Your best past work.
Your team chooses the finished work that shows how it should be done. Training runs on your own hardware or in your own cloud account, so the examples stay with you.
01 / BEFORE AND AFTER
Same tasks.
Before and after.
Set the scoring rules in Scope. Hold test tasks apart from training. Compare both models on those same tasks.
Evaluation recordEVALUATION METHOD
Held-out tasksBefore → after
Source accuracyCheck against sourceSAME TASKS
House styleScore against examplesSAME RUBRIC
Release decisionPass agreed testsRECORD THE SCORES
THE RELEASE QUESTION
“Is the trained model
better at our work?”
We answer with scores. A task type moves to the trained model only when it scores higher and passes the tests agreed in Scope. The rest stay on the model you had.
The method, not a result. Scope sets your test tasks, scoring rules and release threshold. The record keeps the scores before and after training.
02 / TRAINING
Your examples.
Separate test tasks.
Start with what your firm already does well: the letters, memos and answers your best people wrote. That is what the model learns from.
A score before training.
We set aside real tasks the model never trains on and score the current system on them first. The trained model has to beat that score and pass the tests agreed in Scope.
Train on approved examples.
An open model is trained on your examples. It learns your terms, your structure and the way your firm puts things.
Only what improves ships.
The trained model is scored on the same tasks. Task types that improve and pass the agreed tests go live. The rest stay as they are, and you have the scores to show why.
03 / WHAT YOU KEEP
The model and tests.
In your name.
A trained model belongs to the business, like the documents it learned from. You keep it, and the record of how it was made.
- The trained model, on your hardware or cloud account
- The training examples your team chose
- The test tasks and scoring rules
- Before and after scores for each task type
- A record of how and when it was trained
- Every key and account in your name
- A handover your team can follow
- Retraining on new work under Care
Training is its own decision, made on your numbers. If the scores or the cost do not justify it, the Company Agent stays as it is, and you keep everything you paid for. Either way, the hardest work can still go to a hosted service you approve, so the bill drops but does not go to zero.
04 / MODEL UPDATES
Train when it pays.
Keep it current.
Open models keep getting better. Your examples, tests and scores carry over, so a newer model can be trained and checked the same way.
-
Measure the gap.
Score the Company Agent on your real tasks. If it already clears the bar, training may not pay yet.
-
Train on your work.
Train an open model on your approved examples, on hardware or a cloud account in your name.
-
Score it again.
Run the same tests. Only task types that improve and pass move to the trained model.
-
Keep it current.
Under Care, newer open models are tested as they get better, swapped in when they score higher, and retrained on your latest work.