/safety: the maintenance-cost card — sustained deception leaves audit traces (system-claim wording, honest limit stated)

[emooreatx]

What

One new card on /safety, in the page's existing plain register ("Safety Built In / Not bolted on"), stating the deception-cost design as a claim about our system (testable, linkable) — never as a claim about nature.

Draft copy:

A kept-up false story costs effort to maintain — so we built the books to show the effort.
Every decision leaves a record that cannot be quietly changed. Because of that, sustaining a deception inside CIRIS is not one bad message — it is ongoing work, and ongoing work leaves traces in the audit telemetry even when each individual message looks clean. We test for exactly this: the test batteries are public, and the honest limit is stated with them — to date they run on ordinary traffic; the decisive adversarial test runs when the new mesh telemetry ships, and its requirements checklist is already published.

Links: the public safety test batteries (CIRISAI/CIRISAgent tests/safety, already linked from /proof), and the adversarial-readiness checklist (coherence-ratchet, experiments/adversarial_neff/trace_harness/READINESS.md).

Notes

• System-claim framing only ("inside CIRIS", "our telemetry", "we test") — the site never asserts the general principle, which lives with the research record off-site (see editorial-policy issue).
• The honest-limit sentence is load-bearing; the site's credibility register (mtBanner, measured-vs-model tags) is the house style this follows.
• 29 locales.