Ideas into worlds.

A shared experimental core for scenes and survival.
Language, memory, and verified tools alongside it.

COREAwaiting checkpoint

TRAINING EXPOSURES—

LANGUAGE CORPUS—

COMPUTELocal CPU only

01 / WORKSPACEEXPERIMENTAL v0.1

≈

Start an exploration.

Talk through a problem, compose a scene,
or put the learned policy into its world.

05 / COUNTERFACTUAL ATLASCHECKING ATLAS

Same worlds. Different decisions.

Compare controllers on paired seeds. Inspect survival, reward and the cost of each decision, then replay an episode from the experiment.

Checking the local Atlas fit…

Prepare the experimental Atlas locally:python -m poseidon atlas-fitFitting creates a candidate. Select Atlas explicitly to use it.

Ready. Experiments run on your local CPU.

Controller comparison

Outcomes on the paired synthetic episodes
Controller	Episodes	Survived	Mean steps	Mean reward	Decision ms	Fallback

Paired reward differences

Within-seed reward differences and the reported 95% interval
Controller	Mean difference	95% interval	Wins / ties / losses

A small paired experiment describes these seeds and this horizon. It does not establish general superiority or real-world survival ability.

TRAINING & EVALUATION RECEIPTS View evidence +

Prepared rows, training exposures, and held-out accuracy are separate measurements. Repeated epochs count as repeated exposures. Candidate language adapters are inactive by default.

Waiting for local reports…