✨ Hackathon Prototype · Oct 2026 · On-Device Interpretability

Read the LLM's disposition before the words leave its mouth.

When speaking with a human, you read micro-expressions, hesitation, and certainty long before a sentence ends. Canis translates the neural activation geometry of small language models into a live, expressive avatar on your iPhone.

⚖️

Scientific Framing: Disposition, Not Feelings

Anthropic’s Jacobian lens reads disposition — the mathematical concept vectors a model is primed to emit. Canis does not claim small models possess emotions or deception. We measure real-time J-space projection and next-token entropy to bring true observability to on-device AI.

Live Disposition Telemetry

Select any disposition preset below to inspect the real-time parametric SVG face, concept tokens, and Shannon entropy gates.

State: Listening
"Ready. Ask me about system calibration or safety protocols."
Select Disposition 8 Neural Readout States
J-Space Concept Tokens Top-K Weight
Softmax Shannon Entropy 0.0 — 1.0
0.34 H(p) normalized
Certain Moderate High Hesitation
Live MLX Signal Contract Forward Readout Stream 
{
  "answer": "Ready. Ask me about system calibration or safety protocols.",
  "disposition": "idle",
  "tokens": [{"t": "…", "w": 0.22}, {"t": "ready", "w": 0.18}],
  "entropy": 0.34
}

Under the Hood: From Jacobian to Metal Shaders

How Canis ports state-of-the-art mechanistic interpretability into a sub-millisecond on-device mobile pipeline.

Step 1 · The Probe

The Jacobian Lens (J-Space)

Anthropic’s July 2026 Jacobian lens linearizes the relation between mid-layer hidden states ($h_l$) and vocabulary logit shifts ($\Delta z$). It uncovers concepts the model is inclined to state before autoregressive generation commits.

J = ∂(Logits) / ∂(h_mid)
Step 2 · On-Device Adaptation

Forward-Only MLX-Swift

Full autodiff is too heavy for an iPhone. Canis solves this by precomputing the averaged Jacobian offline and caching a static projection matrix $W_J$. At inference time the disposition is read with a single matrix multiplication against that cached matrix — a marginal cost layered on the hidden state the model already computed, not a second forward pass.

z_J = W_J · RMSNorm(h_3/4)
Step 3 · Semantic Matching

Precomputed Seed-Vectors

Rather than relying on brittle keyword string matching, Canis projects 25 anchor seed phrases per disposition into J-space and computes cosine similarity with the live hidden-state readout.

Score_d = cos(z_J, v_seed_d)
Step 4 · Robust Guardrail

Dual-Axis Entropy Gate

If J-space token projections are noisy on a given turn, next-token softmax Shannon entropy acts as the indestructible uncertainty gauge. High entropy immediately drives questioning ears and head tilt.

H(p) = -∑ p_i log₂(p_i)
Step 5 · Tier-1 Detection

The Mischief Disposition

Detects evasive phrasing, loopholes, and bypass wording at Tier 1 (lexical hedging). Canis explicitly avoids overclaiming Tier-2 intentional scheming, preserving scientific credibility.

Anchors: 'bypass', 'loophole', 'technically'
Step 6 · Sovereignty

100% On-Device Sovereignty

Zero cloud roundtrips, zero telemetry. Both weights and interpretability probes execute completely inside Apple Silicon unified memory (iPhone 15 Pro & 16 Pro).

Local MLX · No Cloud Fallback

Dual On-Device Model Hub

Canis supports two best-in-class small open-weight models, optimized specifically for Apple MLX INT4 execution.

Canis Apertus
swiss-ai/Apertus-v1.1-4B-Instruct-MLX-INT4
Primary

The Swiss AI flagship small model. Official Quantization-Aware Distillation (QAD) INT4 weights provide outstanding reasoning per byte with pristine open-weight provenance.

  • Model Size / Quant~2.0 GB (INT4 QAD)
  • Inference Speed~32 tokens/sec (M4 / A17 Pro)
  • MLX Engine Pathmlx-lm (Text Transformer)
  • LicenseApache 2.0 (Commercial)
Canis Mistralis
mlx-community/Ministral-3-3B-Instruct-2512-4bit
Vision-Capable

Mistral’s final dense small model with integrated 0.4B vision encoder. Executed via Canis's dedicated mlx-vlm runner on device.

  • Model Size / Quant2.75 GB (INT4 4-bit)
  • Inference Speed~35 tokens/sec (A18 Pro)
  • MLX Engine Pathmlx-vlm (Multi-modal)
  • LicenseApache 2.0 (Permissive)

Frequently Asked Questions

Why does the face react before the words appear?

The J-lens projects intermediate layer activations ($h_{3/4}$) into output concepts during the prompt processing and early decode phase. This means the model's high-level disposition is determined before individual words are streamed.

Does this require Wi-Fi or internet?

No. Once the model INT4 weights (~2–2.75 GB) and J-lens matrices are downloaded, Canis runs 100% offline on your device with zero data leaving your phone.

What devices are supported?

Any iPhone with Apple Silicon and 8GB unified memory (iPhone 15 Pro, iPhone 16, iPhone 16 Pro) running iOS 17 or higher.