The run ledger

Every full run of the evaluation, whether or not the other pages discuss it. A row is the record; the narrative pages choose what to tell. Smoke runs are not here.

Each run's folder holds the evaluation's report, its JSON, and trajectories.csv: every completed turn of every agent on the run, from Nessy's nessy_agent_turn, with the scenario, the repetition and the exception id joined in. Reader turns carry no case, so their scenario is empty. The queries that read the export are in trajectory-queries.sql beside this page's folders.

Run	Date	Nessy	Models	Passed	Agent settled	Folder
luna-070-retry	2026-10-08	0.7.0	gpt-6-luna, agent and reader	550 of 550	201	20261008-luna-070-retry
luna-070	2026-10-07	0.7.0	gpt-6-luna, agent and reader	548 of 550	200	20261007-luna-070
luna-060	2026-10-07	0.6.0	gpt-6-luna, agent and reader	550 of 550	200	20261007-luna-060

Runs before 2026-10-07 have no folder: their reports were written to target/eval-results, which a clean deletes, and only what the results and how the desk evolved copied out survives. The luna-060 report was lost the same way, hours after the run; its trajectory export survived. That loss is why this ledger exists.

The reports are written by ap-eval; copying them here is still a manual step after each run.