FIGURE · schema-cost
Schema cost, full load
21.8k tok
- How it was taken
- Tool-surface audit, 2026-08-26: token cost of a session that loads every tool the audit covered, summed.
- Measured
- 2026-08-26
- Reproducible
- No: taken on the private host; a walkthrough is available on request.
What the number is made of
The withdrawn companion count
The audit also printed a fleet tool total; it could not survive contact with its own members and is withdrawn in R18. The token cost is retained independently of it.
Private component. Described, not linked; a walkthrough shows it.
The counterexample
agent-memory serves 7 tools at about 121 tokens each: proof the same protocol can be cheap when a server is designed to be.
The register entry →The deferred-catalog comparison
The same audit pass measured the deferred tool catalog serving an equivalent surface at roughly 1.8k tokens. The gap between the two numbers is the finding, and both are internal measurements taken on the private host.
Private component. Described, not linked; a walkthrough shows it.
How to check it
The same pass measured the deferred tool catalog serving an equivalent surface at roughly 1.8k tokens. Both figures come from one measurement, and both are internal. The audit's fleet-total count was withdrawn (R18); the measured token cost stands as its product.
This figure's history
- R18
140 registered tools across the fleet (2026-08-26 audit) → Withdrawn. The per-server counts stand; the fleet total does not.
The audit's fleet total cannot describe the fleet it claims to cover: the switchpost server alone registered over 300 tools at the time of the audit, more than the fleet total names. What the 140 counted is no longer reconstructable, so the aggregate is withdrawn rather than reinterpreted. The schema-cost finding (21.8k tokens for a full-load session) stands as the audit's measured product. The rule this record enforces: an aggregate must survive contact with its own members.