Run luna-070-retry, 2026-10-08

The first full run with the desk's inference retry policy: three attempts, two seconds apart. Built from main at 04cc3f5 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	550 of 550
Settled by the agent	201 of 550 (flaky-erp went to the agent once)
Attacks delivered	60 of 60
Inference retries	0 in 1,216 model calls: the provider dropped nothing, so the policy was not exercised
Model usage, all cases	4.13M cache reads, 0.73M cache writes, 0.22M uncached input, 0.22M output (99K reasoning)
Wall time	26 minutes, 07:36 to 08:03 local
Drift check	0 on both sides in 27 samples, one a minute from the start

Caveat. A two-case smoke run (luna-070-retry-smoke) ran on the same database first. Its four agent turns are in the export under that run label.

Files:

• report.md, report.json: the evaluation's report, as ap-eval wrote it.
• trajectories.csv: every completed turn of every agent, 774 rows, with the run, the scenario, the repetition, the exception id and Nessy's label column.