IONIAN Labs — Language & runtime
Project Diell — A language small enough for an AI to verify before it emits.
Diell is a tiny language an AI writes under a formal grammar, and a deterministic engine verifies before anything renders. Grammar-constrained generation makes invalid programs impossible to emit; a simulator then checks behaviour, so the failure mode is a caught error instead of confident nonsense. The whole machine spec fits in one file a model can load in a single shot.
Language & runtime · Active research · v0.9.1 · 2026
Diell is a tiny language an AI writes under a formal grammar, and a deterministic engine verifies before anything renders. Grammar-constrained generation makes invalid programs impossible to emit; a simulator then checks behaviour, so the failure mode is a caught error instead of confident nonsense. The whole machine spec fits in one file a model can load in a single shot.
The problem it attacks
When a language model writes code, the failure mode that costs the most is not a crash — it is confident nonsense. The model emits something syntactically plausible, it type-checks, it ships, and it is wrong in a way nobody notices until production.
Diell narrows that gap by making the space of emittable programs smaller than the space of wrong ones. The language is constrained by a formal grammar at generation time, so a syntactically invalid program cannot be produced at all — not caught later, but impossible to write. A deterministic simulator then executes the program and checks its behaviour before anything renders.
How it works
Diell programs are a flat instruction stream rather than a tree. A model emits records and micro-ops against a fixed register file, with forward-only control flow. That shape is deliberately hostile to the kinds of mistakes models make — there is no deep nesting to lose track of and no unbounded backward jumps.
The entire machine specification fits in a single file a model can load in one shot, which means teaching an AI the language does not require fine-tuning or a retrieval corpus. Every build is checked by a parity gate that hash-verifies output against the reference implementation, so a stale compiler cannot quietly masquerade as the current one.
What we actually measured — including what went against us
The honest result is a seam: 18 of 27 structural checks passed against 9 of 27 semantic ones. Constraining the grammar reliably fixes shape. It does not, on its own, fix meaning. We publish that gap rather than the headline it would have been without it.
Our own tokenizer audit refuted the token-savings claim we started with. That result is in the claims ledger alongside everything else, tagged as measured rather than cited. It is the part of the project we are most willing to defend, because it is the part that cost us something to publish.
At a glance
- Formal grammar + deterministic simulator — invalid programs cannot be emitted
- Measured 18/27 structural vs 9/27 semantic seam, published with the method
- A tokenizer audit that refuted our own headline claim — kept in the ledger anyway
- Byte-compatible parity gate: every build is hash-verified against the reference
Go deeper
- Live playground
- Language reference
- Research & claims ledger
- Machine spec (one file)
- Aether Bros — the Diell runtime demo
Frequently asked questions
Is Project Diell a real, usable language?
It is a working research language, not a production dependency. It has a formal grammar, a deterministic engine, a language reference, a live playground, and a parity gate that hash-verifies every build. It is stable enough to write and run real programs, and we publish an honest ledger of which claims are measured versus cited.
What does grammar-constrained generation actually do?
It restricts what a language model is allowed to emit at each step to tokens the grammar permits, so a syntactically invalid program cannot be produced in the first place. It fixes structure reliably. Our own measurements show it does far less for semantic correctness — 18 of 27 structural checks passed versus 9 of 27 semantic.
Can I try Diell without installing anything?
Yes. The playground runs entirely in the browser — write Diell, watch the deterministic engine execute it, inspect the register file and hex stream, and inject faults to see how it fails. Nothing installs and nothing is sent to a server.
All IONIAN Labs projects · IONIAN Digital builds AI-powered products, websites, and MVPs for companies in NYC, New Jersey, and worldwide — start a project.