An embedding model turns a sample into numbers. From there it is distance measurement — the kind you could check by hand if you wanted to.
A phrase, a report line, a field note, a label. Plus your API key. Nothing else is required and nothing else is kept.
The text becomes vectors describing the relationships between its tokens. This is the only point at which a large model touches the sample, and it produces numbers, not opinions.
Each pod holds the distribution of its own phrase set. A Nearest Centroid Classifier with diagonal covariance estimation measures the specimen against it using Standardized Euclidean Distance, so dimensions that matter to your concept count for more than dimensions that do not.
Remove the vectors that matched the pod and score what is left. If the remainder is total noise, then the population was the entirety of the specimen and rejection or escalation is advised. Vectors can never be turned back into text, but they can be pulled out of other vectors.
Similarity and distance floats, a heat map of where the signal sat, and the phrase set that produced the reading. The sample text itself is never returned to anyone but you.
Fig. 1 — The classification path. Everything after the embedding step is ordinary geometry on numbers.
Each phrase embedded individually, differentially weighted, holding a real centroid set. High accuracy and fine discernment; adjusting one means partial re-training. Ideal for concepts that do not move much linguistically — a species key, a regulatory clause, a defect class.
One fast embedding pass over a mixed bag of phrases, so it can be updated as often as every second. Lower magnitude means lower resolution — informative, but pair it with corroborating signal rather than wiring it straight to an automated action.
Because monolithic pods have no per-phrase distribution, an erasure re-score against one is a hint rather than a verdict. Research pods support it properly, which is why they carry anything that leads to a consequential decision.
| Stage | What happens |
|---|---|
| Design | Name the pod, write the phrase set, set thresholds, choose whether erasure is available and what the receipt shows. |
| Calibrate | Run known samples through it, watch the distances, tighten the phrase set. This is the whole of the work, and it is readable work. |
| Deploy | Your scoring schema becomes an endpoint on infrastructure we host for you. One call in, scores and a receipt out. |
| Publish | Optionally share the design — scope, phrases, calibration, receipt template — so others can fork it and credit the lineage. |
| Take it home | Export and self-host whenever you want. Nothing about the method depends on us staying in the loop. |
We put the math out front because we want you to understand it. A classifier you cannot argue with is not a classifier, it is a rumour.