Every challenge follows the same structured assessment. The AI mirrors prompt quality — vague in, vague out. Same budget. Same AI. Real measurement backed by session replay.
01LOGIC LIVE
No AI. The candidate writes code alone. Proves structural thinking, edge case handling, and system design from scratch. Every line is theirs — no shortcuts, no prompts.
AI OFFEDITOR EDITABLE
02AI BUILD LIVE
AI enabled. Editor read-only. Same fixed token budget for everyone. The AI mirrors prompt quality — precise request yields precise code; vague input returns minimal stubs.
AI ONEDITOR READ-ONLY
FULL ROADMAP — 4 PHASES
01LOGIC
LIVE
No AI — write the code yourself.
02AI BUILD
LIVE
AI on — direct it with clear prompts.
03DEBUG
LIVE
Find hidden bugs in AI-generated code.
04REVIEW
LIVE
Read-only comprehension questions.
FOR COMPANIES
Stop guessing. Start measuring.
Create an assessment. Send one link. Get AI-scored reports backed by session replay. See how candidates think — not just what they type.
01CREATE
Pick a challenge. Set duration and token budget. Get a unique invite link in under 30 seconds. No setup, no installation.
Challenge AI BuildDuration 30 minBudget 2,000 TOK
02ASSESS
Every interaction is logged: prompts, AI responses, code edits, terminal commands. Replay any session to see how the candidate directs AI.
Session LiveReplay FullEvidence 3-5 q
03ANALYZE
AI scores two weighted signals per session. Every claim is backed by quoted evidence from the chat log and code history. Replay any moment.
Score 0-100Signals 2 dimsFormat PDF + JSON
THE PROOF — REAL SCORE BREAKDOWN
One score. Two signals. Quoted evidence.
Every AI Build session returns an Overall score derived from two weighted signals, with 3-5 quoted evidence moments pulled directly from the session log.
SIGNAL
WEIGHT
SCORE
UNDERSTANDING
Did the candidate describe important behaviors in any words?
“Cagrex caught something I never saw in 8 years of interviewing: candidates who look great on paper burn the entire token budget producing minimal code, and candidates who look junior produce working systems in 30 minutes. Token efficiency is the new interview signal.”
AM
Ahmed Mohcin
Engineering Lead, Beta Testing Partner
Beta testing partner. First 200 developers evaluated. Real sessions, real scores, no marketing varnish.
Two weighted signals per session: Understanding (does the code behave correctly?) and Token Efficiency (how much compute was spent). Every score is backed by quoted evidence from the chat log and code history.
You receive a codebase with 3–5 injected bugs. You have 30 minutes and a 3,000 token budget to find and fix them. Scoring: 30% bugs found, 50% fixes correct, 20% tests pass. Token efficiency is tracked separately.
A comprehension test — 5–10 questions (multiple choice, short answer, code reading, security review). You must demonstrate you understand the system, not just that it passes tests. 40% comprehension, 40% accuracy, 20% depth.
Yes. The score breakdown (Understanding + Token Efficiency + Domain Competence + Cognitive Load) gives you four independent signals. Compare candidates on the same challenges. Export PDF + JSON for your ATS.
No. All evaluation runs locally in your Convex backend. The AI judge (Fable 5.1 via OpenRouter or Bedrock) only sees the final submission and scoring rubric — not your proprietary codebase.
Logic = algorithmic reasoning (token budget + correctness). AI Build = LLM-assisted feature development. Debug = find/fix injected bugs. Review = comprehension test. All four produce comparable 0–100 scores.
READY WHEN YOU ARE
Measure how developers think.
Same AI. Same fixed tokens. Real signal — not vibes. The future of developer assessment is measured, not guessed.
FOR DEVELOPERS & CANDIDATES
Take a live AI challenge.
15 minutes, free, no signup. A real AI-scored report lands in your inbox in ~60 seconds.