Silently wrong answers, made structurally impossible.
A grammar layer for analytical data — built so an AI agent either returns a verified answer, or refuses with diagnostics. Nothing in between.
What makes it different
four platesJoins and aggregations resolve automatically, with a correctness proof.
No more fan-trap, no more chasm-trap. The grammar layer walks the schema for you and emits a query that's either provably correct against the model — or refuses with a diagnostic.
Verified answer, or a refusal with diagnostics. Nothing in the middle.
The dangerous case in analytical pipelines isn't a wrong answer — it's a wrong answer that looks fine. Coframe makes that case structurally unreachable.
An agent describes the question. Coframe writes the SQL — and stands behind it.
Agents shouldn't be hand-crafting joins from scraped table descriptions. Hand off the intent; let the grammar layer produce, verify, and return.
Three tiers — A, AA, AAA — like WCAG, but for analytical trust.
Pick the bar you want answers to clear. The grammar layer marks every result with its level — so you know, at a glance, what you're trusting and what you're not.
Two editions, one grammar
pick your barCoframe Core
The open grammar. Pre-v1.0, evolving in the open, backed by Polars and DuckDB.
Coframe Pro
Multi-backend, audit-grade verification, support contract. For teams shipping analytical answers to humans who'll act on them.
Coframe is being engraved in the open. The grammar, the verifier, and the manual all live on GitHub — pulled apart, debated, and re-cut as the spec settles toward v1.0.