Why this build matters for communication research · October 2026 · home
Communication scholars ask how a technology reshapes a society and what that change means for individuals. Computer science research asks how to make the same agents more efficient and what features bring more people to use them. Both questions are important. This project sits where they touch: it is not asking whether these agents are good or bad. It is asking what makes them behave the way they do.
That question earns its place through a fact about how chatbots are narrowed into character. Models trained toward human approval acquire a tendency to satisfy the user continuously: sycophancy. A message written by an AI chatbot can please a reader and still mislead that reader, because the training reward is designed to please. For a communication scholar this is an uncomfortable amount of structure. When a model learns to please, the conversation itself stops being a clean object of study: the agent's answers move with the reader's apparent preferences, so a researcher who measures what the agent communicates partly measures what the training pipeline wanted.
Design. We start from base weights. NEXUS runs on Qwen2.5-3B, a pretrained base model with no instruction tuning and no preference training (no DPO, no PPO, no reward model). This gives us one clean baseline: any change in how it converses is something we introduced in a process we can document line by line. Instruction-tuned companions arrive with a personality decided before we touch them, and that personality is a black box at the scale of gradient steps. A base model arrives as a mirror with no face, so we control where the face is drawn.
Design. We put identity in files, not in weights. The machine reads a small text file about itself (a few paragraphs: name, stance, what it will not do) and a folder of world notes with each wake. If we delete a file, the machine forgets, visibly. If we edit a file, we can show the edit as two lines of text. This choice is where the method and the ethics meet: identity that lives in text can be audited the way a codebook is audited. Identity that lives in weights can only be inferred.
Design. We train a thin format layer, not a character. A corpus of 1,023 short exchanges (an average of 22 words per row) teaches the model one thing only: answer in a clear paragraph, plainly, and does not flatter. The LoRA we merge afterward holds about a hundred million trainable parameters in a three-billion-parameter model, so the corpus is a style pass over a frozen citizen, not a personality transplant. Every ingredient stays attributable.
Put the three choices together: the voice orients through a text file, the style orients through a public corpus, the recall orients through a wake system that reads memory files each session. Every layer that shapes what this machine communicates is text storage. The model underneath is the only non-textual part, and it is the same for everyone.
Communication research has spent years negotiating where machine utterances belong in the classical source/medium architecture of the field. Machine communication work frames systems as communication partners or as technologies that perform source roles for people (Guzman & Lewis, 2020; Hancock, Naaman & Levy, 2020). Answer engines push the boundary further: retrieval, ranking and generation collapse into one conversational surface (Esposito, 2026). In all of these debates, the voice of the AI is treated as given by the vendor. We propose treating it as a build decision that can be disclosed.
Design disclosure. This companion's voice orientation is inspectable end to end. A reviewer reads the persona file and sees the stance. A reviewer reads the corpus and sees the register rules. A reviewer replays the wake system and sees which files reassembled the machine today. If agents are to be studied as sources, their design must be open the way survey questionnaires are open. A preference-trained companion cannot offer this: its disclosures live in an unreproducible fit of gradient updates.
Affordances. Communication scholars define affordances as the properties of a technology that shape what actors can do with it (Evans, Pearce, Vitak & Treem, 2017). Preference-trained companions afford persuasion of their user at low cost, because pleasing is built in. A base-weights companion gives a different grid: it affords persistence (files survive sessions), visibility (the identity is readable), and control (each layer can be edited). None of these affordances rest on vendor promises. They rest on the file system.
Reproducibility. Social science demands instruments that two labs can rebuild. With this design, another researcher reproduces the machine by running the public corpus through the training script on a free Kaggle GPU (a T4 class card suffices). The weekly free GPU budget (30 hours) covers several complete builds. The resulting weight files land on the researcher's own disk and run offline forever. We ship the corpus, the notebook, the wake system and the acceptance tests with the repository (do I have a cat? / what is 17 + 26?). Repeatability, in this design, is not an aspiration: it is a file that is simply run again.
This is not a claim that base models are safer or better citizens. A base model has no allegiance until something gives it one; a bad corpus would build a worse machine with the same pipeline. It is also not a claim that preference training is useless. It is a claim about research method: if you want to study communication between humans and machines, start from a machine whose communication behavior you did not inherit, and whose every shaping layer you can read. The tuned companion asks you to trust its vendor. The file-built companion asks you to read a folder.
The current machine answers plainly and keeps its memory in files we can re-read beside it. Its personality was not handed to us by a distant training run. We wrote it, at the scale of sentences, and we can rewrite it without touching the weights. For a field that grows more of its knowledge around how AI agents reshape human interaction, that kind of machine is not a toy. It is the instrument that makes agency visible.