How it works

Everything inside an Eigen-Rho interview.

The sealed case, the live twist, the cockpit, the verdicts, and the report, in full. New here? Start on the overview →

How the screen holds up

Proof they can do it. Not that they can prompt an AI.

One platform, built so the answer key never reaches the browser, the question can change live, and you watch the work happen. Here is what is inside.

CPG · The Vanishing SKU
-- four clues, one root cause
SELECT region, SUM(revenue) FROM sales
GROUP BY region ORDER BY 2 ASC;
▶ RUNexplore freelySUBMITgraded on the server

A narrative case unfolds over four clues, each unlocking only after the last one checks out. Candidates explore with RUN; SUBMIT is graded server-side.

Public hook"Sales of FC-2847 fell 23% in six weeks."
The brief
The four clues & hints
The schema seeds
The golden answer

Public surfaces show only the hook. The brief, clues, schema seeds, and answer live behind the server boundary. No expand panel or inspect-element reaches them.

Q. Which region drove the drop?
⚡ Push twist
Q. Now isolate how much is the new carrier vs. seasonality.

Mid-interview the question shifts in a way no question bank or second-screen model anticipated. A rationale gate then asks, in one line, why this query answers it.

human cadencepaste burst
multi_monitorgaze_offtranscript_detected

Typing forensics read the rhythm of an answer, not its content, so off-screen transcription shows up. The camera is server-enforced: no camera, no queries.

14:02:11SELECT region, SUM(revenue)...PASS
14:03:48JOIN shipments ON ...RUN
14:05:02SELECT carrier, AVG(days)...FAIL

Your hiring manager runs the room: every query, result, and flag in real time, push questions one at a time, drop the twist when the moment is right.

✓ CLEAN87/100
Typing cadencehuman
Paste / transcriptionnone
Live-twist responseadapted

One weighted verdict reads like an instrument, always with the signals behind it. A no-show reads NOT ASSESSED, never a false 100 and never a false Clean.

One report, named per candidate and session

A verdict your panel can read in thirty seconds, and defend in calibration.

Every interview closes with one shareable evaluation: the integrity verdict first, then a skill score, the signals behind it with severity, per-clue graded results, and the full timestamped transcript. Export it to PDF and forward it. It reads like a financial verdict, not a gut call.

Candidate Evaluation
FINANCIAL SERVICES · THE FRAUD PATTERN · HARD
✓ CLEAN
87 / 100
Skill score
Low
Integrity risk
Typing cadencehuman rhythm, natural corrections
Paste & off-screen transcriptionnone detected
Live-twist responseadapted in the next few queries
Multi-screen & gazesingle display, eyes on the problem
Queries logged14 runs · full history attached

Illustrative example, not a customer metric.

Four possible verdicts. Each says exactly what it means.

The verdict reads like an instrument: monospace, all caps, never a color without a word behind it. Every interview resolves to one of four states, and a no-show is one of them by design.

CLEAN

Work holds up. No integrity concerns worth flagging.

REVIEW

Mostly solid, with one or two signals worth a second look.

SUSPECT

Signals point to outside help or a pasted answer. Read the evidence.

NOT ASSESSED

No usable session to judge. A no-show is recorded as exactly that, never a false 100 and never a false Clean.

Whichever state it returns, the verdict carries the specific signals that produced it: multi-monitor, paste, gaze off-screen, transcription cadence, twist response. You decide with evidence you can point to.

Nine industries, 155+ live questions

Calibrated for the data roles you actually hire.

Every case is a false narrative: a surface metric that hides a different reality, on its own schema and realistically seeded data, across nine industries from CPG to financial services to paid media. The candidate has to verify the headline, decompose it, find the root cause, and quantify the impact, the way the work actually happens. Easy, Medium, and Hard tiers match the seniority you are hiring; each case below shows only its hook.

The same interview, two screens

What they see. What you see.

The candidate solves on live data. You watch the work and the integrity signals they never see. Here is the same moment from both seats, in real time.

A scripted illustration. Case internals never reach the candidate’s screen, and never this page.

A real interview. In four steps.

You run it live and watch the work happen. No take-home to outsource, no black-box score you cannot explain.

01

Pick a problem

Choose an industry and a case calibrated to the role: Easy, Medium, or Hard, 20 to 40 minutes. You see the full brief, the four-clue arc, and the real client it is calibrated for. The candidate sees only the hook.

02

Share a join link

Send the link on your usual call. The candidate enters a session code, agrees to proctoring, and opens a three-pane workspace against a live, seeded database in the browser. Nothing to install. The live camera is relayed to the interviewer only and is never recorded or stored by us.

03

Push questions, live

Release the clues one at a time; a clue only unlocks after the server checks the previous step. Watch the query feed update: timestamp, SQL, row count, and PASS, FAIL, or RUN. When the moment is right, push a twist and change the problem under them.

04

Get a scored report

The session closes into one document, verdict first: CLEAN, REVIEW, or SUSPECT, each tied to the signals behind it. Then the skill score, per-clue results, integrity events with severity, and the full timestamped transcript. Export to PDF and forward it to your panel.

Where we are different

Most tools detect the cheat after the fact. We make it pointless.

The rest of the market bolts surveillance and plagiarism scores onto coding puzzles whose answers can leak, be memorized, or be generated. Eigen-Rho removes the answer key from the room and changes the question live, so there is nothing to scrape and nothing a copilot can survive.

×

The answer is gradeable in the browser, so question banks leak and must be defended with similarity checks, leak-monitoring crawlers, and rotating variants.

Sealed cases. The brief, clues, hints, schema seeds, and answer never reach the browser; SUBMIT is graded on our server. Nothing to leak, scrape, or pre-feed.

×

Integrity is a probabilistic flag on a static problem the candidate could have rehearsed: a suspicion score, a plagiarism percentage, an event log.

A live twist, pushed by your interviewer mid-interview, changes the problem. A pre-baked answer falls apart in the next few queries; real reasoning adapts in front of you.

×

Classifiers read the submitted code or keystrokes, and at least one vendor publicly admits this misses overlay tools that transcribe from an off-screen AI.

Typing forensics read the cadence of an answer as it arrives, not its content, so off-screen transcription shows up. A rationale gate then asks, in one line, why this query answers the question.

×

You get a skill score in one place and integrity flags in another, and reconcile them yourself. A no-show can quietly read as a zero or a pass.

One weighted verdict reads like an instrument: CLEAN, REVIEW, or SUSPECT, always with the signals behind it. A no-show reads NOT ASSESSED, never a false 100 and never a false Clean.

×

The live interviewer is a contractor on someone else’s roster, or an AI, working through generic software-engineering questions.

Your hiring manager runs the room from a real cockpit, on forensic business investigations on seeded data, at software economics.

Honest caveat: we are SQL and data today, with more languages on the way. We claim depth and structural prevention, not the broadest catalog or a magic detection rate.

For the candidate

For the candidate doing the work.

Most of this page is written for the person hiring. This part is for the person being interviewed, because a screen the strong candidates resent is not one worth running.

This is a hard interview and a fair one. You get a real business problem on real data, a SQL editor you control, and room to think. RUN as many queries as you need to understand the problem; running is never graded. SUBMIT when you are ready to be checked against the answer. The question stays in view, separate from your editor, and every state (idle, running, pass, fail, an empty result, an error) is legible at a glance, because ambiguity under pressure is not fair.

On privacy, plainly: your live camera video is shown to your interviewer to confirm it is you. We do not record it and we do not keep it. Your microphone audio stays on your device. We can tell whether a second display is connected; we cannot see what is on it. None of these checks decide the outcome on their own, and a person reviews the result. The reason for all of it is one thing: so that strong work is recognized as your own.

Two ideas from the math of data, put to work on one decision

Why we're called Eigen-Rho.

An eigenvalue is what a system holds onto under pressure. Rho is whether a relationship is real or a coincidence. A hire turns on both, so the name carries both.

Eigen

The ability that survives the pressure

In linear algebra, most directions get knocked off course when a transformation acts on them. An eigenvector is the one direction the system cannot bend, and its eigenvalue is the exact factor by which it stretches: the property that stays characteristic no matter what you put it through. We screen for the same invariant. Strip away the résumé, the rehearsed answer, and the model on a second screen, then push a live twist mid-investigation, and what is left is real reasoning or it is not. That residue is what we measure.

ρ Rho

The fit the evidence actually supports

In statistics, ρ measures whether two things genuinely move together or just happened to line up once. A high ρ is a relationship you can stand behind; a low one is noise wearing a pattern's clothes. We report fit the same way: not a gut read that a candidate seems strong, but per-clue graded results, integrity signals, and a verdict tied to the specific evidence that produced it. Correlation you can defend in a calibration meeting, not a coincidence you explain after the hire.

Eigen for the ability that holds up under pressure, rho for the fit the evidence supports. One name, one verdict you can point to.

Tailored for your hiring

Tell us what you hire for.

Pick what matches your team. When you log in, you get exactly this: your roles, your case types, your rubric. No two rollouts are the same.

Role
Stack
Volume
What matters most
Your tailored setup

  • Cases
  • Stack
  • Verdict rubric
Request this setup →
A 30-day pilot on your own hires, not a sales demo

See it on your own candidates.

Start a 30-day pilot. We add your company domain to the allowlist, your team runs real proctored interviews with unlimited logins, and every verdict you generate (CLEAN, REVIEW, or SUSPECT, with the signals behind it) is yours to keep and forward to your panel. No commitment.

Sign in

We reply within one business day. Your pilot, your data, your verdicts to keep.