6. A conservative deterministic parser runs before any model

Status: accepted

Context

Many table instructions are explicit ("age > 25", "sort by country then age desc", "20 per page"). Sending them to a model adds cost, latency, a network dependency and non-determinism for no benefit. A parser that guesses, however, produces confidently wrong queries.

Decision

A dictionary-driven parser handles explicit phrasing. It claims an instruction only when every meaningful token is understood; otherwise it declines and the instruction goes to the model (or returns unsupported when none is configured). Values never start at keywords, field names, the table name, recency phrases or prepositions.

Consequences

• Explicit instructions work offline, instantly and identically every time. The evaluation gate asserts exact matches for them.
• Semantic phrasing ("Indian customers", "older than 40", "high-value accounts") is deliberately left to the model; the gate asserts the parser declines it.
• Every parser bug found so far that produced a wrong query (found by the example app and the evaluation dataset) came from claiming too much; each fix made the parser decline instead of guess.