An agent writing code for the platform needs two things a human engineer gets from experience: the grammar of each language and a way to know whether what it wrote is valid. This release serves the first and provides the second, with one rule: validation never runs SQL.
Proofs
- SQL by prepare A statement is prepared and closed, so tables and columns resolve against the live schema, with placeholders neutralised. A test prepares a DELETE and an UPDATE against a populated table and asserts every row survives. The XML dialect is transformed to the engine's SQL first and chained into the same prepare.
- Scripts by parse JavaScript through the GraalVM parser, Python through GraalPy, Groovy through its compiler, and 4GL through its parser with a walk that reports the 26 screen statements that need a terminal. Each says "cannot prove here" when its engine is absent.
Grammars
- The XML dialect Fourteen families and 56 root elements outlined from the schema generated out of the runtime's own classes, with the expression-function catalogue cross-checked against the grammar: 31 functions accepted but uncatalogued and 16 catalogued but not expressions, both ratcheted.
- The scripting library Forty-eight top-level modules and 440 nodes found by introspecting the constrained host-access policy. A proof rejects a library member the runtime does not export, with a nearest-name suggestion.
- 4GL Seventy-seven statement kinds, 31 built-ins and 296 keywords read by reflection from the engine jar.
Grammars written by hand drift from the code. These are read from it at request time, so an agent is told what the server actually accepts.