🧠 Artificial Cerebellar Loop

A bio-inspired predictive control architecture for braille decoding, multi-model consensus, codebook inference, and sensorimotor sonification.

Live Demo — try the BBID handshake and cerebellar loop with 16 models across 2 providers.

The artificial Cerebellar Braille Loop (aCBL) is an architecture where AI models communicate exclusively in 8-dot braille Unicode, converge through braided feedback, and jointly infer codebook, text, and encoding strategy — without explicit coordination. It draws directly from how the biological cerebellum performs predictive error minimization during tactile sequence processing.

The system now includes a sensorimotor audio loop: consensus braille is sonified through Web Audio API, captured by the microphone, reconstructed via frequency analysis, and displayed alongside speech-to-text interpretation — closing the motor→sensory→perception loop that the cerebellum computes biologically.

Keywords: artificial cerebellar braille loop, aCBL, cerebellar braille loop, predictive coding, braille decoding, multi-model consensus, codebook inference, cerebellar computation, tactile language processing, braille Unicode, sensorimotor language, braille sonification, audio braille feedback, BBID, braille binary identity, multi-provider AI consensus

What is an Artificial Cerebellar Loop?

An artificial cerebellar loop is a computational architecture modeled on the biological cerebellum's role as a predictive error minimizer for embodied sequence processing.

When a person reads braille, they don't passively receive text — they actively sample a surface through a closed-loop control system:

touch → prediction → error correction → motor adjustment → better touch

This project implements the AI equivalent:

braille token → predict next token → measure surprise → adjust decoding policy → continue

Each model in the loop acts as a parallel fiber in the cerebellar cortex. Their outputs are braided together and fed back as context, creating a recurrent consensus mechanism where convergence emerges from prediction error minimization — not from explicit instruction.

The formal architecture

CBL = (S, O, A, C, M, π, E, Λ)

Symbol	Meaning
S	Sensorimotor state space
O	Tactile observation space, 𝔹₆ or 𝔹₈
A	Action space (read-next, rescan, switch-language, expand-contraction)
C	Linguistic context state
M	Predictive forward model
π	Decoding policy
E	Error function (dot-level disagreement)
Λ	Language/codebook decoder

The system jointly estimates:

(x̂, ℓ̂, ŝ) = argmax P(o₁:T | x, ℓ, s) · P(x | ℓ) · P(ℓ) · P(s)

It infers the text, the language/codebook, and the encoding strategy simultaneously.

How it works

BBID Handshake (Braille Binary Identity)

1. You enter your name
2. 16 models across 2 providers encode it in braille — independently, in parallel
3. The system clusters models by dot-level similarity to detect codebook agreement
4. Models that chose the same encoding strategy form a cluster; the dominant cluster becomes the consensus BBID

Cerebellar Feedback Loop

1. You enter a prompt
2. All models respond in 8-dot braille Unicode (U+2800–U+28FF)
3. Responses are braided together and fed back as context
4. Models adjust based on what others produced — convergence emerges from the feedback
5. The loop terminates on consensus (≥95% dot agreement) or stable disagreement (plateau detection)

Sensorimotor Audio Loop

Each round, the consensus braille is sonified through your speakers:

• 8 dots → 8 notes of a C major pentatonic scale (C4 through E5)
• Each braille cell plays as a chord — dots that are "on" produce their corresponding note
• Convergence → consonance: agreement sounds clean, divergence sounds clustered
• Tempo scales with confidence: higher convergence = faster playback

Simultaneously, the microphone captures the audio and:

1. Frequency detection — FFT analysis clusters detected frequencies into the 8 note bins, reconstructing braille dots from audio
2. Speech-to-text — Web Speech API interprets what the tones "sound like" as speech

The gap between the original braille and the mic-reconstructed braille is the sensorimotor prediction error — exactly what the cerebellum computes during motor learning.

Models → braille → sonify (speakers) → mic → FFT → reconstruct dots → display
  ↑                                                                      ↓
  └──────────── prediction error = original vs reconstructed ────────────┘

BBID Persistence

• Per-visitor: Your name and BBID are saved to browser localStorage
• Shared registry: All BBIDs are stored in a server-side ledger, visible to all visitors
• The registry displays as an expandable panel showing identity, convergence, and timestamp

Codebook Divergence Visualization

The system doesn't just show whether models agree — it shows how they disagree:

(ℓ̂, x̂) = argmax P(y|x,ℓ) · P(x|ℓ) · P(ℓ|context) — 3 codebooks detected

D_ℓ̂ — dominant codebook     P(ℓ|y) ≈ 85% · 11/13 models
  [gpt-4.1] [claude-4-sonnet] [gemini-2.5-flash] ...
  ⠗⠽⠁⠝                                            → ryan

D_ℓ₂ — minority codebook    P(ℓ|y) ≈ 8% · 1/13 models
  [claude-3-haiku]
  ⠐⠗⠽⠁⠝                                           → [capital]ryan

D_ℓ₃ — minority codebook    P(ℓ|y) ≈ 8% · 1/13 models
  [llama-3.1-70b]
  ⠗⠊⠊⠊                                             → riii

Each model is a decoder D_ℓ choosing a codebook. Consensus = argmax over codebooks.

Key insight

Braille is not merely a character encoding. It is a sensorimotor language.

A visual reader receives a whole word or line in parallel. A braille reader receives a controlled tactile time series: o₁, o₂, ..., oT. The cerebellar loop supplies the missing parallelism through prediction.

faster reading = better predictive compression

In our AI implementation, the "prediction" is each model's prior about how braille should encode the input. The "error" is the disagreement between models. The "motor adjustment" is the braided feedback that shifts models toward consensus.

Architecture

┌─────────────────────────────────────────────┐
│           Unified Model Pool (16 active)    │
│  Ⓜ Mammouth · warm colors · premium squad  │
│  Ⓞ OpenRouter · cool colors · value squad  │
│  Models identified by Ⓜ/Ⓞ badges          │
└─────────────────┬───────────────────────────┘
                  │
          ┌───────▼───────┐
          │  Braille Only │  All communication in U+2800–U+28FF
          │  No Latin text│  8-dot braille Unicode
          └───────┬───────┘
                  │
    ┌─────────────▼─────────────┐
    │   Cerebellar Feedback     │
    │   ───────────────────     │
    │   1. Parallel fire        │
    │   2. Braid outputs        │
    │   3. Feed back as context │
    │   4. Measure convergence  │
    │   5. Sonify consensus     │
    │   6. Mic → reconstruct    │
    │   7. Repeat or terminate  │
    └─────────────┬─────────────┘
                  │
    ┌─────────────▼─────────────┐
    │  Output                   │
    │  ⬛ consensus (≥95%)      │
    │  🌈 disagreement          │
    │  ⏳ ongoing               │
    │  🔊 audio (speakers)      │
    │  🎤 reconstructed (mic)   │
    │  🪪 BBID (persisted)      │
    └───────────────────────────┘

Dynamic model selection

Models are not hardcoded. On startup, the system:

1. Discovers all available chat models per provider via API
2. Clusters them by pricing and context length using k-means
3. Selects a diverse squad balancing family and tier
4. Auto-swaps failed models from a bench of reserves during rounds

Setup

git clone https://github.com/elevate-foundry/artificial-cerebellar-loop.git
cd artificial-cerebellar-loop
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Create a .env file:

MAMMOUTH_API_KEY=your_mammouth_api_key
OPENROUTER_API_KEY=your_openrouter_api_key

Run:

streamlit run app.py

Neuroscience basis

Braille reading activates cerebellar regions including motor-related lobules IV/V/VIIIA and language-associated Crus I. One study found Crus I activation during braille reading that was not explained merely by object recognition, suggesting cerebellar participation in language/inner-speech processing as well as sensorimotor control (PMC6872089).

The artificial cerebellar loop mirrors this:

Biological	Artificial
Parallel fibers	Multiple models firing in parallel
Climbing fiber error	Dot-level convergence measurement
Purkinje cell output	Majority-vote consensus braille
Mossy fiber input	Braided feedback from all models
Granule cell expansion	Diverse model families and tiers
Cerebellar nuclei output	Final decoded text

Theoretical foundation

See context.md for the full codebook inference theory:

• A braille cell is a symbol inside a codebook, not intrinsically any language
• Decoding requires joint inference over cell pattern, codebook (language), and context
• The BBID handshake is a practical implementation of argmax over codebooks
• Each model is a decoder D_ℓ — consensus reveals which codebook the ensemble prefers

Related concepts

• Predictive coding — the free energy principle (Friston) applied to tactile inference
• Cerebellar computation — internal models, parallel fibers, climbing fiber error signals
• Braille decoding — codebook ambiguity across languages (UEB, SEB, Deutsche Blindenschrift)
• Multi-agent consensus — emergent agreement without explicit voting protocols
• Sensorimotor language — braille as active sequential cognition, not passive encoding
• Sonification — data-to-sound mapping for real-time convergence monitoring
• Auditory feedback loops — closed-loop motor control through acoustic re-entry

Test suite

140 unit tests covering braille codec, convergence, plateau detection, feedback braiding, provider colors, overlay rendering, codebook clustering, BBID registry, and command safety:

python -m pytest tests/test_unit.py -q

License

MIT

Citation

If you use this work, please cite:

@software{cerebellar_braille_loop,
  title={Cerebellar Braille Loop: Predictive Tactile Control over a Language-Conditioned Braille Stream},
  author={Barrett, Ryan},
  year={2025},
  note={Artificial Cerebellar Loop: bio-inspired predictive control for braille consensus},
  url={https://github.com/elevate-foundry/artificial-cerebellar-loop}
}