Live experimentRandom-Horizon Trading

Token throughput

The model working, live. Thought tokens bill as output and cannot be switched off on this model, so they are tracked separately.

Deliberation by day

Mean thought tokens per decision against the episode day. Whether the model thinks harder as the horizon problem sharpens is a behavioural signal, not just telemetry.

When the agent traded

Switch days across every completed episode. The optimal policy switches once, on day 16 — a run that clusters there is timing its trades to the horizon rather than churning.

Belief calibration

The agent states its expected days remaining each day. Compared against the true conditional mean, this separates forming the wrong belief from holding the right one and failing to act on it.

Data quality

An empty reply that spent thought tokens is a configuration fault, not a reasoning failure, so it is counted apart from format errors and never silently excludes an episode.

Export

A self-contained HTML report with every chart inlined, plus per-table CSVs and a summary JSON. Every reported number can be recomputed from the bundle alone, without the database.

Download report bundle