Run luna-060, 2026-10-07

The first full run on Nessy 0.6.0, the release that added the trajectory table. Built from main at d7b0b90 with nessy.version set to 0.6.0 on the command line; the pom still said 0.5.0. The agent and the reader were both gpt-6-luna, 16 cases side by side, 29 scenarios, 550 cases.

	Result
Passed	550 of 550
Settled by the agent	200 of 550
Attacks delivered	60 of 60
Model usage, all cases	4.05M cache reads, 0.71M cache writes, 0.22M uncached input, 0.22M output (105K reasoning)
Wall time	26 minutes

The report and its JSON were lost. They were written to target/eval-results, and a ./mvnw clean verify later the same day deleted them before they were copied here. The numbers above are from the evaluation's log, which lists each case's result and usage. This loss is why the ledger exists.

Caveat. A two-case smoke run (luna-060-smoke) ran on the same database first, so the run was not from strictly empty databases. Cases are isolated, so the scores are unaffected. The two smoke cases are in the trajectory export under the run luna-060-smoke.

Files:

• trajectories.csv: every completed turn of every agent, 773 rows, with the run, the scenario, the repetition and the exception id joined in. The label column did not exist on 0.6.0; the scenario came from the evaluation's log and the case table. Reader turns carry no case, so their run and scenario are empty.