ../../shared/approval-gates.md.

3. Search history for the same problem. This runs automatically at write time, not on request, because the moment of detection is the only moment anyone is looking.

Whether the earlier fix held is the highest-value single line in this chain. A recurrence after a closed CAPA is a materially different finding from a first occurrence, and it changes what happens next. Surface it as such.

4. Read the prior tickets, inspection data and production history together, aligned onto one zone-aware timeline. Verify the clocks first; where alignment cannot be verified, refuse to correlate and say which conclusions survive anyway.

5. Propose ranked candidate causal chains, each stated as a chain with supporting and contradicting evidence attached to each link, and the cheapest test that would confirm or kill it. Never one confident verdict: a confident wrong root cause closes the investigation and the fix fails.

6. Draft the 8D. D1 to D5 and D7 populated from the evidence. D6 and D8 stay as structured placeholders, because they need a real-world outcome and a human signature.

7. Produce the customer-facing pack with a redaction pass. Gate 2.

Approval gates

The redaction pass is a reviewable diff, never a silent filter. The internal 8D carries cost data, other customers' names, supplier identity and internal process detail; the customer version must not. A recurrence stays in — removing it is misrepresentation, not redaction.

Output

Every document this skill produces is printed in the reply, in full. Saving a copy to a file as well is fine, as Word or Excel per ../../shared/replies.md, never .txt; a file instead of the reply is not. The person reads the reply, and every delimited block and outbound check lives in it — a document only on disk has been neither seen nor checked. A record handed to another skill is saved and summarised, never pasted as raw YAML; memory blocks are written where the plant's memory lives; and the name on an approval is asked, never taken from the account. All three per ../../shared/replies.md.

The NCR in the plant's controlled template, the containment plan ordered by urgency, the recurrence finding with whether the previous fix held, the ranked candidate causes with their discriminating tests, the drafted 8D, the evidence appendix, and the redacted customer pack.

The figures block — required, inside the evidence appendix. Every measurement, control limit, ratio and count the chain states is printed from the computation and listed with it, and the prose quotes the block. Two member skills doing the same arithmetic must not arrive at two different multiples of the same limit. Per ../../shared/computed-figures.md.

Where the leading candidate points at purchased material, the supplier-caused flag is raised, which is the trigger into Chain 5. See supplier-escape-to-alternate-source.

The Quality command-center page: the Quality Register — open corrective actions, carrying the NCR and the 8D by reference, updated in place per ../../shared/live-dashboard.md. The chain writes no register row of its own. ncr-writeup writes the row at step 2 and root-cause-capa adds the 8D fields at step 6, both keyed <site>_<ncr_ref>, so one chain run leaves one row. The chain reports the register once, at the end: the dashboard line with its URL and the sensitivity line, or the fallback with "The Quality Register was not updated: [reason]." It is reported as updated only when a publish returned a URL in this run, and never as shared. Read ../../shared/command-center.md.

Edge cases and failure modes

• The requirement cannot be located — the chain blocks at step 2. Never invent one. That is a complete and honest outcome.
• Traceability is incomplete — the suspect scope widens and says why. Never narrow it to make the number look better.
• The clocks cannot be aligned — step 4 refuses to correlate. Non time-dependent analysis still runs, and the output says which is which.
• Two candidates are supported equally — present both. A stated tie is a real finding, and it names the test that would settle it.
• No prior history exists — say so. Absence of recurrence is not evidence of a one-off.

Degradation

Check the freshness of every export before correlating on it. A scheduled export that silently stopped produces a timeline with a hole in it, and the hole looks like an absence of evidence rather than an absence of data. State the coverage window of each source in the evidence appendix.

If Limble isn't included, or no CMMS connector is live — minor here. Maintenance history arrives in the same scheduled export as everything else and the maintenance-correlation check runs on it unchanged. What changes is freshness: name maintenance explicitly as an export-sourced input in the evidence appendix rather than letting it read as live, since a repair logged after the last export will not be on the timeline at all.

Chain position

This is Chain 2 itself. Members: ncr-writeup, root-cause-capa.

Metric. CAPA closure time and first-pass yield, both owned by root-cause-capa. ncr-writeup also moves nonconformance closure time and escape response time inside this chain; the chain does not claim them, because it is named for the corrective-action loop.

Reference files