SPECTER LABS

Addenda

Tools, datasets, and experiments that support the research.

A controller that counts its own patternDiffusion settles, the gene network resets, then counts peaks and feeds back. Below is a schematic of the twenty-cell chain used in the wiring benchmark.DiffusionResetCount peaks
Controller and twenty-cell chain · schematic

A pattern that feeds back on itself

A Julia implementation joins diffusion, a peak-counting gene network, and feedback using Catlab wiring diagrams and polynomial interfaces. A separate experiment changes which cells can exchange chemicals and compares the resulting patterns. That comparison is implemented, the injury-and-repair experiment is what we want to run next, and the numerical sweep still needs reproduction.

Explore the controller and wiring test →
K = log10blind search costpolicy cost
Blind search100
Policy10
K = 1.00
Illustration · one shared task, horizon, and cost unit

How much search does a policy save?

K compares the expected costs of an agent and a blind policy in the same problem space. This reference implementation makes the comparison executable: the moves, constraints, goal, horizon, and cost units stay fixed while the policy changes.

Explore the reference implementation →

All addenda

A-013Poly Morphogenesis

Julia implementation of the BITSEY reaction-diffusion controller as composed Poly-style systems.

A-011Design Tokens

Canonical SPECTER design language specification and multi-target token generation.

A-010Equational Theories Distillation

Sub-10KB offline decision procedure for equational implication over magmas, distilled from the 4694-law implication graph.

A-009K Semantics Reference

Compare search policies within the same problem space, horizon, and cost units.

A-008Typst Field Manual

Typst template suite for Specter field manuals and public papers.

A-007Lean Corpus Extractor

Lean-native extractor for deterministic theorem-item JSONL used in Wonton Soup corpus workflows.

A-004Lean Tactic Representation

Algebraic, round-trippable representations of Lean proof-state tactics.

A-003tinygrad Benchmarks

History-mined benchmark harness for reproducible tinygrad contribution tasks.

A-002Lean Sorry Repos Benchmark

Benchmark harness for scoring and verifying Lean sorry tactic proposals.

A-001Lean Sorry Dataset

Deterministic ingestion pipeline for SorryDB-compatible snapshots and benchmark indexes.