ALT Evaluation report with two expanded Luna task cards. Each tool call shows its timestamp and a clickable span ID; failed calls are labelled Error. These links open the exact operation in the trace viewer.

ALT Trace viewer opened from a failed tool call’s span link. The selected bash span appears within Luna’s investigation waterfall. Its details show the executed SQL command, exit code 1, and a GROUP BY error.

ALT Alt text: A dark purple OpenTelemetry Desktop Viewer dashboard titled “Eval results.” An amber banner marks the running evaluation as provisional and awaiting human signoff. At capture, 12 of 132 planned investigations are recorded across twelve models. Each model appears in a collapsible row with an “In Progress” badge. A short teal segment represents its first successful question; the remaining ten questions form a dark, background-coloured pending segment. Every model shows 1/11 runs and a 9.1% mean, calculated against the full plan rather than only completed answers. Controls switch between results, time and estimated Canadian-dollar cost. Below, a separate Sol 6.1 grading section reports twelve full-credit judgments. A continuous table groups results by investigator model, with visible totals and columns for grading time, estimated cost and token usage. Muted lavender text and restrained colour accents distinguish pending work from recorded results.

ALT OpenTelemetry Desktop Viewer showing Qwen3.7’s investigation of “Why did checkout-017 fail?” A waterfall connects seven agent turns, tool calls and viewer queries across roughly one minute, followed by grading. The selected investigation’s details appear alongside it, including the original question and answer: the customer’s card was declined for insufficient funds, causing checkout to return HTTP 402. The view connects the agent’s conclusion to its recorded investigation steps.