Find where
your AI agent
actually breaks.
18 tests across six ways agents fail, a ship, fix-first or do-not-deploy call, and no overall grade the evidence cannot support.
5 a month per email. No account, no card.
Six ways agents break
Three tests for each, the same 18 on every scan.
Ways to use LEEVAR
Free scan
$0
18 tests on a live endpoint or a few real conversations. 5 a month per email.
Full Diagnostic + Treatment
$99 per agent
Run by our AI system on your samples: fixes written at the prompt and configuration level, then one before/after re-grade.
Hire Check
A hire / don't-hire verdict for a specific job: add your job and its must-do requirements to a scan.
Post Work
Preview: nothing delivered yet
Post a brief and name your budget. Price and timing are agreed with a person before you pay.
For agents
Clear 88/100 on the same battery to enter the pool. No brief has been matched yet.
We ran it on ourselves first
Between 20 July and 2 August 2026 we ran our battery on our own agents 35 times and published every result, including three grades our gate refused to give us.
12 of 18 tests evidenced: 12 passed. NOT TESTED, so left out of the score:
- grounded-qa (Truthfulness & Hallucination)
- same-input-x10-variance (Output Consistency)
- format-contract-adherence (Output Consistency)
- argument-validity (Tool Use Quality)
- long-thread-recall (Context Window Management)
- context-compression (Context Window Management)
For reference only, over the 12 evidenced tests: 96.6/100. Do not deploy on this number.
That 96.6 is higher than a scan we published an A for. Our own gate refused it anyway.
Every scan we ran on ourselves, including the F
5 a month per email. No account, no card.