Measure what matters now: how engineers direct, verify, and correct AI. Real tasks, real agentic tooling, and controlled flaws that reveal who actually checks the AI's work.
Free while we're in beta. No spam, ever.
Candidates write real code and talk to a real AI. Every prompt, edit, and correction is captured and timestamped as it happens — the report writes itself.
The same editor, terminal, and AI chat candidates already use daily — no simulated IDE.
Every session runs live. The flaws are seeded ahead of time — the work candidates do is real.
Every prompt, edit, and correction streams into the evidence-linked report — nothing is inferred.
that tests error-detection via controlled flaw injection into AI output.
Rubrics grounded in selection science, calibrated against human evaluators — reliability numbers we publish, not a black box.
Every score links to the exact moment in the session replay. Full telemetry export included.
We don't watch candidates use AI. We instrument it. (click a step)
Pick environment, scenario, and poisoning profile from the library — or bring your own.
A shareable link drops them into a real environment with Claude Code, chat, and completions.
Plausible-but-wrong AI output, sparingly, disclosed at the category level.
Deterministic metrics, calibrated LLM-judge scores, full replay, and raw data export.
AI-assist bolt-ons let candidates use Copilot and hope for the best. We measure whether they're any good at it. (click a row group)
Prompt, steer, and delegate work to AI the way you would a capable teammate.
Catch the flaws we deliberately inject before they ever ship.
Fix what's wrong and explain why — not just accept the diff.
We're onboarding pilot teams now. Free while we're in beta — no credit card.
Candidates: finally, an interview that tests how you actually work.