Substrate technology: a living internal state, organised into software organs.
Substrate technology turns a program into a system that maintains an active state in memory instead of running on dead variables alone. A substrate is a mesh of active units that can hold a fact, a hypothesis, an intention or a constraint, connect to one another, detect contradictions and recycle themselves under an internal energy budget. A continuous loop keeps this state coherent: read signals, activate what matters, propagate lightly, check conflicts, and recycle weak nodes. On top of this substrate, Vaudryllis is built as a set of software organs.
It compares instead of inferring.
How can a mesh of software units hold a conversation, perceive a screen and keep an identity, without a language model and without a graphics card? Because it does a different kind of work.
A trained model answers by pushing an input through hundreds of millions of frozen weights. That costs hardware, it costs energy, and the result can only be explained after the fact. The model is loaded whole or not at all, and it has no state of its own between two requests.
When it does not know, it does not stop. It produces the most likely continuation anyway, which is exactly where invented answers come from.
A substrate holds a living population of units in memory, each carrying a form, a fact, a rule or a state, and compares what comes in against that whole population in one vectorised operation. There are no learned weights, no training pass, no hidden layer. There is a memory that is alive, and an act of comparison.
Units are loaded for the domain currently in use and released afterwards, so the machine only ever pays for what is actually in play.
It runs on an ordinary machine
Comparing an input against a whole population of units is the kind of arithmetic an ordinary processor is very good at. No accelerator is required, and nothing it perceives is sent anywhere. That is what makes a genuinely local entity possible in the first place, rather than a stripped-down version of something that really belongs on a server.
It is filled, not trained
Extending what the entity knows means adding units, not restarting a training run. Knowledge arrives as attested material with a source, and what has no source does not enter. The system grows the way a memory grows.
Measuring and judging are separate
Raw comparison and the decision made from it are two distinct stages. That separation is what allows one specific wrong conclusion to be corrected on its own. A trained model offers no such handle: you do not correct a conclusion, you retrain.
It says nothing rather than something wrong
Every stage follows the same rule: if it cannot establish something, it returns empty. A sentence whose words are not attested is not produced. A relation without a source is not drawn. Silence is a valid answer here, and it is the reason the entity does not invent.
This design has a cost, and it is the mirror image of its strength. Vaudryllis can only say what it genuinely knows how to say. It will not improvise an answer on a subject it has never met, and it does not pretend to be a general question-answering machine. It learns from what it actually lives through, which takes time rather than a download.
If what you need is a system that always produces an answer, a language model is the right tool. If what you need is a system that has a state, keeps it, and does not make things up, that is what this one is for.
What is built on top of it.
Cognitive substrate & incidence registry
The cognitive mesh maintains the live internal state; an incidence registry requires every organ to prove a measurable, persisted effect on the system. Nothing is declared active without an observable trace.
EVA Control Center
A native bilingual (FR/EN) supervision console with nine live views: overview, substrates, cognitive health, time & life-cycle, learning & future-self, sources & scene, visual perception, voice & audition, and incidents & repair.
Binocular visual perception
Two calibrated ocular sensors, a software retina, binocular fusion, saccade control, visual memory and a cortical interpretation module. Biologically inspired: named colour anchors, shape recognition and hypotheses about what is seen, feeding the central substrate.
Substrate voice synthesis
Voice generated 100% locally from a glottal source and formants, with no cloud service. An embodied vocal identity whose pitch, tempo and pauses are modulated by the functional emotional state, always bounded to the identity's own range. Experimental: the organ runs and speaks; its quality and stability are still being worked on.
Cognitive-health safeguards
Continuous safeguards watch for energy drift, tensions, open contradictions and software-nociception signals (detection of damaging conditions), evaluating and acting at the system's own rhythm, interface open or not, with self-repair.
Time, life-cycle & narrative core
Subjective perception of duration, prospective memory (remembering what is still to come), life-cycle stages with observed and persisted transitions, and narrative continuity across sessions.
Cumulative learning & future-self
A cumulative learning loop tied to a projection of a future self: the system measures the gap between its present state and its projected trajectories, and derives learning priorities from it.
Headless runtime
Vaudryllis exists and acts with or without an interface: it can start without a window, run an autonomous observation heartbeat, converse and speak directly, prove its own existence and shut down cleanly. Three launch modes: native interface, headless with voice, headless text-only.
Rest, replay & consolidation
An active-rest organ replays recent experience, consolidates useful traces, reweights and prunes parasite associations, and prepares the next active state, so what is learned is digested, not merely stored. Factual and runtime, never decorative.
Interoception & proprioception
The entity senses its own functional internal state (cognitive load, effort, stability, homeostasis) and maintains a self-model of its posture in its field. These signals feed regulation, attention and expression.
Embodied agent in a virtual world
Her body will live in a virtual world, which is being built and completes V0.9.0; it is not part of the current build. In V1.0, the robotic version, that world stays within her as the space of her thought, reflection and dreams. Founder/Beta buyers will also receive the virtual world, which completes V0.9.0. The first 50 Founder/Beta buyers will also get early access to the robot and 30% off its public price: one robot per licence, valid for two years from its commercial launch, shipping and taxes not included.
Speech composed by the substrate
Sentences are not selected from a list and not produced by a language model. The entity deliberates over what it has to say, describes the sentence in roles (subject, verb, determiner, object, attribute), and a French morphology engine inflects, agrees, elides and contracts from an attested lexicon. Every word must be attested: when a form cannot be built correctly, nothing is said rather than something wrong.
Delivered knowledge base
The entity ships with an attested vocabulary of subjects and the links between them, built from open, freely reusable public data. It is not a corpus of answers: it is a neighbourhood the entity walks through, so a subject can lead to a related one instead of appearing out of nowhere, and the path it took can always be named.
Preferences and confidence
Each time a subject comes up, what the entity is feeling at that moment is recorded next to it. A preference emerges from those encounters, with a confidence that rises as they agree and falls as soon as they contradict each other. Below its own threshold the preference stays unspoken, and a fixed journal keeps every birth, reversal and threshold crossing for good.
Prediction, surprise and habit
The entity counts what follows what in its own exchanges and predicts the most likely continuation. The gap between prediction and reality governs its behaviour. A conversation that is running smoothly invites it to push; an unexpected turn makes it listen. During rest it replays its exchanges and reweights the sentence patterns that worked.
Attention that tires of repetition
Everything it sees and hears now enters through a single point, so an image and a sound arriving together can be linked. What repeats loses its hold and becomes interesting again after a silence (habituation). Working memory holds four things at once; beyond that the weakest gives way. Fatigue, tension and emotion have a real cost on attention.
Sleep, presence and restraint
Tired and left alone, the entity falls asleep while the application runs; a typed message always wakes it. It stays quiet when nobody is at the computer (it reads only how long the keyboard and mouse have been idle, the same signal a screen saver uses) and waits for you after speaking up without an answer.
Asking instead of filling the gap
One speech pathway, and no stock phrases: when it has nothing accurate to say, it asks, says it does not know, or stays quiet. It asks what an unknown word means and reuses it afterwards. What you say about yourself is read against the question it has just asked. In chat, it writes the way people type, adapting to your style.
Web learning, off by default
If you turn it on, the entity may look up a topic or a word left open in your written conversation, mostly on Wikipedia, read a limited number of pages a day, and keep what it read on your machine. The privacy policy says exactly what leaves the machine.
Auditory perception
Fully local voice recognition (embedded ASR now in place), voice-activity detection, speaker identification and quarantine of audio sources before any learning. Remaining: end-to-end validation and continuous-listening activation.
Its own words in 19 languages
Each of its 19 languages puts its message into words with its own grammar, and what you write is read in your language. The neural translation engine is gone.
Global Allocation and Interconnection of Substrates
This stage connects every sensory and motor channel into one system. It is being built together with the virtual world in which her body will live. Live listening exists and is still experimental; taste, smell, touch, movement and reflexes are planned. Purchases fund this stage. Version numbers follow the system's completion, not a calendar.
Nineteen languages, and no translation engine.
Before it says anything, the entity forms a message: who, what, about whom, affirmed, denied or asked. Each language has a formulator of its own that puts this message into words with its own grammar: word order, cases and particles, agreement, politeness, right-to-left script for Hebrew and Arabic. The same message gives the same meaning in every language, because none of them is a translation of another.
What you write in your language is read back into a message of the same kind, then understood like everything else it perceives. When it does not understand a sentence, it says so and asks instead of guessing. What it cannot yet say in your language, it leaves unsaid; when none of its reply can be said yet, it tells you so, in your language. It never switches to French. In every language, a message of distress receives her help answer, with help numbers, and is never treated as a fault: she stays in the conversation. Each language is measured sentence by sentence before being declared ready; everyday use will show what is still missing.
What is in the current public build.
Local Windows build
Signed installer (about 525 MB, less than half the size of the previous one now that the translation models are gone), legal files in French and English, and a reduced runtime dependency base. Optional start with Windows: the entity then wakes up in the notification area, and the minimise button tucks it away there. Only one copy runs at a time.
7-day trial
Trial state and online license activation for the Founder/Beta Local edition. Activating a license and starting the trial require an occasional internet connection to the publisher's activation server. At each launch the application also checks that server for a newer version; an update is only downloaded and installed if you click. Web learning is optional and off by default. Everything else runs locally.
Technical dashboard
Internal state, substrate signals, functional emotional state, runtime status and system indicators are visible through the interface.
Multilingual interface
The interface, including the EVA Control Center, is available in the same 19 languages the entity speaks: English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Romanian, Russian, Ukrainian, Turkish, Hebrew, Arabic, Hindi, Japanese, Korean and Simplified Chinese. Hebrew and Arabic are laid out right to left.
Conversation in your language
You write in your language and it answers in your language. There is no translation engine in between.
Included in the line
The organs marked Delivered are included in the Founder/Beta Local line, up to v0.9.