10 to the 23 AI logo

Field Guide / Seeing the system

ConceptConceptual framing

Records as living memory

Every record system an AI touches — case management systems, knowledge bases, caches, training buffers — is a memory that future decisions will draw on. Records hold verified truths, unverified notes, and quiet contamination side by side, and the proportions change with every write, purge, and audit.

A useful mental model is a filing cabinet with four kinds of contents: records that are contaminated, records that are clean but never verified, records that were verified and locked, and records that were verified but can drift stale as the world changes. Every write adds to one drawer; every purge, audit, and re-check moves material between them.

Two properties make record systems governance-critical. First, persistence: a person who adopts an error might be corrected tomorrow, but a contaminated record can sit quietly for years before someone — or something — reads it and acts on it. Second, amplification: a record is read many times. One bad write can seed many bad decisions, and in retrieval-augmented systems it can seed the model's own future outputs.

There is also a trap here that intuition sets for well-meaning governance: deleting records without regard to their content can make contamination worse, not better, because blind purges remove the benign material that dilutes the bad. Only content-aware cleanup — actually verifying what is true — reliably shrinks the contaminated share. That counterintuitive result is one of the recurring lessons the PAN Lab lets you experience directly.

For institutions in human services, the record is often the product: assessments, case notes, eligibility determinations. Which is why several of the strongest governance patterns in the Practice Library are about what may be written, by whom, with what verification, and how it is labeled when it has not been checked.

This page is conceptual framing — a way of seeing, not an empirical claim. Documented real-world events appear in the Domain Atlas with citations; testable versions of these ideas live in the PAN Lab.