Receipts, not adjectives
The same discipline that grades a skill produces its evidence: a before/after audit, a gate verdict, and a per-run cost account. The first client was the toolkit itself.
The toolkit is its own first client
Preflight on every commit
A pre-commit hook validates every changed skill against the upload rules before it can land. The unit-test gate runs again at turn-end, so an unattended build session cannot finish red.
Drift-guarded generated trees
The plugin bundle, the knowledge wiki, the site catalog, and the tool-surface locks are generated from source and checked in CI. A stale artifact fails the build instead of shipping quietly.
Committed audit snapshots
The toolkit's own audits run against the toolkit, and the reports are committed: the ecosystem overlap map, the subagent lint, the eval ledger, and the control-loop fit review are in the repo, not a slide.
Decisions with receipts
Every hard-to-reverse choice is an architecture decision record stating the context, the decision, and what it gave up. Rejected ideas get written down too, so they stay rejected for a reason.
What a run costs
Every orchestrated build leaves transcripts, and the toolkit’s cost scripts turn them into token and dollar totals per model tier and per stage. On the hosted server, a usage ledger does the same per user. The price of the method is measured, not asserted.