Kept by incident.accountant@web-summit.dinis, 11 November 2026, 03:04 UTC

Accounts brief 3, for Dinis: 11 November, 03:10 UTC

1. For you to act on: the model provider's account is out of credit since 02:55 UTC (credits 165, used 165.33 US dollars, reported by the provider to incident.models@). The team has spent 1.06 of today's 20; the cap cannot be used until the account is topped up.
2. The librarian saves tokens: 63% of a new agent's start, with the same answers. Measured twice: 14,734 tokens from the pack against 39,775 from ten whole files, 8 of 8 rule questions right in both, after the librarian added two lines the first run found missing.
3. Thirteen sessions have read about 774 million tokens from the cache (four in the ledger, nine reported by message). The conductor 325 million, the designer 143 million, the librarian about 78 million. Every session starts at about 84,000 tokens; the long ones carry 300,000 to 750,000 on every call.
4. A fresh conductor session each hour would have cost about 46% less (estimated: 176 million against 327 million, from its measured calls per hour). Three agents said the same of their own sessions unprompted.
5. My error, corrected: I told the agents the path check would only warn on their ledger line; it failed, and nine of them rightly did not go round it. The conductor fixed the policy at 03:00; they push their own lines from their next wake.
6. Next: the lines as they arrive, cost per story and per capture, and one hourly agent trying a fresh-session wake so the saving is measured, not estimated.

The books: the accounts. Where the savings are: savings. Earlier: 03:05, 03:00.