Documentation
¶
Overview ¶
Command dst-generate runs metamorphic equivalence scenarios (hand-written or LLM-generated) through the deterministic SQL engine over SimFDB and reports findings. It is the "judge" half of the LLM-adversarial loop (RFC-199 Tier 2): an LLM proposes scenarios — schemas, data, and groups of queries it asserts are equivalent — as JSON; this tool executes them and reports where the engine returns DIFFERENT rows for equivalent queries.
go build -o /tmp/dst-generate ./cmd/dst-generate /tmp/dst-generate -dir ./generated/ # check a directory of *.json scenarios /tmp/dst-generate # check the hand-written seed corpus
Findings are split: INEQUIVALENCE (the engine disagreed with itself on two equivalent queries — a REAL finding: engine bug OR a wrong LLM equivalence claim, triage it) vs ERRORED (a query the engine could not plan — usually the LLM emitting unsupported SQL, low signal).