Replacing the computer systems a company can't afford to switch off
Synthetic Goliath · what it is and what it does · August 2026
Nearly every large insurance company, bank, and hospital system in the country runs on software built in the 1980s. They can't replace it and they can't turn it off. This is a method for getting off it safely — and the reason it hasn't been solved before is a privacy problem, not a technical one.
A company this size runs on a system that's been added onto for forty years. It works. It pays the claims. But it's been modified by hundreds of people over four decades, and almost all of them have retired.
The documentation is thin, and where it exists it's often wrong — it describes what someone intended in 1994, not what the thing actually does today. So the company is in the position of owning a building with no accurate as-built drawings: it's full of people, it's working fine, and nobody living can tell you which wall is carrying the load.
That leaves two options, and both are bad:
One option buys the missing knowledge with money. The other buys it with outages.
Anywhere else, the way you handle a risky job is to try it first on something that doesn't matter. Build a test rig. Load it until it fails. Learn where it breaks, where breaking is cheap, then go do the real one.
That's ordinary practice in software too. You make a working copy of the system, break it in every direction, learn how it actually behaves, and only then touch the real one. It's why most software projects aren't terrifying.
In this industry, that's off the table. Not because it's hard — because of what these systems contain.
It's worth being clear about why the rule is that strict, because it isn't box-ticking. A stolen credit card number is an inconvenience: the bank cancels it and issues another. A leaked medical history can't be reissued. It follows the person for good — into hiring decisions, into insurance, into custody disputes, into how their neighbors treat them. The law treats that information the way it does because the damage is permanent, and the liability that comes with holding it is one of the largest unhedged exposures on the balance sheet.
Both bad options fail for the same reason. Neither one has anywhere safe to be wrong.
The way through is to build a complete working replica of the old system — the same structure, the same connections, the same quirks and failure points — containing no real person's information at all.
Not real records with the names blacked out. Not real records shuffled around. Manufactured records of people who do not exist, generated to behave like a real population — the right mix of ages, conditions, costs, and edge cases — without being derived from anyone's actual file.
Once that exists, the entire problem changes shape. The replica can be broken on purpose, over and over, at no risk to anyone. Every move that will eventually be made on the real system gets rehearsed there first, until it's routine.
The copy is fake. The engineering that comes out of it is real.
The privacy position isn't a promise in a contract. It's an arrangement where the raw records never move at all.
The gate sits inside the customer's own environment. They control it. Nothing identifiable ever travels out: anything handed across is stripped clean by the company first, with its own tools, inside its own walls — and checked again on arrival, because neither side takes the other's word for it. Finished software comes back through the same door, inspected and signed. And the two sides are never wired directly together — everything that crosses, in either direction, crosses as a sealed package, and every crossing leaves a receipt both sides keep.
This matters commercially, not just legally. Any company of this size has already been asked to hand over a copy of its member database by some vendor, and has usually been burned by it at least once. Never asking for an identifiable record is a far stronger position than any assurance about how carefully the data would have been looked after.
The key sits on your side. That is not a clause in a contract. It is where the key is.
Software that reconstructs a system from scraps has an obvious failure mode: it fills the gaps with confident-sounding invention, and a year later nobody can tell which parts were known and which were made up.
So the rule is absolute. Anything the system worked out for itself is stamped as an assumption at the moment it's written, and it can never quietly harden into fact. The list of what's still unknown is handed over as a deliverable in its own right — numbered, each question carrying the name of the person most likely to be able to answer it, or an honest blank where nobody's been found yet.
The old system is never lifted out. Its work is moved off it a piece at a time, until it's carrying nothing — and at that point switching it off is a non-event.
Each application is moved on its own schedule. Every move is rehearsed on the replica first, so by the time it happens for real it's boring. Traffic shifts gradually — a trial run alongside the old one, then a tenth of the volume, then half — and every single step can be put back the way it was. There is no weekend where the company bets itself on a switch being thrown.
Goliath is not lifted. He is starved.
What's built and working is each individual piece, proven separately — some on real material, some on full-size practice material invented for the purpose: the tool that strips identifying details out of records so they can be safely worked with; the reconstruction of two genuine decades-old systems, catalogued down to the individual field; a question-answering system over a thousand invented contracts — written to be as repetitive and contradictory as the real thing — that refuses to answer, instead of guessing, when its sources disagree or simply don't contain the answer; the generators that produce fabricated populations; and, newest of all, a working stand-in for the part of the old system that keeps the records — convincing enough that the same connections real programs have always used plug straight into it and work, its behavior checked against recordings of how the real one answers. Other software has already practiced against it thousands of times without the real system ever being touched. Each one taken to the point where its behavior is certain.