What this page does not claim
Three altitudes
Not three levels of difficulty, and not a course you are enrolled in. The three positions are marked by what a reviewer is holding at each: nothing yet, then your file, then a record they can recompute. Nothing on this page is inferred from what you read.
- 01Where you already areThe reviewer holdsnothing yet
Nothing has been submitted and nobody outside has read anything. What is decided here is whether the thing you built is a device at all, and which of the files on your laptop are already records.
4 of 4 written
- 02The seamsThe reviewer holdsyour file
The file has crossed the desk. These are the places where a machine-learning practice and a regulated practice describe the same object and disagree about what must happen to it — which is why the collision lives here and not in the first altitude.
3 of 5 written · 2 not yet
- 03The artefactThe reviewer holdsa record they can recompute
The far end is not a certificate. It is a record of what ran, on what, against criteria fixed before the result was seen — one a reviewer can check without trusting the party that produced it.
3 of 3 written
Where you already are
Nothing has been submitted and nobody outside has read anything. What is decided here is whether the thing you built is a device at all, and which of the files on your laptop are already records.
The seams
The file has crossed the desk. These are the places where a machine-learning practice and a regulated practice describe the same object and disagree about what must happen to it — which is why the collision lives here and not in the first altitude.
The artefact
The far end is not a certificate. It is a record of what ran, on what, against criteria fixed before the result was seen — one a reviewer can check without trusting the party that produced it.
Words that mean something else here
The ledger pairs a file with the name a reviewer gives it. These pair a word with the meaning the regulated estate gives it — the same double-naming, applied to vocabulary, because reading a SPEQ page with the wrong sense loaded is a worse failure than not reading it.
- reference standard
The ground truth a model’s output is scored against — whatever the labels say is true.
A characterised reference material of established quality, used to calibrate a method and qualify a result. Its identity is a property of the substance, not of a judgement. where SPEQ defines it →
- labelling
The annotation a human drew on a training image, and the process of drawing it.
Everything printed on or accompanying the product — the instructions for use, the carton, and, in the FDA sense, promotional material. It is a regulated output in its own right. where SPEQ defines it →
- validation
The held-out split you tuned against, between train and test.
Evidence that the finished thing meets the user’s need and its intended use — the question your eval notebook does not answer, which is design verification instead. where SPEQ defines it →
- specificity
The true-negative rate: of the negatives, how many were called negative.
In analytical method validation, the ability to assess the analyte unequivocally in the presence of everything else in the sample. Not a rate at all. where SPEQ defines it →
The left column is SPEQ synthesis — no standard defines these senses, which is the whole problem. The right column points at the page where SPEQ does define the word.
Where the line is not drawn
Two rows in the ledger are dashed and two seams in Altitude 02 have no page behind them. They are the ones most likely to matter to a team building imaging AI: SPEQ has written nothing on how a training corpus and its labellers become a controlled dataset, and nothing at all on DICOM or PACS integration — that second gap was measured across the whole content estate, not assumed.
What SPEQ has not written yet
- Controlled dataset, and who drew the labelstrain/ · val/ · test/
- Not yet writtendicom/ · PACS integration
- Design inputs for a learned systemA requirement you can trace to a fitted parameter, when the parameter was never specified in the first place.
- Training data as a controlled artefactProvenance, consent, and who drew the labels — the dataset as an object under change control.
Drawn dashed rather than left out: a gap you can see is worth more than a list that looks complete. The first people in the room shape what gets written first.
Where to go next
Each solid row in the ledger is a link into the standard that gives the reviewer-side name its force. Two are worth reading first: the design-control clauses that turn your weights and your eval notebook into a design output and a verification record, and the change-control plan that decides what you may retrain without a new submission.