Live class attendance, measured honestly

Every teacher who has run a live class online has been asked a version of the same question, usually by somebody who was not in the room. A parent asks whether their daughter actually attends. A student contests a detention. A funder wants to know what their sponsorship bought. An internship coordinator wants to know who is still turning up in week six. The question is always the same shape — how much of this class did this person attend, and what was it about — and for most teachers the answer lives in memory, a WhatsApp thread, and a screenshot of the participant list taken on a good day.

The obvious fix is to have the software count it. That is where the trouble starts, because attendance in a live class is deceptively easy to compute badly, and a number computed badly is worse than no number at all: it looks defensible right up until somebody checks it.

Before deciding what to measure, it is worth being honest about who reads the result. An attendance figure is not a dashboard metric that lives and dies inside your own head. It gets forwarded. In practice it has to survive four different readers, each of whom will attack it from a different angle:

• The student who disagrees. They will say they were there. If your record cannot show when they joined, when they left, and whether they came back, you have no answer beyond your recollection.
• The parent who is paying. They want a pattern, not a day: is she attending, roughly, most of the time, and is it getting better or worse.
• The institution or funder. They want totals across a cohort and a range, in a file, with names and email addresses that match their own list.
• You, six weeks later, deciding whether to award a certificate. This is the reader everyone forgets, and the one who most needs to know what each class was actually about.

Every design decision below follows from that list. Nothing here is about a prettier chart; it is about a number that holds up when it is challenged.

The easiest thing to record about a participant is when they first appeared and when they were last seen. Subtract one from the other and you have a duration. It is one line of code and it is wrong more often than people expect.

Consider a student who joins at 10:00, realises they have the wrong link open, leaves at 10:05, gets distracted, and comes back at 10:55 for the last five minutes of a one-hour class. First seen 10:00, last seen 11:00. Measured as a span, that is sixty minutes and a perfect attendance record. Measured honestly, it is ten minutes.

So presence has to be accumulated as stretches, not as a span. Each stretch opens when somebody arrives, extends while they are still being seen, and closes when they actually leave. A return opens a new one. The time credited is the sum of the stretches; the gap between them is nobody's class time.

This has a pleasant side effect: because the stretches are kept, the report can say how many times somebody rejoined. One rejoin is a dropped connection. Five is a student who is not really in the room, and it is worth knowing which is which before you write to them.

The second mistake is quieter and does more damage. A percentage needs a denominator, and the obvious one is the class length you scheduled. But scheduled classes routinely do not run to their scheduled length. You start eight minutes late because the previous session overran. You finish early because the material ran out. You cancel the last twenty minutes because half the cohort is sitting an exam.

Measure against the sixty minutes you booked and a class that actually ran thirty-two reports a full, attentive room at 53% — and every one of those students has a permanent record saying they missed nearly half of something they did not miss.

The defensible denominator is the time the teacher was in the room, because that is when the class existed. If you were not there, it was not class. It also handles the awkward case honestly: if your own connection drops for twenty minutes and comes back, the class is the union of the stretches you were present for — not the span from your first login to your last goodbye. Nobody should be marked absent from twenty minutes during which nothing was being taught.

The third gap is not arithmetic at all. A great many people join a live class by opening the link somebody sent them. They never signed into the platform, they may not have an account, and any attendance mechanism that runs inside the student's app never hears from them. For an internship cohort or an open workshop, that can be most of the room.

The fix is to make the host the recorder. The teacher's browser is the one participant that can see the entire room, including anonymous guests, and the name and email each person supplied on the way in ride along with them. So the register is written from the host's view of the room, once a minute, for everyone in it — signed in or not. It means a workshop attendee who joined from a forwarded link still lands in the record with a reachable email address next to their name.

The tiles are the answer to “how is this cohort doing” in one line: how many classes ran, how many distinct people attended anything, the average attendance across every line, how long you actually taught for, and — the one nobody expects — how many of those classes have a topic recorded against them.

Two charts, because they answer different questions. The day line is for the pattern — is this cohort drifting, did something change after the mid-term. The per-class bars are for the outlier: one bar noticeably shorter than its neighbours is a class worth thinking about, and because the topic is recorded you can see immediately whether the room emptied during the database session or during the one you ran at 8am on a Saturday.

This is the view that settles an argument. Not a percentage on its own — a percentage with the times that produced it, so a student who says “I was there” can be shown that they were, for half of it, from half past. Where somebody left and came back, the row says so, with the number of rejoins next to the minutes.

Here is a thing worth pausing on. Nearly every tool that records attendance in a live class records the room, the date, the people and the duration. Almost none record what the class was about.

That sounds like a nicety until you look at what an attendance record is used for. “Ravi attended 47 of 52 minutes on 3 September” is not, on its own, evidence of anything. Forty-seven minutes of what? The answer sits in the teacher's head, and by the time somebody asks — a certificate decision, a parent meeting, an audit of a sponsored programme — the head has moved on to next term.

The reason it usually is not stored is that people assume it already is. The scheduled session has a title, so surely that is the topic. But the title is the name of the recurring room — “Full-Stack Internship · Batch 45” — and it is identical for all forty-five sessions. Planning boards have topic cards, but a card is an intention: what somebody meant to cover, on a board, not what actually happened in Tuesday's class after the first twenty minutes went on questions about last week's assignment.

So it gets its own row in its own table, keyed by room and date — one topic per class occurrence, edited in place, rewritable when the wording is wrong. Deliberately not attached to the scheduled session, because a recurring room has one session record and forty-five different lessons.

Where it earns its place is on every line downstream. The topic rides on each attendance row, into the CSV, into the PDF. It also becomes searchable: typing “pointers” finds the class about pointers and everybody who was in it, which is a genuinely useful thing to be able to ask a register.

Search on an attendance register has a failure mode worth naming: when a search for a misspelt name returns nothing, it does not read as “you spelled it wrong”. It reads as “she never came to class”. Teachers type quickly, and names transliterated from another script have several plausible spellings — Sheikh, Shaikh, Shiekh — none of which is a typo exactly.

So the search is deliberately forgiving. It matches on exact text first, then prefixes, then per-word with a tolerance for a character or two out of place, and finally on the letters in order — so “dstr” finds “Data Structures”. It searches names, email addresses, room codes and the topic taught, and it narrows the charts along with the table, so what you are looking at and what the numbers say cannot disagree.

The filters are in the URL, which sounds like a technicality and is not: a teacher who has narrowed to one room and one week can send that link to a head of department, and they will open exactly the view that was described to them.

Two exports, and one rule about both: they obey the same filters as the screen. An export that quietly ignores the search box is how somebody emails a spreadsheet of the wrong cohort to a funder, and does not find out for a month.

The CSV is the one that goes into somebody else's system — one row per attendance line, with the topic, the join and leave times, the minutes and the percentage. The PDF is the one that goes to a human: the charts, a table of classes with what was taught in each, a table of students, and every line underneath.

One detail in that PDF is worth copying whatever tool you use: the footer says how the numbers were arrived at. “Attendance is measured against the time the teacher was in the room, and counts only the stretches a person was actually present.” A report that states its own method can be argued with. A report that just shows 17% invites the reader to assume the worst about either the student or the software.

There is a category of live-class analytics that measures whether a student's tab was in focus, whether their camera was on, whether their mouse moved. It is technically straightforward and we are not going to do it.

Partly because it is wrong about people — a student listening carefully on a phone while walking is not disengaged, and a student with a camera pointed at a wall may be the most attentive person in the room. Partly because it is wrong about bandwidth: in most of the cohorts we see, cameras off is the difference between attending and not attending at all. And partly because a number that cannot be shown to the person it describes without embarrassment is not a number worth collecting.

Time present, honestly measured, is a defensible thing to record. Attention, inferred from a browser, is not.

1. Use one room per cohort and let it repeat. The register keys on room plus date, so a recurring room gives you one clean row per day rather than a pile of one-off sessions to reconcile.
2. Open the room yourself, and stay in it. Your presence is the denominator; a class you dial into ten minutes late is a class that is ten minutes shorter for everybody, which is correct but worth knowing.
3. Write the topic the same day. It takes five seconds while you are looking at who turned up, and it is the single field that makes the whole record answer questions later.
4. Check the outliers weekly, not the average. The average tells you nothing you did not already feel. One student at 30% for three weeks is the thing worth a message.
5. Export before the term closes, not after. A PDF with the charts is the artefact that survives; a dashboard is only useful while somebody has a login.

An attendance register is not a metric. It is evidence, and it gets used at exactly the moments when relationships are strained — a detention, a refused certificate, a conversation with a parent who has been paying for something they suspect is not happening. Evidence that overstates attendance protects nobody, because the first person to check it stops trusting the rest of your records too. Evidence that understates it punishes students for your scheduling.

The whole job is to make the number mean what it says: this person was in this class, for this long, out of this much class, and here is what the class was about. Everything else on the screen is arrangement.