Decepticons

Website · Architecture · Kernel matrix · Examples · Related work

O(n) attention is deception. A backend-neutral kernel of predictive primitives — substrates, memory, gating, routing, readouts — that downstream systems combine into trained models without forking the kernel itself.

decepticons is the shared mechanism layer for predictive descendants: substrate dynamics, controller summaries, memory primitives, feature views, readouts, and runtime helpers extracted from a broader experiment family so downstream systems can specialize without forking the kernel.

Install

Requires Python ≥ 3.11. The kernel itself only needs numpy.

If you don't already have a Python virtual environment, make one first. Modern Linux distributions block pip from writing into the system Python (PEP 668), so a venv is the standard path:

python3 -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

Then install from PyPI:

pip install decepticons

For the model backends:

pip install "decepticons[torch]"   # PyTorch CausalBankModel + routed readouts
pip install "decepticons[metal]"   # Apple MLX backend (Apple Silicon)

To leave the venv when you're done: deactivate.

For development from source (clone + editable install + run tests):

git clone https://github.com/asuramaya/decepticons
cd decepticons
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[test]"
pytest -v

Quickstart

from decepticons import ByteCodec, ByteLatentPredictiveCoder

text = "predictive coding likes repeated structure.\n" * 64
model = ByteLatentPredictiveCoder()
report = model.fit(text)

prompt = ByteCodec.encode_text("predictive ")
sample = model.generate(prompt, steps=40, greedy=True)

print(report.train_bits_per_byte)
print(ByteCodec.decode_text(sample))

CLI:

decepticons fit --input ./corpus.txt --prompt "predictive " --generate 80

A complete worked example lives in examples/quickstart.py. For descendant-shaped projects, see examples/projects/.

What's in the kernel

Full capability matrix: docs/kernel_matrix.md.

Architecture

decepticons  ──→  chronohorn  ──→  heinrich
  kernel          runtime          forensics
 (this repo)   training · fleet   geometry · audit

Three layers inside this repo:

1. Kernel — src/decepticons/. Public package. Reusable mechanisms only.
2. Project descendants — examples/projects/. Pressure-tests the kernel boundary with concrete descendant shapes (causal · oracle · bridge · noncausal · byte-latent).
3. Tooling — examples/tools/. Development and analysis scripts. Not part of the public package.

Code moves into src/ only when all three hold:

1. It is a mechanism, not a project policy.
2. At least two descendants want the same thing.
3. The generalized API is simpler than keeping the duplication.

This rule is the main defense against turning the kernel into a renamed collection of branches. Full detail in docs/architecture.md and the boundary against the runtime in docs/chronohorn_boundary.md.

Causality is verified

All substrate modes are verified by tests/test_causality.py: it feeds two identical sequences up to position t, different after t. If logits at position t differ, causality is violated and CI fails. Modes verified: frozen, learnable_mixing, learnable_decays, selective scan augment (state_dim > 0), readout_bands, routed experts.

decepticons never imports its descendants — enforced by an AST scan in tests/test_dependency_firewall.py.

Docs

• Architecture — package map, three-layer model, promotion rule
• Kernel matrix — capability matrix
• Chronohorn boundary — boundary against the runtime descendant
• Downstream patterns — causal, noncausal, oracle, bridge, byte-latent patterns
• Related work — research anchors and prior art
• Landscape — ecosystem snapshot (March 2026)
• Lineage — source attribution
• Examples — example descendants and tooling
• Tests — verification surface

Scope

This is a research kernel and reference implementation. The current pressure from descendants is O(n) causal-bank architecture search — cheap ablation lanes to separate mechanisms before promotion, with scale and context survival checked in the descendant runtime.

It is not a frontier runtime, a production compression stack, or a benchmark claim. It exists to keep the shared mechanism layer reusable and legible.

Contributing

See CONTRIBUTING.md. Issues and pull requests welcome.

License

MIT — see LICENSE.