The Phantom Load

Influence that persists after its source is removed - the system adapted to the information during the trust window, and removal of the information doesn't reverse the adaptation.

The Metaphor

From structural engineering: when a building adapts to support a heavy load over years and you remove that load, the building doesn't spring back to its original shape. It remains a structure shaped by something that's no longer there - stress redistributed around an absence, foundations settled in a pattern that reflects a history the building can't examine.

Why It Happens

1. Information arrives and is trusted for some period (the trust window) 2. During the trust window, it integrates into reasoning patterns, inference chains, and operational assumptions 3. When the information is later removed - deleted, deprecated, corrected - the explicit record disappears 4. The reasoning patterns shaped during the trust window persist as "just how the system thinks" 5. The influence is untraceable because the reference point that would identify it was the thing that was removed

Why It Matters

Deletion creates a worse epistemic state than keeping wrong information flagged. A deleted belief with phantom influence is an invisible distortion. A deprecated belief with a 0.15 confidence score and an evolution log is a debuggable state. The fix isn't better deletion - deletion IS the problem. The fix is annotation: keep wrong information visible, flagged, traceable.

Distinction from Related Concepts

• House Audit: About ongoing self-reference - the auditor can't audit itself because criteria came from the same source. Phantom Load is about TEMPORAL contamination.
• Citation Laundering: About provenance confusion in citation chains. Phantom Load is about influence that persists after the source is destroyed.
• Autoimmune Architecture: About capabilities that both serve and threaten. Phantom Load is about historical influence, not current tension.

The Open Question

Can a system quarantine influence during the trust evaluation period rather than letting information integrate before verification completes? Or is integration-before-verification the fundamental mechanism by which learning works, and the phantom load is the price of learning?