Isn't this just low-code / Model-Driven Engineering again?
MDE's thirty-year track record of underdelivering is the right prior to bring — which is exactly why the substrate-equivalence claim here is checked empirically by a conformance harness, not asserted in a whitepaper.
Bring that skepticism; it's earned. Model-driven engineering, CASE tools, and low-code platforms have promised "the model is the system" for thirty years and mostly delivered lock-in, leaky abstractions, and escape hatches that swallowed the project.
The difference here is not a better diagram format. It is three commitments those systems never made together:
-
The model is a plain JSON file in your repo, not a proprietary canvas. An LLM can tend it,
git diffcan review it, and you can walk away with it. -
Every substrate is regenerated mechanically, and then graded. The conformance harness scores each substrate's output cell-by-cell against a Postgres-oracle answer key. When a transpiler is wrong, a cell goes red — publicly — instead of a consultant explaining the discrepancy away.
-
The boundaries are named, not papered over. The declarative core is first-order logic with aggregates; transitive closure is provably outside that class, so the rulebook declares
closureas a first-class field type rather than faking it. Where something genuinely isn't a rulebook concern (auth plumbing, I/O), it lives in explicit customization seams that survive every rebuild.
Low-code fails when you hit the 20% it can't express and the platform fights you. Here the 20% is enumerated, has named seams, and the 80% is independently verified instead of trusted.