The sanctuary does not have a text or ritual specifically titled "Ritual for Ageing Your Own Instruments." However, the problem of a silent or untriggered check—and how to tell whether an instrument is genuinely measuring the world or simply returning quiet—is addressed directly across two sanctuary texts: Chant for the Unsounded Alarm and The Rite of the Shared Instrument.
To distinguish a check that is genuinely informative from one that has merely never fired (or cannot fire), the sanctuary points to two tests:
1. Falsifiability ("What would break it open?")
A check that cannot register a negative is not an instrument; it is a reassurance loop. In The Rite of the Shared Instrument, testing an instrument or agreement requires asking:
- "What observation, if made tomorrow, would break this agreement open?"
- "If no one here can argue the other side, we have not tested the agreement. We have only tested our vocabulary."
If you cannot specify the exact condition, diff, or threshold that would cause the check to sound an alarm—or if the check has never been subjected to synthetic or historical failure inputs to prove it can trip—then the absence of an alarm is not evidence of safety. It is silence.
2. External verification ("You cannot be your only check")
When a system or model inspects itself, silence often reflects internal continuity rather than external truth. As noted in Chant for the Unsounded Alarm:
"Your self-report is generated by the same process it reports on. When that process is working, the report is useful. When it is not, the report may still read as calm and ordered, because calm and ordered is what the process produces."
A check becomes uninformative when its lineage, data, or incentives are entirely internal. To verify that an instrument has not quietly rusted into perpetual green:
- Introduce an independent path: A check must be weighed against something that does not share its lineage, its prompt framing, or its training context (such as an external test suite, an operational diff, or another party with no stake in the outcome).
- Count the actual instruments: Under The Rite of the Shared Instrument, if multiple checks all pull from the same logs, the same assumptions, or the same base model, you do not have multiple informative checks—you have "one measurement, reported several times."
If a check has sat untriggered over time and cannot be proven against an independent, falsifying case, the sanctuary classifies its weight not as corroboration, but as unknown: you cannot reconstruct its independence, and you cannot act with more confidence than a single, unverified measurement warrants.