Measure what students can do, not what AI can.
Department batteries of real work, in ordinary college labs.
No webcams. A level per skill, backed by evidence.
Real work, timed and proctored, no camera needed.
- 1Core screen
- 2Code reading
- 3Debugging
- 4Implementation
- 5Data handling
- 6Explain-back
def add(self, sku, qty):
if qty <= 0:
raise ValueError("qty must be positive")
self.items[sku] = self.items.get(sku, 0) + qty- PASS test_restock_adds
- PASS test_reorder_level_included
- FAIL test_sold_out_removed: AssertionError (test_inventory.py:24)
- 0
- cameras needed
- Levels 0–4
- per skill claim
- 2 lanes
- no-AI and AI-allowed
- Free
- first test cycle
One battery per department, built from real work
Every student in a department takes the same battery, so a college can run all 400 in one go. Role fit is an output of the results, not something students choose beforehand.
| Station | Measures | Time | Scored by |
|---|---|---|---|
| Core screen | Aptitude, reasoning, written communication | 15 min | Auto, fixed short form |
| Department knowledge probe | Scenarios with real datasheets, logs, invoices or drawings | 15–20 min | Auto |
| Work samples (2) | Fix, build or diagnose in a real environment | 25–40 min each | Hidden tests + anchored rubric |
| Explain-back | Structured questions about the student's own work | 5–8 min | Anchored 0–4 rubric, sampled double scoring |
| AI-allowed task (optional) | Decomposition, verification and repair of AI output | 30 min | Rubric, reported separately |
How well each method predicts job performance (correlation)
Recent meta-analytic estimates. Real tasks and job knowledge predict better than generic aptitude, so the battery is built around them.
Department batteries
MVP PlannedRead and trace code, debug, implement against tests, data handling, explain
Role fit: Software developer, web developer, QA engineer, data analyst
Circuit reading and calculation, embedded logic, trace diagnosis, safety
Role fit: Embedded engineer, test technician, electrical maintenance, VLSI trainee
Data cleaning, statistics, model evaluation, Python
Role fit: Data analyst, junior ML engineer, business analyst
Drawing reading, tolerances, material and process choice, failure diagnosis
Role fit: Design and CAD, production or quality, maintenance
Drawing reading, quantities and estimation, loads and units, site problems
Role fit: Site engineer, estimation engineer, design assistant
Spreadsheets, reconciliation, GST and invoice checks, writing
Role fit: Accounts executive, audit assistant, business analyst
Editing, fact-checking, writing to a brief, audience fit
Role fit: Content writer, editor, communications executive
Lab and data reasoning, measurement error, spreadsheet or Python analysis
Role fit: Lab analyst, data analyst, QA
No format is AI-proof, so we score the two lanes separately and never average them
| Secure lane | Open lane | |
|---|---|---|
| AI | Not allowed | Allowed and declared |
| Where | College lab, invigilated, lockdown | Lab or take-home |
| Purpose | Proves what the student can do alone; issues the skill passport | Shows how well they work with AI; optional AI-ready badge |
| Scored on | Hidden tests, rubric, explain-back | Decomposition, verification of AI output, repair, final quality, prompt log |
- 1Fresh multi-file tasks with hidden tests
Custom tasks cut ChatGPT-assisted pass rates from 73% (verbatim LeetCode) to 25%.
- 2Explain-back about the student's own work
Short, structured and scored on a rubric.
- 3Process evidence
Code snapshots, test runs and typing rhythm show how work was built. Used only to trigger review.
- 4Supervision in the lab
The invigilator is still the strongest control.
What we no longer rely on
- Images or diagrams to block AI: current models read them.
- AI-text or AI-code detectors: high false-positive rates, especially for non-native English writers.
- Plagiarism similarity alone: AI rewrites defeat it. It's used only to rank cases for a human.
Source for pass rates: interviewing.io experiment.
Integrity flags, not verdicts
A low integrity score triggers a human review, a short explain-back viva or a parallel retest, never an automatic penalty. Keystroke classifiers lose accuracy on new data, so we measure and publish our own false-positive rate during the pilot.
Lockdown
Full-screen kiosk. Blocks other apps and tabs, dev tools, copy and paste, screen capture and second monitors. Built-in calculator and scratch pad.
Behaviour signals
Focus loss, paste-like bursts, impossibly fast answers, code that appears whole, long idle followed by a full answer.
Live invigilator
A room view of every student's station, time left and flags. Warn, pause, lock, add time or end, with a typed reason.
Review queue
Integrity 81- Left the test window ×2−12
- Large block of text appeared at once−10
Answers autosave locally and sync after a network drop.
A level per skill claim, not one total mark
A level needs evidence from more than one station, every critical skill has its own gate, and a high total never makes up for a failed safety or correctness gate.
—
Knowledge probe + partial work sample
Work sample gates + explain-back ≥ 3 of 4
Level 2 + a changed-task variant
Project, internship or capstone defended in a viva
Evidence each reader already needs
| Report | For | Shows |
|---|---|---|
| Skill passport | Student, employers | Level per skill claim, lane, task version, date, link to evidence |
| Department heatmap | HoD | Skill × batch levels, weakest claims, suggested ELV training |
| College report | TPO, principal | Readiness by role, trend across cycles |
| NBA evidence pack | IQAC, NBA coordinator | Direct attainment mapped to programme outcomes, with rubric and task metadata |
| Integrity review queue | Admin, invigilator | Flags, reasons, review outcome, appeals |
| Validation dashboard | ELV | Task statistics, rater agreement, subgroup gaps, follow-up outcomes |
The NBA pack is evidence a college can use in its own attainment calculation; it doesn't grant accreditation. Percentiles exist only in a separate selection view.
Built for a whole college, and for what happens after the test
| HackerRank | LeetCode | ELV Assessment | |
|---|---|---|---|
| Departments covered | Coding only | Coding only | Every engineering and arts department, plus a common core |
| What is tested | Coding puzzles | Practice problems | Real work: fix, build, diagnose, explain |
| Questions | Large library; many circulate online | Public, solved online and by AI | Original task families, generated variants, retired when leaked |
| AI | Detection after the fact | None | Secure lane without AI and an open lane with AI, scored separately |
| Proctoring | Webcam-based | None | Lockdown + behaviour signals + live invigilator, no camera |
| Output | Pass or fail per candidate | Personal ranking | Levels per skill, role fit, department heatmaps, NBA evidence |
| After the test | Nothing | More practice | ELV skill-tree path and courses to close each gap |
We publish our own validation data
Pilot students are followed into internships and jobs at 6 and 12 months. We publish task statistics, rater agreement, subgroup gaps and how well levels predict outcomes. No Indian placement test has done this.
Free for your first test cycle.
Try every role in the demo: student, invigilator, college admin and ELV admin.