72:["$","div",null,{"data-slot":"skeleton","className":"animate-pulse bg-muted rounded-none motion-reduce:animate-none h-3 w-16"}] 73:["$","span",null,{"className":"sr-only","children":"Loading leaderboard…"}] 75:T4f2,

GLM-5.3-Flash on tau2 airline

This page shows two rows from the existing Based Models tau2 airline runs. Harbor did not re-run them.

Both rows use the tau2 1.0.1 airline base split: 50 tasks, 4 trials, user simulator gpt-5.2, temperature 0, max steps 200. Tasks, policy, tools, user simulator, and grader are the standard tau2 setup.

Model	What it is	Pass^1	Pass^4
GLM-5.3-Flash	Base model, called directly. Model organization Z.ai.	82.0	66.0
GLM-5.3-Flash:Causal-Reflection	The same model, called through the Based Models API. The API keeps a per-trial record of tool outcomes and adds a stable prefix plus a short readout tail. Model organization Based Models.	87.5	76.0

Pass^2 and Pass^3 are on the Leaderboard tab. Trajectory files are linked from that tab.

The Tasks tab is empty on purpose. This dataset does not publish the 50 airline tasks as runnable Harbor tasks, so harbor run -d based-models/tau2-airline has nothing to execute.

6d:["$","div",null,{"className":"prose prose-sm dark:prose-invert max-w-none","dangerouslySetInnerHTML":{"__html":"$75"}}]