latent-anything

Latent Understanding, Manipulation & Execution Network

A Python framework that treats latent space as a first-class object: load latent representations from models, inspect them, manipulate them, and execute small runtime pipelines.

Beta Scope

0.1.0-beta.1 is a pre-1.0 core-framework beta. It includes:

• Core primitives: LatentSpace, LatentValue, and Trajectory
• Layer A introspection methods: PCA, UMAP, and SAE
• Layer B manipulation methods: Lerp, SteeringVector, and ActivationPatch
• Built-in adapters: VAE, RandomProjection, HiddenStateAdapter, and GaussianRendererAdapter
• Registry/config construction for built-ins
• Concrete analysis and manipulation pipelines
• First runtime helpers: batching, in-memory cache, async wrappers, and profiling hooks
• Script-level demos and tracked release artifacts

This beta does not claim the full Latent Anything thesis is implemented. Probing/TCAV, clustering, feature attribution, trajectory similarity, rollout, planning, discrete latent adapters, streaming runtime, external plugin discovery, and interactive visualization remain future work.

APIs are still pre-1.0 and may change under normal 0.x SemVer expectations.

Registry configs now use adapter, analysis, and intervention kinds. The beta method_a and method_b spellings remain supported with a migration warning until 0.9.0; run uv run python scripts/report_config_migration.py <config.json> to inspect repository-owned JSON configs without rewriting them.

Installation

The package is not published yet. Clone and install locally using uv:

git clone <repo-url>
cd latent-anything
uv sync --locked

Quick Start

import numpy as np

import latent_anything
from latent_anything import LatentSpace, Trajectory
from latent_anything.methods import PCA

print(latent_anything.__version__)
# 0.1.0b1

rng = np.random.default_rng(42)
space = LatentSpace(dim=4)
points = rng.normal(size=(12, space.dim))
trajectory = Trajectory(data=points)

pca = PCA(n_components=2)
projection = pca.fit_transform(trajectory.to_numpy())

print(projection.shape)
# (12, 2)

Demos

Start with the release demo index:

• Release demo index
• Showcase summary
• Showcase plot
• Async runtime summary

Representative scripts live in scripts/, including:

• scripts/end_to_end_showcase_demo.py
• scripts/end_to_end_manipulation_demo.py
• scripts/end_to_end_gaussian_renderer_demo.py
• scripts/end_to_end_async_runtime_demo.py

Release Gate

Before tagging a package release, run:

uv sync --locked
uv run ruff check src tests scripts
uv run ruff format --check src tests scripts
uv run pyright
uv run pytest

Recommended release tag:

git tag v0.1.0-beta.1
git push origin v0.1.0-beta.1

The GitHub Release workflow also accepts plain tags such as 0.1.0-beta.1, but the v prefix is recommended to keep package releases visually distinct from theory deployment tags such as theory-v*.

Project Structure

latent-anything/
├── src/
│   └── latent_anything/   # Main framework package
├── tests/                 # Pytest suite
├── scripts/               # End-to-end demos
├── artifacts/             # Demo outputs, audits, and task summaries
├── docs/                  # Architecture, theory, and sprint docs
├── latent-anything-theory/ # Standalone theory research sub-project
├── .agents/               # Agent rules, skills, and memory
├── .github/workflows/     # CI, theory deploy, and release workflows
├── pyproject.toml
└── CHANGELOG.md

Documentation

• docs/IDEA.md - Vision and motivation
• docs/ARCHITECTURE.md - Core primitives and layer design
• docs/THEORY.md - Theoretical foundations
• docs/PLAN.md - Incremental project plan

License

MIT - see LICENSE.