LatentReasoningWorks

Measure the shape of a thought. Get the abstraction back.

Thoughts have structure. Whether they form in a brain or in a model, they take a shape set by what is doing the thinking: its inductive biases, how sparse its wiring is, what it was built to prefer. That shape can be measured. Latent Reasoning Works builds the tools to measure it, to find the abstractions a reasoner forms, and to make those abstractions concrete enough to name, compare, and build on.

Research in progress

The codebases

Released as they are ready. Each carries the flavor of a problem worked on here.

manylatents

one geometry, learned through many latents

A latent has a shape, and a dozen methods will each show you a different one. Which structure belongs to what was embedded, and which to the method that embedded it? manylatents runs the methods and measures the geometry of every result with one shared suite, so the structures can be compared on the same terms.

open
source / pypi / docs
extended by manylatents-omics for genetics and genomics

manyagents

coordinate, dispatch, aggregate

A model's thought forms somewhere you cannot see from its output. manyagents talks to the models, one adapter per provider, and captures the trace and the hidden states while the thought is forming, so its structure can be measured the way a dataset's can.

opening
not yet public / github

manyruns

point it at your data, get the shape back

manyruns reads your data, runs the experiment loop, and records each step, so you can retrace how a result was found and run the experiment again.

opening
not yet public / github

Come build

Help build the tools and explore the questions behind them.