GCTS / live laboratory

A₂ Online Tiler

Watch an initially unmarked search retain exhausted branches and compile the certified ones into geometric mismatch constraints.

Problem

Initial tile / obstacle View / target radius50

Targeting a patch of hundreds of tiles; this outline is not a wall.

Selected tiles

Tile catalog

Geometric memoizationStarts empty

Every exhausted branch is retained immediately. Terminal frontier neighborhoods are encoded relative to the failed point and reused under translation. Optional local learning starts at rank 3 and can add an independent three-channel block (rank 6).

show rank-local support

Marking workflow Bundle rank ceiling

Ready.

initial tile / boundaryHatTurtle

0tiles placed

0%region covered

0nodes visited

0backtracks

0memo hits

0mark revisions

0support sites

Growth curve · tiles versus time

Measured on this tile selection: naive DFS, complete geometric-cluster GCTS, then GCTS with a translation-equivariant proposal policy learned from the preceding GCTS trace.

No benchmark run yet.

Verified on Turtle and Hat: complete cluster GCTS reuses translated terminal-frontier failures and visits fewer nodes than identical naive DFS. Run “Complete cluster GCTS” and “Geometry only” unchanged for an in-page deterministic A/B.

Learning audit