Specter Labs

Addenda

The actual Catlab wiring diagram for the ten-cell reaction–diffusion phase
A-013 / Poly Morphogenesis

A pattern that feeds back on itself

A reaction–diffusion pattern, a gene network that counts its peaks, and the feedback between them form one controller. The diagram is the wiring the code exports through Catlab.

The comparison of wirings runs today. The later numerical sweep has not been reproduced yet, and the injury and repair experiments are still proposals.

Explore the controller →
Two source-generated paths on a seven-by-seven grid: random walk and shortest-path policy
A-009 / K Semantics Reference

How much search does a policy save?

The start, goal, legal moves, and cost unit stay fixed. A blind walk and a shortest-path policy take different routes through the same problem.

The paths are two seeded replays. K = log₁₀(τ_blind/τ_agent) comes from a separate exact calculation that caps each walk at 500 steps, so its value of 1.22, about 16.6 times less expected cost, is a lower bound for this grid.

Explore the executable comparison →
A-010 / Equational Theories Distillation

A countermodel you can inspect

We condensed the known implications among 4,694 equational laws into a short sheet a model can use offline. Small operation tables give explicit counterexamples when one law does not imply another, and rules keyed to the starting law prove some of the cases where it does.

XOR’s two-element Cayley table: commutativity holds, while idempotence fails at one
An exact kernel rule excerpted from the shipped equational decision sheet

More addenda

The real web color palette with IBM Plex Serif and Berkeley Mono type specimens
A-011Design Tokens

The colors, type sizes and spacing used by this site, our papers and our emails, written once in TOML and generated into CSS, Typst and email styles.

The title and abstract of the actual Typst paper example, rendered by Typst
A-008Typst Field Manual

Typst templates that typeset our papers and field manuals in the site’s fonts and colors.

Nat.add_comm and the extractor’s documented corpus record
A-007Lean Corpus Extractor

Pulls theorems out of Lean libraries into the fixed, reproducible lists that Wonton Soup searches over.

A recorded Wonton goal and its next tactic in the training-pair format
A-006Lean MCTS Learning

Trains small models on Wonton Soup search logs to score proof states, rank tactics and suggest the next one.

The actual Panel and LineChart API alongside the four-point example series and interval from its README
A-005Specter Viz

Rust chart and panel components for our research dashboards, built with egui and compiled to WebAssembly for the browser.

The toy interpreter’s actual nested constructor/apply program
A-004Lean Tactic Representation

A research plan and a small interpreter for writing Lean tactics as parts that combine and convert back without loss.

The LLM evaluation protocol: a historical issue and pre-PR code enter an isolated, git-free offline workspace; the model patch is tested and compared with the human patch from the merged PR, held outside the workspace
A-003tinygrad Benchmarks

Coding tasks taken from merged tinygrad pull requests, each asking whether a model can write a patch that passes the check the original change passed.

A checked Lean proof swaps the two components of a conjunction, with crossing proof wires; exact h is rejected and the reordered pair closes the goal
A-002Lean Sorry Repos Benchmark

Asks a model for the next tactic at an unfinished step of a real Lean proof and checks the answer with Lean.

A real dataset test fixture: x : Nat, goal x = x, and its index metadata
A-001Lean Sorry Dataset

Turns snapshots of SorryDB, a catalogue of unfinished proofs in public Lean projects, into a fixed benchmark index.