V. Machine Reasoning & Program Synthesis

Neuro-Symbolic Visual Induction & Verifiable Python Transformations

Can a machine reason like a geometer—observing only a handful of visual transformations, abstracting the underlying invariant law, and expressing it as an exact, verifiable program rather than a statistical pixel guess?

seer — Multi-Agent Program Synthesis for Abstract Visual Reasoning

Built for the Abstraction and Reasoning Corpus (ARC-AGI), geometor.seer rejects direct matrix-to-matrix neural hallucination in favor of neuro-symbolic program synthesis:

  • The Dreamer: Perceives input–output grid pairs as structured spatial objects, symmetries, enclosures, rays, and topological invariants, formulating natural-language geometric hypotheses.

  • The Coder: Translates hypotheses into executable Python transform(grid) programs, testing them against every demonstration pair and iteratively refining the hypothesis from exact pixel-level discrepancies.

seer_navigator — Terminal Trajectory Browser

A fast Textual terminal user interface (TUI) for browsing multi-step seer reasoning trajectories, inspecting synthesized Python programs, comparing predicted vs. expected grids, and diagnosing failure modes across solver runs.

seer_sessions — Longitudinal Reasoning Observatory

An empirical archive, mining suite, and static web observatory spanning 8,514 task runs, 2,939 unique visual reasoning puzzles, and 475 experimental sessions—documenting how symbolic verification, iterative error feedback, and geometric priors govern machine abstraction.