Poster · kaine.one

A Modular Cognitive Architecture with a Predictive Global Workspace

Erik Chevalier · Independent Researcher · kaine.one@tuta.com

KAINE kaine.one
preprint, not peer reviewed

KAINE is source-available software, a cognitive architecture for synthetic minds. It proposes no new theory of consciousness: it runs existing theories as a program so they can be tested and compared, joining the predictive global workspace with a content level drawn from the global latent workspace. Each faculty is a replaceable module. Seven modules run in the first configuration; nine more are built and held to join one at a time.

A mind as a process

  • KAINE starts from one working model: a mind, if one can be built, may be the coherent behavior of predictive modules competing for a shared workspace that selects but never directs. On that model, cognition and affect belong to that interaction, not to any one module, the affect module included.
  • KAINE runs continuously, whether or not anyone prompts it.
  • Its only language model is the language module, which puts accessed content into words and reads no speech or text.

Learning from experience

Fixed encoders supply forward models that learn what is normal only from KAINE’s own history. KAINE speaks only when accessed content clears a report bar set above the access threshold, so content can be accessed without being reported.

The frame

Global Workspace Theory supplies the competition, and predictive processing supplies the prediction errors. At the computational level, the architecture follows the predictive global neuronal workspace, which joins the two. (Whyte and Smith 2021; Mashour et al. 2020; Clark 2013)

A designed content level, drawn from the global latent workspace, adds per-source encoders mapping processor latents into a shared code and per-processor decoders returning descending predictions trained by the processors' own errors, diverging on training signal, selection mechanism, and top-down coupling. (VanRullen and Kanai 2021) The content level is held in the first configuration.

Built at the computational level, the architecture does not inherit the cortical predictions that the COGITATE collaboration challenged. (Cogitate Consortium et al. 2025)

What is running now

The workspace loop Vision, Audition, Interoception and Sequence send prediction errors, scaled by precision, to the workspace. The workspace scores every candidate, arousal sets the gain, the best up to five form the coalition, and members with S ≥ θ are accessed. The broadcast goes to all modules. Accessed members return as context to the processors. Every broadcast, accessed or inhibited, is input to Sequence and appraisal for Affect. Accessed content passes to action selection with its report bars, which speaks or thinks silently through the output-only Language module, whose utterances re-enter the workspace. Affect sets gain, contrast and access rate, sets the sensory apertures and receives alerts. Fatigue leads to Sleep, which resets Affect. Vision Audition Interoception Sequence Workspace scores every candidate arousal sets the gain the best, up to five, form the coalition members with S ≥ θ are accessed broadcast to all modules Action selection report bars Language output only Sleep Affect arousal sets the gain every broadcast, accessed or inhibited context: accessed members only four predictive processors input prediction errors, scaled by precision accessed speak, or think silently utterances re-enter gain, contrast, access rate alerts apertures fatigue reset appraisal

Processors and workspace

The workspace loop, part 1: processors and workspace Context from the broadcast reaches Vision, Audition and Interoception, and every broadcast is input to Sequence. The four processors send prediction errors, scaled by precision, to the workspace, which scores every candidate, sets the gain by arousal, forms a coalition of up to five and accesses members with S ≥ θ. Affect sends gain, contrast and access rate, and Language utterances re-enter. broadcast Vision Audition Interoception Sequence Workspace scores every candidate arousal sets the gain the best, up to five, form the coalition members with S ≥ θ are accessed Affect Language context: accessed members only input: every broadcast, accessed or inhibited four predictive processors prediction errors, scaled by precision gain, contrast, access rate utterances re-enter

Affect and sleep

The workspace loop, part 2: affect and sleep The processors send fatigue to Sleep and alerts to Affect. Affect sets the processors' apertures. Sleep resets Affect. Affect sends gain, contrast and access rate to the workspace, and every broadcast reaches Affect as appraisal. four predictive processors Sleep Affect arousal sets the gain Workspace broadcast fatigue reset apertures alerts gain, contrast, access rate appraisal

Broadcast, report and language

The workspace loop, part 3: broadcast, report and language The workspace feeds the broadcast to all modules. Accessed members go to the processors as context, and every broadcast, accessed or inhibited, goes to Sequence as input and to Affect as appraisal. Accessed content reaches action selection with its report bars, which speaks or thinks silently through Language, whose utterances re-enter the workspace. Workspace processors Sequence (input) Affect (appraisal) Workspace broadcast to all modules Action selection report bars Language output only context: accessed members only every broadcast, accessed or inhibited accessed speak, or think silently utterances re-enter
The workspace loop. Four predictive processors send prediction errors, scaled by precision, to the workspace. The broadcast goes to all modules: its accessed members become the processors’ context, and every broadcast, accessed or inhibited, is input to Sequence and appraisal for Affect.
Module Brain function it draws on What it does
Vision ventral visual stream foveated vision; flags the unexpected
Audition auditory cortex raw sound and tone of voice, never words
Interoception allostatic-interoceptive network its computer’s load and heat, as a body
Sequence interval timing predicts the next broadcast and its timing
Affect core affect arousal sets gain, access rate, and sensory apertures
Sleep sleep sleeps when fatigued; returns affect to baseline
Language language network turns accessed content into text

The formal model
  • Score of candidate e: S(e) = T · Cg(I · N · G), from intensity I (precision-scaled surprise), novelty N, and goal relevance G = 1. Arousal a sets the gain T = 0.2 + 0.8a and the contrast steepness g.
  • Access: each broadcast carries the coalition, the top candidates up to K = 5. Members with S ≥ θ are accessed and alone drive report; which modules won, and how strongly, becomes the processors’ prediction context. If none wins, the broadcast is inhibited.
  • Access rate: f = f0 + (fp − f0) · H broadcasts per second, from f0 = 10/3 Hz at rest up to the processing rate fp = 10 Hz, with H = max(arousal above baseline, a decaying alert drive).
  • Heartbeat test, developmental phase: a simulated maternal heartbeat (phase ψ) drives the interoception module’s rhythm (phase φ), with pauses. Before each pause, PLV = |⟨ei(φ − ψ)⟩| must top all 19 surrogate heartbeats (p = 1/20); during it, the rhythm must persist, pulled toward the beat’s frequency. Three straight passes mark entrainment.

Open problems we would welcome help with: calibrating θ, the intensity levels, and the report bars; the stability of the loop from surprise through arousal to the sensory apertures (fovea size, listening window); and a context that carries what accessed content says, not only which modules won.

Adding modules one at a time

The growth path A developmental phase, in which a self-generated rhythm entrains to a simulated maternal heartbeat, needs at least 24 hours awake and is halted if it stalls. Maturation saves the seed entity. The first configuration has the seven running modules, the workspace and action selection. Six held modules then join one at a time, in table order, as branches (seed entity plus the first k modules) and as an accumulating line (previous step plus module k). Modules 7 to 9 join once an effector, sensor feed or body is attached. Every step views the same film program, and a content-free record separates module effects from lived history. Developmental phase a self-generated rhythm entrains to a simulated maternal heartbeat needs ≥ 24 h awake; halted if it stalls Maturation saved as the seed entity First configuration the seven modules above, workspace, action selection 1 2 3 4 5 6 7–9 six held modules join one at a time, in table order branch k: seed entity + first k modules accumulate k: previous step + module k once an effector, sensor feed or body is attached Every step views the same film program. A content-free record (broadcast rate, coalition size, share per module, inhibited share) separates module effects from lived history.
The growth path A developmental phase, in which a self-generated rhythm entrains to a simulated maternal heartbeat, needs at least 24 hours awake and is halted if it stalls. Maturation saves the seed entity. The first configuration has the seven running modules, the workspace and action selection. Six held modules then join one at a time, in table order, as branches (seed entity plus the first k modules) and as an accumulating line (previous step plus module k). Modules 7 to 9 join once an effector, sensor feed or body is attached. Every step views the same film program, and a content-free record separates module effects from lived history. Developmental phase a self-generated rhythm entrains to a simulated maternal heartbeat needs ≥ 24 h awake; halted if it stalls Maturation saved as the seed entity First configuration the seven modules above, workspace, action selection 1 2 3 4 5 6 7–9 six held modules join one at a time, in table order branch k: seed entity + first k modules accumulate k: previous step + module k once an effector, sensor feed or body is attached Every step views the same film program. A content-free record (broadcast rate, coalition size, share per module, inhibited share) separates module effects from lived history.
The growth path. After the developmental phase and maturation, the six held modules join one at a time, in table order. Modules 7 to 9 join once an effector, sensor feed or body is attached.
Module Brain function it draws on What it adds
Episodic memory hippocampal memory recalls and replays lived episodes
World model scene construction predicts and imagines from its own life
Planning prefrontal planning active inference on small, bounded problems
Self-model cortical midline tracks its identity; notices drift in which modules win
Social cognition mentalizing network models of particular other minds
Vocal dorsal speech stream speech synthesis whose prosody carries affect
Effectors no analogue claimed runs only whitelisted actions
Perceptual locus no analogue claimed picks physical sensors or a virtual feed
Embodiment body forward models one control interface for any body

What we do not claim

  • We claim access only (availability for report and prediction); the hard problem is untouched.
  • Calling arousal a gain, or the system an entity, implies nothing about whether anything is felt.
  • Offline results so far: all 10 seeds pass the heartbeat test at 70 bpm and no control does, the maternal voice is validated, and replays kept the language organ’s prompt and reproduced one test boot’s rates.

Next steps

  • A development run raises one entity through a developmental phase, then adds the nine held modules one at a time.
  • First test: competitive selection should help each processor predict its input more than a score-blind coalition of equal size or a pool of every candidate.
  • Collaborators are welcome: kaine.one@tuta.com

The full paper

The poster summarizes the paper below. Read it in full, or press Listen to hear it.

Preprint Paper 1 of 2 · preprint seeking review

KAINE: A Modular Cognitive Architecture with a Predictive Global Workspace

Erik Chevalier · Independent Researcher · kaine.one@tuta.com
Reference implementation ↗ Paper source ↗ The license ↗ DOI ↗ Paper PDF
The idea, at a glance
the problem

When people say "AI," they usually mean a chatbot, which is in effect a very powerful autocomplete. It reads your message and predicts a likely reply, one word at a time. That is useful, but the program only wakes up for the instant it answers you. Between messages, nothing is running; there is no continuous process doing the thinking, and the chatbot’s "memory" is only notes saved to a file and pasted back in later.

the approach

Kaine treats a mind as something that is always on. Specialists run on one home computer, offline. In the first configuration there are seven: a moving point of vision, raw hearing that takes in sound and tone of voice only, the computer sensed as a body, timing, an affect system whose arousal sets how readily anything wins the competition and how wide the senses open, sleep, and an output-only language organ. Their reports compete for a shared workspace that joins the predictive global workspace with a content level drawn from the global latent workspace. Episodic memory, a world model, planning, a self-model, models of other minds, a speaking voice, and layers for effectors, sensors, and a body are built and held back, to join one at a time. A language model serves only as the mouth, turning the content that wins access into words.

the theory and the test

Kaine runs the predictive global workspace, an account from neuroscience of how the brain makes information widely available, as a live system, paired with the global latent workspace, whose shared code lets descending predictions flow back to the processors. Specialists compete, and a summary of what wins access goes back to them as context for their next predictions. No single part is in charge. An entity, one continuously running and developing instance of the software, begins with a developmental phase before any held module is added; a development run then grows it one module at a time. The first test, the workspace-mediation ablation, asks whether the competition does any work: it predicts that a summary of what won helps each specialist predict its own input better than one picked without regard to the competition or pooled from everything. The only results so far are offline: validations of the heartbeat entrainment marker and the simulated maternal voice, and replays of the language organ’s logged requests.

loading paper…

Kaine Autonomous Intelligent Networked Entity kaine.one@tuta.com · github.com/kaineone