Out-of-tolerance notices, drafted from the as-found readings your technicians recorded.

A technician records the as-found readings, sees one or more points outside the tolerance the job was run to, and flags the work order. Then someone on the quality desk gathers the instrument details, the failing points, what was adjusted or repaired, and the account's written terms on how and when they want to hear about it, and writes the notice for the quality manager to sign.

In the proof of concept, AI does that gathering on copies and drafts each notice: the instrument, the readings exactly as recorded with the tolerance and its source, what happened to the instrument afterward, the customer's notification term it answers, and a list of anything missing. It sends nothing. Your reviewer compares every draft with the notice your lab sent for the same failure.

Out of whose tolerance? The customer may not judge the instrument by the limits you ran it to. The equipment calibration SOP published by the Frederick National Laboratory for Cancer Research defines out of tolerance against its own specified level of calibration, which can differ from the manufacturer's, so failing the maker's specification may not count. It says plainly that an out-of-tolerance condition its calibration vendor reports may not be one under its process tolerance, and that its quality engineer decides. So the draft names the limit every reading was judged against and where that limit came from: the maker's specification, a customer tolerance on the PO, or your procedure.

The notice feeds someone else's remedial action. Drug makers calibrate under a written program that 21 CFR 211.160 says must include specific directions, schedules, limits for accuracy and precision and provisions for remedial action when those limits are not met, and an instrument that doesn't meet its specifications may not be used. Device makers have worked under FDA's Quality Management System Regulation since February 2, 2026, which brings ISO 13485:2016 into Part 820 by reference. Frederick's SOP expects its vendor to notify it and offer options when an out-of-tolerance instrument can't be put right. Each account's terms decide who gets the notice, by when and in what form, and the draft quotes the term it is answering.

Condition on arrival is part of the story. A nonconformance guide from one of NIST's own calibration divisions lists shipping damage among the causes, has the client told when an item fails inspection, allows repair only on the client's written request, and has an item that behaves against its design reported to the client at once. For labs that seek recognition under it, the optional ANSI/NCSL Z540-1 annex in NVLAP's handbook calls for tamper-resistant seals on adjustments whose removal would invalidate a calibration, and for instructions on equipment whose seals are damaged or broken. NIST's example certificate shows the matching fields: conditions of the items as submitted, such as fingerprints and dents, and any cleaning done before calibration. The draft carries those notes exactly as recorded.

Some failures have no as-found reading at all. Frederick's vendor letter asks for a written statement with the certificate when a unit arrives inoperable, and its SOP needs no as-found data for new equipment. ILAC's uncertainty policy normally leaves transport effects out of the stated uncertainty and asks the lab to raise them at contract review if it expects them to matter. So a reading taken after a rough trip is reported as recorded, with the stated uncertainty copied from your data, and nothing recalculated. A blank as-found field goes on the missing list, never filled in.

Facts and opinions travel separately. A2LA's rules on references to its accreditation want any opinion or interpretation outside the accredited work marked with a disclaimer, and prefer it on a separate letter without the symbol. The impact on the customer's product is the customer's question: Frederick's SOP has each out-of-tolerance event investigated for its effect on product inside its own program. The NIST guide also has a client who accepts nonconforming work say so in writing, kept on file and noted in the report. Your reviewer will check that each draft states facts only, names the customer's term, and lists what the file doesn't hold, such as an account with no notification term on record. These rules are summarized for context, not as legal or technical advice.

After a yes from your reviewer, new failures get drafted the same way for roughly 30 days.

Each notice waits for your quality manager, who edits, signs and sends it the way your lab does now.

Scoring continues on the live notices. Any second job is a new decision at its own price.