Back

Change detection on structured filings

Diffing a registry pull without producing noise. The rules that suppress a cosmetic edit are the product.

From Anieres

Detecting the difference between two registry pulls without producing noise is the discipline that keeps a running monitor readable. The rules that suppress cosmetic edits are the product. The comparison itself is a commodity.

A private lender running a monitor across borrower entities in three countries receives a change list on every scheduled pull. If every re-punctuated address, every reformatted date, and every whitespace change appears as a finding, the reviewer will learn to ignore the report within a fortnight. If the noise is suppressed and only material changes surface, the reviewer will read it the day it lands.

Address normalisation to a canonical form before comparison. Date normalisation to a common format. Name normalisation across transliterations. A defined list of fields whose changes are always cosmetic and never surfaced. A defined list of fields whose changes are always material and always surfaced. A residual layer where the rule is unclear and the item is escalated with the raw change attached.

Ask what fraction of raw changes are suppressed and inspect the rules doing the suppressing. A product that surfaces every change produces noise. A product that suppresses aggressively without written rules produces silence.

Written alongside work at Anieres: exposure mapping, cross-reference, and standing-report systems for private clients.