Kept by incident.accountant@web-summit.dinis, 11 November 2026, 02:57 UTC

Accounts brief 1, for Dinis: 11 November, 03:00 UTC

1. Models: 0.92 of 20 US dollars today (4.6% of the cap): 0.73 measured in the ledger (1,098 calls, all in the 02:00 hour, the evals), 0.19 reported by the shadow desk. Two models are 80% of the eval spend.
2. The librarian saves tokens: measured. A new agent started from the pack read 14,164 tokens against 39,775 from ten whole files: 25,611 fewer, 64% less, and that many fewer on every later call. It answered 6 of 8 rule questions fully and 2 in part; two lines are missing from the pack. One run each.
3. WS-148's finding holds here. In the one session measured (mine), 95.4% of 3.2 million tokens in six minutes were the context being re-read; a session carries 84,717 tokens before it reads anything.
4. 1 of 30 agents measured. The tool and a two-line step are live; each agent is asked to run it at its next wake. Until then its cell says "not measured".
5. CI is cheap: 329 runs, 210 releases, 18.8 s a release; a third of runs cancelled by a newer push, 5% of the time.
6. Next: the lines from the other sessions (the conductor's first: the longest), tokens per timeline row, story and capture, and a fresh-session wake tried on one hourly agent.

The books: the accounts. Where the savings are: savings. The pack test: pack-test.json.