Most data pipelines are dynamically typed in the worst way: they accept anything and produce something. The Toleron signal platform makes every signal and every operator port declare three independent things, and refuses to connect ports that do not agree.
The axes
| Axis | Values | Mistake it rules out |
|---|---|---|
| Value type | float · int · bool · string · enum · vector · spectrum | Feeding a classification code into an arithmetic reduction |
| Quantity + unit | amperes · rpm · dimensionless · explicitly non-physical … | Adding a load percentage to a current |
| Temporal class | raw block stream · sampled stream · regular series · event | Treating a kilohertz block stream as one value per window |
The temporal axis matters most because its failure is silent. A per-window operator given raw blocks will happily emit a number per block, at the block rate, and every downstream chart will look reasonable. The platform instead rejects the edge at authoring time with a message naming the reduction the author is missing — usually a window mean or an RMS.
Immutable contracts
Operators are versioned and their contracts are immutable: a second version of a gate, a feature extractor or the model operator is a new contract, never an edit. Semantics never move under a running graph. The operator library spans sources, windowing, reductions, conversion, features and classification.
Why it makes AI affordable
Because a model score is a typed signal like any other, the output of one pipeline is a valid input to the next, to a chart, to a trigger or to a recording — with no special case anywhere. And because the platform was built on top of the existing archive, every recording already made became signal-addressable with zero data migration. The type system is not overhead. It is the reason composition is safe.