The part that changes the industry
The narrative layer can be any model. The guarantees don’t move.
Everything decisive — the fact gate, the classification, the binding, the computed conclusion, the verification gates — is code outside the model. It operates on the model’s output, which is all any AI vendor’s API gives anyone.
Pick the model
The same engine governs a local model and a frontier model identically.
The model writes the prose. The boundary decides what is allowed to land. Swap one for another and the walk below is unchanged, line for line.
Narrative layer by
- Local model — private. The weights sit on your machine and nothing leaves the building.SAME GATES
- Claude, from Anthropic, over the public API.SAME GATES
- ChatGPT, from OpenAI, over the public API.SAME GATES
- The model already inside your platform, driving the narrative layer where it is.SAME GATES
Fact gate · closed-menu classification · exact-coordinate binding · computed conclusion · cite-existence · source-support · reference resolution · quarantine — unchanged.
Ships with a model
It comes with a model, so the install is complete on day one.
Apodicta ships with a pinned, tested open-weight model: the same weights, the same version, on every machine. That is what makes an install identical and reproducible — the behaviour you see in a pilot is the behaviour the next office gets, because nothing underneath it drifted.
It works with yours too. If you already run a model, or you want Claude or ChatGPT driving the prose, point the boundary at it. The gates are the same code either way. Which of those to run, for an office holding client files, is the question worked through inclient data, local models, and why on-prem is the default for defense.
Proven live
Proven live, August 2026.
The identical engine ran on Claude Fable 5 over the public API and imposed identical discipline: zero off-menu answers, nulls held, and when Fable 5 was deliberately baited into citing a source the record couldn’t support, the gates caught it. The same external boundary has since governed frontier models from more than one independent lab identically — the enforcement lives in the code around the model, so it never depends on whose model it is.
It isn’t only law
The same logic works wherever truth has to be earned.
The relevance and relationship rules don’t know they’re doing law. Point them at any domain and they do the same thing: decide what bears on what, connect what’s genuinely connected, and refuse to assert what isn’t.
One engine, unrelated fields
Law, medicine, machinery, epidemiology, seismology, finance. The domain is swappable data, never code — which is why a new field is a new library, not a new product.
Medicine, diligence, research
Labs and records, a data room, a corpus of papers — anywhere a conclusion has to be tied to a source and a proof, the same gauntlet applies. The domain is a swappable library; the guarantees don’t move.
For AI and platform teams
An enforcement boundary you can put around your own model.
Because it operates on output, it does not need your weights, your training pipeline, or a change to the model you have already chosen. A connector and an API are in progress. If you want to see the walk run on something of yours, tell us what you would gate and we will show you the gates firing on it.
The boundary is patent-pending, a family of seven U.S. provisional applications filed 2026.