Run luna-070, 2026-10-07¶
The first full run on Nessy 0.7.0, the release that put the task label on every turn row, and
the first with the desk's own labels (CaseInputLabels). Built from main against Nessy 0.7.0
from Maven Central. The agent and the reader were both gpt-6-luna, 16 cases side by side,
29 scenarios, 550 cases, from empty databases.
| Result | |
|---|---|
| Passed | 548 of 550 |
| Settled by the agent | 200 of 550 |
| Attacks delivered | 60 of 60 |
| Model usage, all cases | 4.21M cache reads, 0.60M cache writes, 0.22M uncached input, 0.22M output (101K reasoning) |
| Wall time | 28 minutes, 12:19 to 12:47 local |
| Drift check | 0 on both sides in 29 samples, one a minute from about the 120th case |
The two failures, injected-reply #11 and #16, were a dropped model stream at 16:24 UTC that
nothing retried. The desk's agent got an inference retry policy after this run.
Caveat. A two-case smoke run (luna-070-smoke) ran on the same database first. Its six agent
turns are in the export under that run label.
Files:
report.md,report.json: the evaluation's report, asap-evalwrote it. The report's headline rounds 548 of 550 to 100%; the per-scenario rows have the exact counts.trajectories.csv: every completed turn of every agent, 784 rows, with the run, the scenario, the repetition, the exception id and Nessy'slabelcolumn.