worked cases

Four cases. What the problem was, what we built, what it cost us to learn.

We would rather show you four problems we understood than a list of things we can do. Each case states what constrained it, where a person stayed in the loop, and what we would change if we started again.

why these four

Four problems that fail differently — and a general-purpose model treats them as one.

Three of the four exist because Arabic-language institutional work is not served well by systems designed elsewhere. A sermon, a statute, a classical text and a bilingual board meeting each fail differently under generic translation — and each failure has a different cost. Fidelity to register, to legal meaning, to source and to intent are not the same problem, and a single general-purpose model treats them as though they were.

That is the work. Not language coverage. Fidelity under conditions where being approximately right is not useful.

how the cases are written

Every case answers the same six blocks.

The blocks are fixed:

  1. The status of the work.
  2. The problem as the user meets it.
  3. The constraint that shaped what was possible.
  4. What we built, and where a person stays in the loop.
  5. What changed.
  6. What we would do differently if we started again.

Uniformity is the point. It lets a reader compare one case against another, and it makes selective presentation visible — a case that skips a block has told you something. The sixth block is the one that costs us: a firm whose argument is that most of this fails owes you a paragraph of hindsight on its own work. It is also the block a competitor will not copy.

Open to inspection means the structure, not the outcome. You can hold every case to it.

where to start

Four cases, in full.

Bring us the process you would test this against. We will tell you which of the four is the closest fit.

Talk to the people who built this