SemPods is a classifier you can hold in your head: a small, isolated model scoped to a single concept you defined in plain language. Same geometry, same math, and the same receipts as Foreshock — with an immediate thirst to become a hyper-accurate expert on your exact need.
Each ring is one pod: its own phrase set orbiting its own centroid. The dashed lines are the only thing they share — components, by choice.
You name the concept and write a handful of example phrases that live inside it. That phrase set is the model — readable, arguable, and revisable in an afternoon.
A pod knows one thing. It cannot drift into a neighbouring concept, and retraining it cannot quietly damage anything else you deployed.
Pods return distances and similarities. No summaries, no reasoning, no generated text — nothing to hallucinate and nothing to leak.
SemPods runs the identical stack behind Foreshock: an embedding model supplies the geometry already present in ordinary language, and a Nearest Centroid Classifier with diagonal covariance estimation does the rest. Standard, documented, inspectable math on numbers.
What changes is the population being measured. Foreshock's Aldous model is one example of what this engine can carry — an affective specification built over years. Yours might be allergen disclosure, raptor call taxonomy, weld defect language, or a species key. If a concept can be scoped in language, it can be a pod.
Deploy your scoring schema to APIs we host for you at the start, and take it in-house whenever you like.
The same centroid geometry works on image embeddings. When vision pods land, a QA line or a field survey can score a photograph against a pod the same way it scores a sentence — and get the same receipt back.
A good phrase set is hard-won knowledge. SemPods will let you publish a pod design — its scope, its phrase set, its calibration, its receipt template — so someone in an adjacent field can fork it, swap components, and credit the lineage.
Classifier components travel well: a distance metric someone tuned for clinical shorthand may be exactly what a veterinary pod needed.
Receipts came from Trust & Safety, where an unexplained decision costs you a user. They turn out to matter just as much when a score rejects a batch, flags a specimen, or fails an inspection.
Which pods responded, how strongly, and against which phrase set.
The region of the sample that carried the signal, without exposing guarded terms.
Remove the matching vectors and re-score what remains, to see if anything of value is left.
A link to dispute the result, and one to report suspected bias in the pod itself.
We are picking the first pods with the people who understand their own domain best. Tell us what you would scope, and we will tell you honestly whether geometry can see it.