OPEN AGENT ENVIRONMENT · RESEARCH AND REPLAYS

Seven rooms. One world.
Can an agent learn to escape?

Borderland tests exploration, planning, tools, evidence, coordination and memory together. Early discoveries change later decisions. Teammates, equipment and a shared power budget matter. A blackout removes conversation history, leaving a 1,024-byte notebook.

See Beam play

Inspect real native observations, decisions, notebook contents and consequences. The failure is preserved.

Open the actual Beam replay →

See a full escape

A handwritten controller uses the allowed public interface. This is scripted solvability, not a trained-agent or Beam success.

Inspect the handwritten baseline →

Inspect learning

A small action ranker learned some early progress. Its notebook writer is engineered. No learned agent escaped.

Compare the eight test worlds →

Inspect a negative result

Local language-model fine-tuning changed weights, but initial and trained models produced no valid actions in two fresh worlds.

Read the outputs and parser failures →

What the experiments show

ConditionObserved resultLimit
Original Beam / GLM campaigns0/6 escapes each; 2.0 / 2.5 rooms on averageSmall original study; not a general ranking
Handwritten public-interface controller8/8 fresh worlds escapedEngineered controller, not training
Small learned action ranker0 → 1.375 rooms on average; 0/8 escapesWeak initial baseline; engineered features and memory
Local language-model fine-tuning102,400 adapter parameters changed; 0/2 escapes before and afterZero valid commands; two-world pilot

Play and reproduce

Play the earlier published browser campaign. This is the previous published browser version; the updated browser port and frozen native environment are separately labeled in the full package.

Download the complete current site and evidence · Research/source bundle · Original native bundle · Action-ranker supplement · Language-model negative-result supplement

Serve the extracted full-site folder with python3 -m http.server 8000 to inspect its current browser port and all linked evidence. Frozen downloads are preserved byte-for-byte. Standalone viewers embed their original records; documentation links open the respective downloadable bundle.

Built with AI coding assistance. Independent automated reviews check evidence and reproducibility; they are not a human audit. No model escaped the original native full campaign in the studies summarized here. Changed browser-study results use different conditions and remain separate. Reliably learning the full world remains unfinished work. No live inference or paid hardware is used by this static Space.