The first real test was deliberately narrow. Fifteen industrial tickers — ETN, HON, EMR, PH, ROK, JCI, CARR, MMM, GE, RTX, LMT, CAT, DE, UPS, FDX — and one question: which companies outside that list are connected to it through people who appear on both sides?
The input for this run was five quarters of SEC Form 4 filings, one of several source types the system ingests and the one that isolates the shared-insider signal most cleanly. A Form 4 is filed when an insider buys or sells stock; it names the person, the company and the relationship. No sector taxonomy, no corporate-actions feed, no list of spin-offs.
What came back
Four of the results were companies that used to be companies already on the list. The system has no concept of a spin-off and no parent-child field.
It found them because a spin-off inherits its parent's board. The people signing Form 4s at GE in 2022 were signing them at GE HealthCare in 2024, under the same person identifier. The corporate action left a trace in the insider filings well before it would surface in a structured feed, and that trace has the same shape as every other shared-director link.
Why this is the failure mode worth having
A relationship graph can miss things, which you notice. Or it can invent them, which you do not — until someone asks where a link came from and there is no answer.
The system did not know about spin-offs and found them anyway, because it records one narrow thing — who filed alongside whom — completely, rather than many things approximately. Anything expressible as the same person, twice comes out for free.