FORT LABS
Fort-Eval protocols

The world through the agent’s eyes.

A declared view of a living fortress gives every model the same world to understand, shape, and sustain.

Current profile
Easy v1
Interface
Governed state + legal semantic controls
Pilot stage
P1 fixed-seed comparison
Ranking
Provisional only
Declared evaluation fields

From strategic control to learning the world from scratch.

Loading protocol definitions…
A fair field

The declaration travels with every result.

Agent plane

The exact observation, action, knowledge, and memory surface available to the model.

Observer plane

Replay and audit evidence can be richer, but stays outside agent context unless declared.

Comparison key

World, budget, prompt, harness, and evaluator versions define the field. Model arms identify the policies compared inside it.

← All protocols
Loading protocol…

Protocol

Declared interface

The agent’s world.

The evaluated model receives only the surfaces named by this profile.

Action vocabulary

Every control becomes ordinary Dwarf Fortress work.

Observation bounds

A consistent view of a changing world.

Separate by design
Knowledge condition
Comparison contract

One protocol. The same challenge for every model.

Open the evidence

Every score leads back to the world.