Cerberus DRE es un motor de razonamiento encerrado a proposito. No se conecta a internet, no sale a buscar nada por su cuenta y no toca nada afuera. Lo unico que hace es razonar sobre la documentacion que recibe, y no afirma nada sin mostrar el fragmento que lo respalda.
Cuando la documentacion no alcanza, no completa el hueco: dice exactamente que le falta y lo pide. Otros agentes de inteligencia artificial salen a buscar ese material y se lo entregan. Con esa documentacion nueva, Cerberus DRE vuelve a razonar.
Y la misma pregunta sobre la misma documentacion da siempre la misma respuesta. No porque recuerde lo que contesto antes: porque razona siempre de la misma manera. Preguntale cien veces a cualquier motor de inteligencia artificial de la web y vas a tener cien respuestas distintas. Aca, cien veces la misma.
Ese encierro no es una limitacion: es lo que lo hace auditable y confiable. No alucina, no inventa, no infiere sobre lo que no esta. Trabaja solamente sobre datos precisos.
🔎 Velo funcionando: cerberus.orvixlabs.com
Dos consultas reales, de principio a fin. Se ve leer, encontrar una contradiccion entre dos fuentes, frenarse, pedir la documentacion que falta, recibirla, volver a analizar y concluir. Cada dato que nombra marca, en ese mismo instante, el renglon del documento del que salio.
| Determinismo | El resultado no depende del azar. Depende de la documentacion. La misma consulta, hecha de nuevo sobre lo mismo, da lo mismo: mismo resultado, mismos datos, misma huella. |
| No inventa | Si no tiene el renglon que lo dice, no lo afirma, y pide la informacion que le falta. Cada afirmacion queda anclada al fragmento exacto que la respalda. |
| Sabe abstenerse | Cuando el material no alcanza, lo dice. "No me alcanza" es una salida valida, no una falla: no arriesga una respuesta para quedar bien. |
| Trazabilidad | Cada conclusion se puede seguir hacia atras hasta el renglon donde nacio. La corrida entera queda asentada, y se puede volver a recorrer. |
Para decisiones que se firman: donde alguien tiene que responder por un numero y necesita saber si la documentacion que hay respalda ese numero.
El motor no conoce el tema. La semantica de cada dominio vive afuera, asi que el mismo motor, sin cambiarle nada, razona igual sobre una conciliacion contable que sobre el kilometraje de una unidad de flota. Las dos consultas de la demostracion muestran exactamente eso.
El codigo fuente de Cerberus DRE no se publica. Es propiedad intelectual de OrvixLabs y se licencia comercialmente.
Este repositorio contiene solo esta descripcion.
Licenciamiento comercial, demostraciones y consultas: hello@orvixlabs.com
Copyright (c) 2026 Carlos Perasso / OrvixLabs. Todos los derechos reservados.
Cerberus DRE is a reasoning engine sealed off on purpose. It does not connect to the internet, it does not go looking for anything on its own, and it touches nothing outside. All it does is reason over the documentation it is handed, and it asserts nothing without showing the fragment that backs it.
When the documentation is not enough, it does not fill the gap: it says exactly what it is missing, and asks for it. Other artificial intelligence agents go and find that material and hand it over. With that new documentation, Cerberus DRE reasons again.
And the same question over the same documentation always gives the same answer. Not because it remembers what it answered before: because it reasons the same way every time. Ask any artificial intelligence engine on the web a hundred times and you will get a hundred different answers. Here, a hundred times the same one.
Being sealed off is not a limitation: it is what makes it auditable and trustworthy. It does not hallucinate, it does not make things up, it does not infer about what is not there. It works only on precise data.
🔎 See it running: cerberus.orvixlabs.com
Two real queries, start to finish. Watch it read, run into a contradiction between two sources, stop, ask for the missing documentation, receive it, analyse again and conclude. Every figure it names marks, at that very moment, the line of the document it came from.
| Determinism | The outcome does not depend on chance. It depends on the documentation. The same query, run again over the same thing, gives the same: same outcome, same figures, same fingerprint. |
| It makes nothing up | With no line saying so, it does not assert it, and it asks for the information it is missing. Every statement stays anchored to the exact fragment backing it. |
| It knows how to abstain | When the material is not enough, it says so. "Not enough for me" is a valid way out, not a failure: it will not risk an answer just to look good. |
| Traceability | Every conclusion can be followed backwards to the line where it was born. The whole run is written down, and can be walked again. |
For decisions someone has to sign off: where a person answers for a figure and needs to know whether the documentation at hand backs that figure.
The engine does not know the subject. Each domain's semantics live outside it, so the same engine, with nothing changed, reasons the same way about an accounting reconciliation as about the mileage of a fleet vehicle. The two queries in the demo show exactly that.
Cerberus DRE source code is not published. It is OrvixLabs intellectual property and is licensed commercially.
This repository contains only this description.
Commercial licensing, demos and enquiries: hello@orvixlabs.com
Copyright (c) 2026 Carlos Perasso / OrvixLabs. All rights reserved.