Full arena replay
Share a step-by-step record with engineering leads and clients: what they asked the model, what broke, and how they fixed it.
AI policies don't tell you who can think.The Constraint does.
A technical skills benchmark for the AI era. Score how people think with AI—not whether they hid a tab.
Shortlists backed by evidence, not vibes.
[01]The shift// Why constraints reveal skill
You can't prevent AI cheating. You can design it out.
In an unconstrained environment, AI compensates for gaps—signal becomes noise. Under fixed limits, only candidates who plan, prioritize, and recover actually win.
- 01
Unlimited compute hides gaps
When anyone can brute-force prompts, you can't tell who understands the problem. Caps force tradeoffs—and tradeoffs reveal judgment.
- 02
Measurement beats prevention
Proctoring measures evasion. The arena measures output and process: what they shipped, how they got there, and what broke along the way.
- 03
Same constraints, fair compare
Every candidate gets the same model, token budget, and context window. You're ranking skill—not who had a better cheat sheet.
[02]How it works// Four steps to a defensible shortlist
Put every candidate in the same arena. Review the run.
You set the role challenge once. They build with AI inside fixed limits. You get a dossier you can share with engineering and the client.
- 01Scope
Pick the challenge for the req
We help you map the role to one arena task—agent workflow, fix under a token cap, or ship a feature with a pinned model. One bar per opening.
- 02Run
Candidates work with AI—on your terms
No blockers. Same context window, token budget, and model for everyone. What varies is skill: planning, prompting, recovery, and what they ship.
- 03Review
You read the trace, not a screenshot
Hiring managers and recruiters get prompts, tool calls, failures, retries, and final output. See who thinks under constraints—not who gamed a test.
- 04Place
Shortlist with evidence
Rank side-by-side, forward traces to the client, and move to onsite only when the run backs the hire. Fewer false positives. Faster debriefs.
Sample run + dossier
Example arena session and the dossier your team reviews.
// Sample trace
> challenge: agent-workflow · token_cap: 64k · model: pinned
> step 03: tool_call failed · recovery in 2 retries
> step 07: output shipped · rubric: recovery ★★★★☆
> export: ready · share_with: eng-lead, client-ta
[03]Sample dossier// Deliverables per candidate
// Sample dossier excerptarena-run · candidate-0042
> challenge: agent-workflow · token_cap: 64k · model: pinned
> step 03: tool_call failed · recovery in 2 retries
> step 07: output shipped · rubric: recovery ★★★★☆
> export: ready · share_with: eng-lead, client-ta
Skill under constraints
Score token discipline, tool use, and recovery—not whether they alt-tabbed. AI is in-bounds; sloppy work isn't.
Ranked shortlist
Same task, same caps. Stack five finalists and pick on evidence—recruiters can defend the slate in one email.
[04]Why not the usual stack// Measure vs. prevent
Prevention is an arms race. Measurement is a standard.
Help your team argue for evidence-based assessment—not another blocker subscription.
| Approach | Proctoring / AI detection | Live coding (no AI) | Generic take-home | The Constraint |
|---|---|---|---|---|
| Measures real skill with AI in bounds | No | Partial | No | Yes |
| Comparable across candidates | No | Yes | Partial | Yes |
| Client-ready evidence | No | No | Partial | Yes |
| Scales async | Partial | No | Yes | Yes |
| Arms race with AI cheating | Yes | N/A | Yes | No |
Measures real skill with AI in bounds
- Proctoring / AI detection
- No
- Live coding (no AI)
- Partial
- Generic take-home
- No
- The Constraint
- Yes
Comparable across candidates
- Proctoring / AI detection
- No
- Live coding (no AI)
- Yes
- Generic take-home
- Partial
- The Constraint
- Yes
Client-ready evidence
- Proctoring / AI detection
- No
- Live coding (no AI)
- No
- Generic take-home
- Partial
- The Constraint
- Yes
Scales async
- Proctoring / AI detection
- Partial
- Live coding (no AI)
- No
- Generic take-home
- Yes
- The Constraint
- Yes
Arms race with AI cheating
- Proctoring / AI detection
- Yes
- Live coding (no AI)
- N/A
- Generic take-home
- Yes
- The Constraint
- No
[05]Who it's for// Hiring, trust & FAQ
If you own the req or the shortlist, this is for you.
Replace blocker tools and anti-cheat add-ons with one assessment your team and your clients can read in minutes.
Hiring Managers
Compare five finalists on the same task—with traces, not gut feel.
Technical Recruiters
Defend the shortlist to the client in one email with artifacts.
Staffing & RPO
Back every placement with a dossier the client can audit.
// Trust
Fairness
Same arena, same caps
Identical model profile, token budget, and context window per challenge. What differs is the candidate—not the playing field.
Pilot
Q2 cohort
Limited pilot slots. Scoped challenge in days. Direct support for challenge design and rubric alignment with your eng team.
Questions teams ask before the pilot.
Yes—that's the point. AI is in-bounds under fixed constraints. You measure how they plan, prompt, recover, and ship—not whether they hid a tab.
Turn off the blockers. Turn on the arena.
- 01Intro call & req scoping
- 02Challenge + rubric in days
- 03Candidate links & trace review