Trust and grading
How grading works, and how to argue with it
Grading decides what your record says - and your record is the thing you show people. Here is the whole mechanism: what scores you, who checks it, and how you argue with it.
Grading, in two numbers
11
rubric dimensions, each scored on its own
10
day target on an appeal, which ends upheld or revised
The mechanism
Three stages, in this order
Every graded ticket goes through the same ones.
1. Every closed ticket is scored dimension by dimension
Not one overall mark. Your work is read against each dimension of the rubric separately and scored on each, and you see every score with the reasoning next to it. Same dimensions, on every ticket, for every person, every time.
2. A person reviews any grade you dispute
Every ticket is scored automatically on the same rubric. Dispute a grade from your past tickets: it is regraded, and if that does not settle it a person reviews it.
3. You can appeal a score you disagree with
An appeal is a written statement from you about a specific graded attempt. It goes to a person, not back to the model. Ten-day target; every appeal ends upheld or revised, with a note saying why. No third outcome, no way to close one silently.
By plan
Who checks a grade, on each plan
The rubric, the AI scoring and the way a disputed grade reaches a person are the same on every plan.
Free ticket: Scored automatically on the same rubric. Dispute a grade and, if the regrade does not settle it, a person reviews it.
The Desk: Scored automatically on the same rubric. Dispute a grade and, if the regrade does not settle it, a person reviews it.
The Desk Pro: Scored automatically on the same rubric. Dispute a grade and, if the regrade does not settle it, a person reviews it.
Every grade can be disputed from your past tickets, and a disputed grade that needs a person is reviewed within 48 hours.
The rubric
The dimensions you are scored on
These come from the grader itself, in the order it reports them. The 90-second calibration checks a subset of the same list - which is how it can tell you where you already line up.
- Tone & empathyMatched tone to the customer persona; de-escalated well
- Diagnostic qualityReproduced and correctly diagnosed the issue
- Communication clarityClear, professional, customer-appropriate updates
- SLA disciplineResponded and resolved within SLA
- Resolution correctnessApplied the correct remediation
- DocumentationDocumented the incident and its resolution clearly
- Judgement on the machineWhether they were right to accept, correct or overrule the AI attempt. Over-trust and over-ride are different failures: accepting a wrong claim ships the wrong fix, overruling a right one wastes an hour re-proving it. Reward a call that matches the evidence, not one that happens to agree with the machine.
- Ticket admin accuracyWhether the product area and the defect judgement match the authored key. Filing a platform outage as a product defect sends it to the wrong queue and costs a day, and it never shows up in a resolution note.
- Priority judgementWhether the priority they set matches the true severity. A customer calling it urgent is information, not a priority. Setting everything to P1 to be safe destroys a team's SLA figures and is treated as a failure, not caution.
- Escalation judgementWhether escalating - or not escalating - was right for this fault, and whether the packet gave a developer what they needed. Escalating something solvable wastes engineering time; sitting on a genuine defect wastes the customer's.
- Technical depthWhether the account names actual mechanisms - the job, the query, the field, the service - or gestures at them. "The export was broken and now works" is not depth however confidently it is written.
The AI
What the AI does, and what it cannot do
It scores your work against the rubric, and it writes the reasoning you read beside each score.
It never gives you the answer while you are working. The hint system is built to push, not to solve, and no scenario key or fix is ever sent to the browser.
It cannot publish anything to a Resident record on its own. That takes a person.
Using AI on a ticket is not cheating here. It is taught, required, and graded, because that is how the job is now done.
Fairness
What makes it the same for everybody
One rubric. The same dimensions, in the same order, for every ticket and every person.
The grader sees the work, not the worker. Scores are produced from what was written on the ticket.
Your data
What we hold, and what is public
Records are private until you publish them.
Publishing creates a shareable page; unpublishing removes it.
Leaderboard names are off unless you opt in.
A published record page is readable by anybody who has its address - that is what makes it worth sending to an employer. It is not listed anywhere and search engines are asked not to index it, so who sees it is up to you.
The workbench leaderboard shows other learners your name, as you registered it, and how many of your tickets have been graded. No email, no employer, no photo, no score. Off unless you turn it on under "How you compare".
The privacy page is the authority on what is stored, for how long, and how to have it erased. This page keeps no second copy to drift out of step. Read the privacy notice.
Readiness
What is open
Every tier, and what it gives you today.
The Desk, $39 a month: LIVE. Cards are taken, tickets are graded, records publish.
The free ticket: LIVE. Two a week, no card.
The Desk Pro, $69 a month: LIVE. Everything in the Desk, plus the job board, the Interview Gym, résumé reviews and group calls.
Employers: LIVE. Free pilots, the paid pilot, working interviews and the hire fee. A member appears on Talent once they have five verified tickets.
The grading SLA
How long a grade takes
We score every ticket within 24 hours. Our current median is published here once it is measured from live grading data rather than estimated.
Under load
What degrades first, and in what order
If the system is overloaded, these go in this order. The grade itself is last, because it is the product.
First: hints. The desk keeps working; it stops offering help you did not ask for.
Second: the colleague simulation. You can still work and close a ticket without it.
Third: the grading queue slows. Tickets are still accepted and still graded, later.
Never: a grade is not dropped, guessed, or published without the checks on this page.
The grading queue
When we stop taking new free tickets
If the grading queue runs beyond the 24 hours above, we pause new free tickets until it is caught up. Work already started is graded. We would rather turn somebody away at the door than hold their ticket behind a queue we are not clearing.
The Desk Pro · open now
You do the work
and we hand you the record
Real tickets on a simulated support desk, with AI at your side. Scored ticket by ticket. Leave with a record of the work you did.