Three Flows

Right. So there are three flows.

They all touch patient notes.

They do not talk to each

other. Obviously.

Flow One wakes up twice a year,

Engineer on VPN, atmosphere of fear,

Dump every table,

CSV parade,

Names get blanked

— lovely job, privacy made.

Phone number? Gone.

NHS number? Gone.

Address?

Gone.

Free-text clinical note saying exactly who you are?

Yeah,

that's coming along.

THREE FLOWS! (three flows!) ZERO FRIENDS!

(zero friends!) Everybody scrubs where somebody else ends.

One hides columns, one hunts words,

One sends the whole bloody lot to the robots.

Three flows! Built apart!

Privacy by distributed works of art.

Nobody calls, nobody checks,

Everybody's solving a different problem next.

Flow Two appeared in twenty-twenty-five,

Python scrubber, briefly alive.

Names and dates and addresses too,

The only one actually reading what humans wrote through.

One researcher. One project.

Handful of runs.

Never connected to Flow One.

And sometimes nineteen-twenty-five

becomes twenty-twenty-six.

Time itself has now failed anonymisation.

But surely Flow One

calls

Flow Two?

No. (no!)

Flow Two processes Flow One's export?

No.

(no!) The live system uses either?

Absolutely not.

FLOW THREE IS LIVE.

Clinician hits button, Patient history comes in,

Raw notes, ward, vitals,

Names if somebody typed them.

No marker pen. No scrubber stage.

Straight to OpenAI, summary comes back.

Then we save that summary into the database—

—which Flow One will

eventually

export. Un-scrubbed.

THREE FLOWS! ROUND WE GO!

One doesn't scrub what the other one knows.

One has a scrubber nobody deploys,

One's in production making AI noise.

Three flows! Same estate!

Different people, different year,

different fate.

The tools exist!

The pieces fit!

They've simply never been introduced

to each other.

So the architecture is actually quite straightforward.

Once you accept that none of it is an architecture.