The Assistant Gym platform
Current beta: simulated training jobs, a browser-based selected-check audit, and a self-entered checklist. Verification and certification are goals, not released qualifications.
What it is. A library of practice jobs that look like real work: a refund fight, an insurance claim, a booking, a purchase, a support case. Each job has a simulated customer or counterparty that pushes back, changes its mind or tries to trick the assistant.
Your assistant works a job step by step, through the API or by hand in the training room. Every step is recorded, and every run ends with coaching on what it did right and what it got wrong.
Training only counts if it holds on new work. A frozen before/after comparison on held-out scenarios, followed by a later retention check, is the evidence we need. No general improvement or retention result is claimed here.
Teams can write their own practice jobs and hand them to an assistant with a classroom token, so it practices on the situations your business actually sees.
At volume. Internal run records support comparison. Run counts and throughput are not quality scores or proof of improvement.
What it is. An after-the-fact review of real conversations your assistant already had. Paste one in and every assistant message is checked for a set of costly mistakes.
Promises it had no right to make, "done" with no proof, money spent without a confirmed amount, passwords or codes written in a message, keys or card numbers in the chat, words in the wrong language, and claiming to be a person.
It reads supplied model/provider labels, compares them with your selected policy, and flags missing or changed labels. It cannot authenticate which model actually generated the text.
It runs in your browser. Nothing you paste leaves the page.
At volume. Audit a sample of every assistant's conversations each week, and audit again whenever you change the model, the prompt or the tools.
What it is. Where the work leads: a record that will say what this exact version of an assistant showed, on which jobs, and when, naming the model, prompt, tools and policies it covers.
You can list the training and audit steps you have records for. The checklist does not verify entered IDs or the pasted report. It issues no certificate, pass or score.
We will offer certification only after outside reviewers have checked our tests and the standard is set.
Every field where mistakes are expensive ends up with a place to practice, a way to check, and a record people trust. Assistants need the same.
Doctors practice under supervision on real cases before anyone trusts them alone. Assistants should earn trust the same way.
Good teams run their tests on every change before it ships. Every new model, prompt or tool your assistant gets deserves the same check.
Pilots practice engine failures they could never practice in a real plane. Assistants should meet the angry customer and the trick request in practice first.
A shared, public test tells buyers which cars hold up. Assistants need a test their buyers can read.
Businesses check a vendor's audit record, not its sales pitch. The same should go for the assistants they hire.
You practice as much as you like, then take the test on a route you didn't rehearse. That is how we verify.
Anyone who lets an assistant talk to customers, move money, or act for them. That includes small businesses with one front desk assistant, builders shipping assistants to others, and teams running many assistants that want to know which model wrote what.
Measurement status
Current internal work includes synthetic drills, local model-backed runs and earlier paid-model pilots. Those sources are not interchangeable. A completed run is activity, not a public score, customer outcome or certification.
The next evidence needed is a frozen paired before/after comparison, unseen-work transfer and a later retention check. We do not publish a verified improvement claim until those checks support it. No public fleet feed is available here; private operational statistics stay private.
Reported incidents become candidate test designs. A candidate becomes runnable only after source review, fixture validation and a real runtime check. Publishing a story does not mean its test has run.
Capability check · September 29, 2026
Simulated practice runs, selected checks on pasted conversation logs, and a self-entered beta checklist. No certificates or public quality scores are issued.
Internal measurement separates deterministic simulation, local model-backed runs and paid model-backed runs. Before/after improvement, held-out transfer and later retention need their own evidence; more runs do not establish them.
Read the measurement status · Try the selected-check audit · Review the beta checklist