The member who looks synthetic but isn't.

Every signal in this series is a description of what an ordinary, legible economic life looks like, and a note that synthetics fail to reproduce it. The uncomfortable truth is that plenty of real people have economic lives that are not legible in the way those signals expect. If you are not careful, the same reading that flags a fabricated customer flags them too. Pretending otherwise would make the method dishonest.

Who trips the signals without deserving it

• Gig and cash-tip workers. Income that arrives irregularly, from shifting platforms, in amounts that never repeat, does not look like the steady payroll cadence provenance expects. A rideshare driver or a restaurant worker paid partly in cash can look thin on inflow while earning a real, full living.
• Recent immigrants. Someone who arrived recently genuinely has a short domestic history, a thin file, and funding from sources a domestic model has never seen. Thin and new is exactly the provenance profile a synthetic presents, and here it describes a real person starting over.
• Cash-heavy households. Families who run much of their life in cash leave a transaction record that looks sparse and low-texture, not because little is happening, but because most of it never touches the account you can see.
• Single-purpose account holders. People who keep one account for one job, a specific bill, a savings goal, and live out of another institution, look thin and optimized for reasons that have nothing to do with fraud.

These are not rare edge cases. They are large, ordinary populations, and they are disproportionately the people a thin-file system is most likely to treat unfairly. A method that quietly punished them would not just be inaccurate; it would be aimed at exactly the members an institution should be most careful with.

Why "against its own history" is the safeguard, not a slogan

This is the real reason every signal in the series is read against the account's own past rather than a population average. A population cut asks "does this account look like a typical member," and against that question the gig worker, the recent arrival, and the cash-heavy household all look abnormal, because they are atypical, and being atypical is not a crime. Reading against the account's own history asks a different, fairer question: is this account behaving consistently with itself, or did it change character in a way its own past does not explain?

A real gig worker's irregular income is irregular in a stable way, year after year. A recent immigrant's thin file thickens along a plausible trajectory as a life is built. What the signals are actually tuned to find is not "unusual," but "inconsistent with itself in the specific way a cultivation-then-bust-out produces." Judging each account against its own baseline is what keeps an unusual-but-real member from being treated like a synthetic.

Why a human makes the final call

Even done well, this is pattern reading, and patterns produce false positives. A signal, or several signals together, should raise an account for review. It should never, by itself, condemn one. The distance between "this account's history has the shape we look for" and "this is a fabricated identity" is exactly the distance a person has to close, with context a model does not have: the local economy, the product, the plausible innocent explanation, the member's own account of themselves.

So the design commitment is simple and it is not negotiable. The signals prioritize where a human should look. A human decides what is actually there. A retrospective review that skipped that step would trade one hidden harm, synthetic losses booked as ordinary charge-offs, for another, real members quietly mistreated because their lives did not fit a template. The point of reading history carefully is to avoid both.