LOGICAL
ARBITRAGEExplore ARC-3

INDEPENDENT AI RESEARCH & ENGINEERING

Intelligence
is learned
in the loop.

Observe. Act. Learn. Repeat.
We build and study agents that learn through interaction—with ARC-3 as our research ground.

See the work

Real environments. Inspectable experiments.

RESEARCH ENVIRONMENTARC-3 / LS20
LS20 practice board: gray passages, a blue marker and an orange-and-blue target.
LS20 · practice environmentView guide ↗
SELECT AN ENVIRONMENT

Actual images from our ARC-3 learning guide. Teaching examples, not live agent play.

NOT JUST A MODEL.

An entire system for asking better questions.

Our research approach

01 / THE WORK

Built around the hard part:
figuring things out.

ARC-3 asks an agent to interact with an unfamiliar visual world. Our work connects the environments, the agent and the tools needed to understand what actually happened.

A / ENVIRONMENTS

Worlds worth learning.

A playable research-game platform, with game versions, development histories and human review. Synthetic games support research and training—not just a collection of demos.

Explore the games
B / AGENTS

From seeing to doing.

Visual, tool-using agents that observe a board, choose actions and work with the consequences. We investigate the harness, memory and interaction loop around the model.

Inspect a recorded action
C / EVIDENCE

Make the work inspectable.

Run inspection, replay analysis and reproducible experiments. Our internal evaluation tools connect outcomes to the exact model, configuration and recorded run.

Read the public research

02 / THE RESEARCH LOOP

The answer is not
always in the prompt.

In an unfamiliar environment, progress comes from trying something—and noticing what changes. The agent, its tools and the system serving it all matter.

  1. 01

    Observe

    Read the visual state. Identify what is known and what still needs explaining.

  2. 02

    Act

    Use a tool, try an action, test a hypothesis about how the world works.

  3. 03

    Learn

    Compare the outcome with the expectation. Carry useful evidence forward.

  4. 04

    Evaluate

    Inspect the whole run. Separate a useful change from a lucky result.

Our harness work builds on Tufa Labs’ Duck Harness, connecting agent experiments, integrations and run analysis.

RECORDED PLAY / BP35

A guess is not
a result.

Two moments from a published Astra replay show why we inspect what changed—not just what an agent said it would do.

MOVE 2 / THE FAILED TEST

A wall stays a wall.

The character is below a dark ledge. Astra clicks the plain dark wall inside the ring. The wall and character stay in place: no path opens.

BP35 before move 2: character below a dark ledge, before a wall click.
Before the wall click
BP35 after move 2: the wall and character remain in place.
After: no path opens

The published note guesses a grappling action. The visible outcome does not support that guess.

MOVE 6 / A DIFFERENT RESULT

A block opens a path.

Later in the same replay, Astra clicks a green block inside the ring. The block disappears, the character rises into the space above, and the view follows.

BP35 before move 6: character below a green block that Astra clicks.
Before the green-block click
BP35 after move 6: the green block is gone and the character has risen.
After: the character rises

There is no written explanation for this move. The frames establish the change, not how the idea was discovered.

Source: our October 5, 2026 learning-guide analysis of a published Astra replay. These are recorded frames; the descriptions are our analysis of external model play.

03 / OPEN A NOTEBOOK

Evidence over
assertion.

Go beyond the final score. Explore the games, inspect published replay analysis and see how learning from a run differs from simply ranking it.

LEARNING GUIDE

What can a replay actually teach us?

Practice images, level-by-level lessons and references back to published play.

REPLAY ANALYSIS

Read the evidence. Notice the gaps.

Our review of published Astra notes: what they explain, and what remains unresolved.

HUMAN BASELINES

A human reference, with provenance.

Dated best-known action records from published ARC Prize results—not claims of optimal play.

These resources include analysis of external models and published human results. Those results belong to their original contributors; the comparison and learning tools are part of our research work.

EXPERIMENT NOTE / SEPTEMBER 22, 2026

Less coaching.
More useful play?

We tested whether repeated instructions about how to observe, plan and act were helping a full-history agent. In timed local evaluations, removing that coaching improved overall results in both repeats against the clean-return baseline.

The tempting conclusion—“shorter prompts are better”—did not hold. Further deletions reversed the gain, and the hardest games did not improve consistently.

Scope: 25 games, seven parallel lanes and a fixed 132-minute gameplay budget per run. This historical local experiment is not a current competition score. The lesson is to compare complete runs, repeat the result and keep the unsuccessful variants.

Inspect the evidence at action level ↗

04 / UNDER THE HOOD

Better agents need
better systems.

Inference engineering is part of our ARC-3 work: context and cache behavior, host and kernel performance, and the difference between generating tokens faster and completing more useful work.

We compare serving behavior against real agent workloads, using complete runs to check whether a systems change helps an agent finish useful work.

WORKLOADAgent interaction
SYSTEMContext · serving · tools
EVIDENCETask outcomes, not token counts alone

EXPLORE THE RESEARCH

Start with a world
you haven’t figured out.

Enter the games catalog