The idea came from a simple distinction: an agent can *extract* semantics from an interface, or the application can *publish* the semantics it already knows about its own state.
Semantic browsers make extraction much better, but they still cannot recover domain semantics that were never present in the UI. SIL is an attempt to explore the other side: an application publishes its state, actions, constraints, and events directly to an agent, using STF as the underlying representation.
This is an early, open specification rather than a finished standard. In particular, authenticated interfaces, permissions, stale state, and consistency are still open questions.
I’m the author of SIL and STF.
The idea came from a simple distinction: an agent can *extract* semantics from an interface, or the application can *publish* the semantics it already knows about its own state.
Semantic browsers make extraction much better, but they still cannot recover domain semantics that were never present in the UI. SIL is an attempt to explore the other side: an application publishes its state, actions, constraints, and events directly to an agent, using STF as the underlying representation.
This is an early, open specification rather than a finished standard. In particular, authenticated interfaces, permissions, stale state, and consistency are still open questions.
I wrote a longer explanation of the extraction vs. publication distinction here: https://dev.to/bswvladimirbeep/stf-from-semantic-inference-t...
I’d be particularly interested in criticism of the architecture and in comparisons with WebMCP and semantic browsers.