● Coming soon
01 — Order
The sequencing is deliberate.
Most attempts start from an architecture and go looking for data to feed it. That is backwards where the binding constraint is the measurement.
The loop closes when the model chooses the next experiment better than we would. Each stage has to earn the next.
02 — Design
The design is the asset.
What is perturbed, at what doses, on what schedule, and what is measured in parallel — that is what determines whether anything downstream is learnable.
Measurements are designed from the outset as training data, not as an experiment that happens to be deposited afterwards.
03 — Workstreams
Four tracks, run in parallel.
-
/ 01
Generation. Designed perturbation series, read across multiple modification layers on the same material.
-
/ 02
Representation. Modification state as a quantity the model predicts, with its own uncertainty.
-
/ 03
Evaluation. Held-out, prospective, scored against baselines chosen in advance.
-
/ 04
Agents. Systems that carry the loop. The interesting part is the model earning the right to choose what gets measured.
04 — The core
Not public yet.
The architecture, the benchmark and the generation design are the three things worth protecting at this stage. They stay closed until there is something to point at.
● Coming soon
The benchmark
● Coming soon
The data engine
05 — Principles
What we hold to.
-
Measured
Every claim traces to something measured or published. Where we do not know, the page says so.
-
Baselined
A result is reported against the simplest thing that could have produced it, not against nothing.
-
Falsifiable
The thesis makes predictions that could come back negative. If they do, we publish that.