Training Environments
A model gets better by attempting a task thousands of times and being told each time whether it succeeded. Nobody can grade a million attempts. So the grading has to be a program that runs unattended.
Building that pair, the task and the program that scores it, is what this page is about.
An expert task and an environment are one step apart
Our experts already write problems that end in an exact, checkable answer. An environment asks for the same thing one step further along: not the result, but the work that produces it.
“Calculate the Year 3 debt service coverage ratio for this leveraged buyout.”
Did the model say 1.40? One check, right or wrong.
“Build the model that produces it.”
Is it 1.40, and is that cell a real formula rather than a number the model worked out and typed in?
That second check is the entire point. A model that computes 1.40 in its head and types it has not built a financial model, it has done arithmetic, and the moment an input changes the answer is wrong. A lab training on that is teaching its model a habit it will later have to unlearn.
Three parts
The task
A problem written by someone credentialed in the field, precise enough that a machine can decide whether it was done correctly. This is the part that does not transfer between domains, and the part an expert is paid per unit to produce.
The sandbox
An isolated container with the real tools a professional would use, and no network access. The model works the way a person would. Blocking the network matters: a model that can search mid-attempt looks up the answer, and the score stops measuring capability.
The verifier
A program that opens whatever the model produced and returns a score between 0 and 1. No human grader, no second AI acting as judge. Written once per field, it scores a million attempts unattended, and gives the same answer today as it will in a year.
Two of the three transfer. The sandbox and the record a lab receives are built once and reused in any field. The verifier is written once per field. What never transfers is the task and the reference answers that prove the verifier works, which is exactly the part that needs someone who knows the subject.
Every field invites its own shortcut
A grader that hands full marks to wrong work is worse than no grader at all, because the lab trains on it and nobody finds out until the model has already learned the wrong thing. So before anything is written, we name the shortcut that field rewards.
Memorise the coordinates instead of reasoning about the geometry.
Dimensions vary per task, and no two tasks share the same combination, so memorising does not pay.
Work the answer out and type the number into the cell.
The cell has to hold a formula that computes the right value from the inputs the task named.
List every plausible code so the correct one is somewhere in the set.
Precision is scored, not just coverage, and each record names the specific miscodes that must not appear.
Every task we write ships with deliberately broken answers that have to fail. If one of them scores full marks, the task does not ship. That check runs automatically on every change, and it is the reason we can say the grader works rather than hoping it does.
What We Don't Claim
This is a new product line and we would rather describe it accurately than impressively.
Want one built in your domain?
Tell us the task and how you would know it was done correctly. That conversation is the whole design.