Keep the exact baseline.
Record the assistant configuration, job, instructions, allowed actions and the real result. No baseline, no improvement claim.
How it worksPrepare for the work.
Prepare for the work.
Learn from the result.
You define the job. Your assistant attempts it. You record the baseline, make a change and test fresh work. The browser tools examine specific material you supply; they do not run the whole loop or improve the assistant for you.
The job sets the standard.
A fluent answer tells you very little about whether the right part was selected, the booking happened or the person was authorized to approve it.
The Gym's approach starts with the outcome, the instructions and the mistakes that would make the result unacceptable.
- Choose the work.
Name the task, the people and systems involved, and the limits the assistant must respect.
- Keep the setup clear.
The model, instructions, tools and permissions all affect the result.
- Decide what counts.
Check the evidence of completion and the errors that a good-looking average can hide.
A work loop, not a badge
Find the miss. Make the change. Test again.
The company is built around one hard question: did the assistant get better at this work, or did it just sound better on the next attempt? The full connected loop is in development. For now, you run the comparison and use the bounded tools where they fit.
Fix the cause.
Change a permission rule, workflow, instruction, tool or handoff. Make the intervention visible instead of hiding what changed.
Run an unseen retest.
Test a fresh case, inspect the actual outcome and see whether the result survives beyond the original scenario.
01 / The Gym cycle
Train
Practice the difficult part before it carries a real cost.
Good practice includes missing information, a confused handoff, a failed tool, a persuasive counterpart and the moment the assistant should ask for help. The goal is useful experience with the work.
The broader training design connects a discovered weakness to a deliberate intervention: a better instruction, a corrected tool description, a playbook or another change you can identify.
A place to begin
Play through a hand-played scenario and examine the choices it forces.
02 / The Gym cycle
Test
Find out whether this assistant can do this particular job.
A test needs a defined task and a clear way to tell success from failure. It should preserve the user's instructions, check the actual result and catch mistakes that make the work unacceptable.
When you change the assistant, test the changed setup on fresh cases. Seeing an answer twice is a different challenge from solving a new problem.
A place to begin
Use a focused check on structured output or compare replies from your own workflow. You choose the cases and run the wider test.
03 / The Gym cycle
Verify
Look for the evidence behind “done.”
A claimed booking needs a booking record. An agreed refund needs a corresponding action. A product choice needs specifications that match the request. Verification asks what the available evidence actually supports.
The evidence has to fit the claim. A record of what was said can help explain a decision; system state is what tells you whether a system action happened.
A place to begin
Compare structured output against values and restrictions you supply with Decision Audit.
04 / The Gym cycle
Certify
Make a specific capability worthy of someone else's reliance.
Our ambition for certification is a bounded record of demonstrated capability: a particular assistant setup, a particular job, a defined standard and inspectable evidence.
Qualification is the independent assessment of a fixed assistant setup against a defined job standard. Certification would record a valid result and its limits. A change to the assistant or its environment may require a new assessment.
A place to begin
The readiness worksheet organizes information you enter. It helps you prepare your evidence; it does not issue a verified certificate.
05 / The Gym cycle
Audit
Reconstruct the work so the next attempt can be better.
What did the assistant know? What was it told? Which tools did it use? Where did the work drift? An audit connects the conversation, the actions and the result.
A failure can become a new practice case, a corrected playbook or a fresh test. That is how the cycle comes back to the work.
A place to begin
Inspect a supplied conversation log for the selected issues covered by Assistant Audit.
Improvement has to survive a fresh test
Find the failure. Change something that matters. See whether it holds up.
You run the baseline, make the change and retest your assistant on work it has not seen. The available checks inspect selected outputs. They cannot establish lasting improvement on their own; that takes fresh evidence across different situations and over time.
For buildersStart where the work breaks.
You may need one check, one practice case or a closer look at a conversation.