Five hundred learners, one paper, one evening. What goes wrong is rarely the software. It is the small decisions nobody made in advance: which paper each learner sees, what happens at 7.02 pm when a learner’s power cuts out, who may reset an attempt, and what you tell the learner who says the timer ate ten minutes.
This is a working plan for an online exam for 500 learners, from the fortnight before to the review after results: preparation, question banks and randomisation, timing and time zones, device and bandwidth checks, invigilation options and their trade-offs, learners who drop off mid-paper, and the post-exam review.
Two weeks out: settle the rules on paper
You do not need a large team. You need a one page exam note, circulated to faculty, your support desk and your learners, that answers every question a learner asks at 6.55 pm.
- Date, start time, last permitted start time and duration of one attempt.
- Who is eligible, and what happens to a learner who joined late.
- Number of attempts, and whether a re-attempt is graded or capped.
- What is allowed on the desk: calculator, rough sheets, notes, nothing.
- The identity proof a learner must show, and to whom.
- Who a learner calls when something breaks, on which number.
- Who may reset an attempt, and on what evidence.
- When results are published, and how a learner queries a question.
Point seven is the one institutes skip and then regret. If three people can reset attempts and none of them logs why, your result sheet is not defensible.
Question banks and randomisation
A single fixed paper sent to 500 people is on a group chat within four minutes. Randomisation is not about making the exam harder. It makes a shared screenshot useless to the next learner.
Build the paper from a bank rather than typing it out. Tag each question by topic and difficulty, then define the paper as a recipe: eight questions from Topic A at medium, four from Topic B at hard. Randomise in three places: which questions are drawn, their order, and the order of options inside each question. Option shuffling alone kills the crudest copying, where one learner passes along “1-C, 2-A, 3-D”.
Do not shuffle options where one of them is “both of the above” or part of a sequence, and keep the bank deep enough that two learners sitting together do not get a heavily overlapping draw.
The Public Examinations (Prevention of Unfair Means) Act, 2024 covers notified public examinations, such as those of UPSC, SSC, IBPS and the National Testing Agency, not your internal tests. Its list of unfair means is still a good checklist for your own rules: leakage of a question paper or answer key or part of it, solutions supplied by an unauthorised person, and tampering with a computer network or computer resource.
Timing, windows and time zones
Choose between a fixed start and a window. A fixed start is easier to invigilate and harder on learners with unreliable power. A three hour window is kinder, but the first learners out talk to the last learners in, so it needs a deeper bank.
India runs on one time zone, so for a domestic batch you only need to say that 7.00 pm means IST. With learners in the Gulf, Singapore or Europe, state IST and local time in the same sentence, and check your server’s time zone before the day. Set the timer per attempt, not per clock: a learner starting at 7.04 pm with a 60 minute paper must end at 8.04 pm. Set a hard close for the exam as a whole, and make sure the two do not contradict each other for a late starter.
Device and bandwidth checks
Do not assume. Send a five question practice test, open for three days, with the same settings as the real paper. It is a device check disguised as revision, and learners who cannot open it will not open the real one either.
There is no official speed floor to quote at learners. TRAI does not prescribe any minimum download or upload speed for broadband; it requires providers to declare a typical speed and to ensure at least 90 per cent of measured samples meet or exceed it. So the real test is not a number. It is whether the practice paper loads and submits on that learner’s own connection and device.
- One charged device per learner, with a backup power plan.
- Mobile data as a fallback, and the learner knows how to switch.
- Browser updated, other tabs closed, notifications off.
- Support number saved in the phone, not on a notice they will lose.
Invigilation options and their trade-offs
There is no free option. Each choice buys assurance and costs you money, staff time, learner goodwill or privacy exposure. The UGC defines a proctored examination as one supervised by an approved person or by technology enabled proctoring that confirms the identity of the test taker and the integrity of the test environment, with Indian students identified by Aadhaar or another government recognised ID.
| Option | What it catches | What it costs | Sensible for |
|---|---|---|---|
| Open book, no invigilation | Nothing, by design; the paper must test application | Rewriting the bank | Formative tests, upskilling |
| Randomised paper, tight timer | Casual copying between two learners | A deeper bank | Low stakes certification, large batches |
| Live invigilator on video | Impersonation, a second person in the room, phone use | Around one invigilator per 20 to 25 learners | Final exams, small cohorts |
| Automated proctoring with recording | Flags for review, not proof | Per learner fees, storage, a privacy duty | High stakes exams you can staff |
| Physical centre or your own lab | Most of it | Venue, learner travel, a seating plan | Terminal exams, licence linked programmes |
If you record faces, screens or rooms, you are processing personal data. The Digital Personal Data Protection Rules, 2025, notified on 14 November 2025 with an eighteen month phased compliance period, require a clear, separate consent notice explaining the purpose, and an answer to any request for access, correction or erasure within ninety days. Decide your retention period before the exam.
When a learner drops off mid-exam
This will happen, several times at 500 learners. Your system should save each answer as it is given, so a learner who reconnects resumes at the same question with the clock where it was. Test that on the practice paper by killing the connection halfway.
Then set one rule for everyone: reconnect within ten minutes and you continue on the remaining time; if you cannot, raise a support ticket during the exam window and a named person decides on a fresh attempt that night. A verbal request next morning is not evidence. A ticket at 7.21 pm is. Keep a six line incident log: learner, time, what happened, what was done, who approved it.
The post-exam review
Before publishing results, spend twenty minutes on question level data, looking for broken questions rather than hard ones.
- Almost everyone wrong: usually a typo, an ambiguous stem or a wrong key.
- Strong learners wrong more often than weak learners: the key or the wording is at fault.
- Suspiciously fast average answer time: the question may be circulating.
Drop or fix the broken ones, say in the result announcement which questions were dropped and why, then send the bad items back to the author with the data attached. An assessment engine that reports at question level turns this into a twenty minute job. Then check your server graphs for the last two minutes, when 500 submissions land at once. That peak, not the start, is what you size for, one of the practical differences in our note on self-hosted and cloud hosted systems.
Frequently asked questions
How many invigilators do I need for 500 learners?
It depends on the method. With live video invigilation, most institutes plan on one invigilator watching 20 to 25 learners, so a single sitting needs 20 to 25 people. That is usually the point at which institutes move to staggered sittings, or to a randomised paper with a tight timer instead.
Should I allow a re-attempt if a learner’s internet fails?
Yes, but on a rule rather than on sympathy. Publish the rule beforehand, require a support ticket raised during the exam window, and let one named person approve every re-attempt. If the re-attempt draws a fresh paper from the same bank, nobody gains an advantage.
Is automated proctoring worth it?
Only if you have staff to review what it flags. It produces flags, not verdicts, and an unreviewed flag is worse than none because it looks like evidence. It also brings recordings of learners into your care, with the duties that follow.
Running it in one place
All of this is easier when the bank, the paper, the timer, the attempt log and the result analysis sit in one system instead of four. Quipu LMS keeps question banks, randomised papers with timers and auto grading, roles and cohorts and progress analytics together, on your own servers.
