Saying numbers and terms clearly: every value with its unit, in a form the line can't flip

Say every value with its unit, in a form a poor line can't flip; say each term as its dictionary or project does.

Before you read: predict

ExamHall mock round on a voice call; your screen is not shared. The incident dashboard in front of you reads p95 890 ms · pool 20/20 · waiting 380 · error rate 4.2%.

The interviewer: “Read me what you're seeing.” Which reading do you give?

The 30-second version

On a call the interviewer has only your words, and writes down what they hear. A number without its unit is half a value. Some number words collide on a bad line, and a few technical terms are said differently from how they are spelled. Check both before the interview, not during it.

Say every value with its unit, in a form a poor line can't flip; say each term as its dictionary or project does.

How it works

Step 1 of 11

The listener has only your words. A value is a number times a unit, so say both. [S1]

Analogy: The bingo caller

Four hundred players sit with dabbers in a noisy hall, and they mark only what they hear; from the back tables nobody can see the board behind the caller. Across the chatter, fifteen and fifty blur, so the caller calls “one and five, fifteen”.

In the hall	In the interview
Players mark what they hear	The interviewer writes down what you say
The back tables can't see the board	The interviewer can't see your dashboard
Fifteen and fifty blur across the hall	-teen and -ty words collide on a poor line
“One and five, fifteen”	Backing up a number that can collide

Breaks down when: every bingo number runs from 1 to 90 and has no unit, while half an interview value is its unit; and a misheard call costs a prize, not follow-up time. Wrong inference: "back up every number." Only a number that can collide needs a backup.

In real terms: before you read a number aloud, check that someone who can't see it could write it down from your words alone.

Examples, run for real

EXH-4520 is a synthetic ExamHall mock on the row's hook: asked to read the incident dashboard aloud, the candidate's first take says the numbers alone. MR !4820, her second take, gives each value its unit. It reads numbers from words. On a line that flattens stress, it hears a number said as one bare -teen or -ty word as either partner. It writes "unit not said" when no unit word comes with a value. And it keeps only the readings on which a value's second form agrees.

EXH-4520 take 1 · “Read me what you're seeing.”
  p95 890 ms      “P ninety-five, eight ninety” · heard as 890 (unit not said) · at risk
  pool 20/20      “the pool, twenty of twenty” · heard as 20 (unit not said) of 20 (unit not said) · at risk
  waiting 380     “three eighty waiting” · heard as 380 (unit not said) · at risk
  error rate 4.2% “error rate four point two” · heard as 4.2 (unit not said) · at risk
MR !4820 take 2
  p95 890 ms      “P ninety-five is eight hundred and ninety milliseconds.” · heard as 890 ms · as the dashboard
  pool 20/20      “Twenty out of twenty connections are in use,” · heard as 20 connections of 20 connections · as the dashboard
  waiting 380     “and three hundred and eighty requests are waiting.” · heard as 380 requests · as the dashboard
  error rate 4.2% “The error rate is four point two per cent.” · heard as 4.2 % · as the dashboard

Take 1 gets every number right, and the interviewer still can't use one of them: “eight ninety” could be milliseconds or requests, and “four point two” a percentage or a count (steps 1 and 2). Take 2 says nothing new; it only says the units.

The model checks what could be written down, not whether the reading answers the question: a take that read every value perfectly and stopped there would still need to say what the numbers mean.

Try it

Answer first, then check. Nothing is saved.

Question 1 of 7

Media-streaming round. Your change moved the video player's manifest to the CDN edge. The player dashboard shows p95 startup time 2.1 s in the week before the change and 840 ms in the week after it.

The interviewer asks: “What was the startup time before and after?” Which answer do you give?

Question 2 of 7

Media-streaming round on a video call with live captions. You said that the transcoding workers take jobs from a queue, that the player's test suite runs on every merge, and that the edge cache serves the video segments. The captions on the interviewer's side read: “…take jobs from a Q, the player's test suit runs on every merge, and the edge cash serves…”.

Which caption shows a word you said differently from the way the dictionary gives it?

Question 3 of 7

Media-streaming round, voice only, on a line that flattens stress and drops word endings. The player dashboard reads startup p95 1.8 s · rebuffers 15 per 1,000 plays · bitrate 4,300 kbps · CDN hit ratio 93% · errors 0.4%.

You read it aloud: “Startup p ninety-five is one point eight seconds, rebuffers are fifteen per thousand plays, the bitrate is forty-three hundred, the CDN hit ratio is ninety-three per cent, and errors are nought dot four per cent.” Under the lesson's listener model, how many of the five values could the interviewer write down differently from the dashboard?

Question 4 of 7

Media-streaming round. You explain that the player's “mark episode watched” call is idempotent, and you say it “id-em-POH-tent”. In her next question the interviewer, from the company's US office, says it “eye-dem-POH-tent”.

What do you do with the word for the rest of the round?

Question 5 of 7

Write, word for word, how you would read this player dashboard to an interviewer who cannot see it: seek p99 1.4 s · requests 6,000 per second · bitrate 2,500 kbps · error rate 0.25%. Write four lines, each starting with its label: Seek:, Requests:, Bitrate: and Errors:, with every number in words.

Write the four labelled lines.

Your answer (optional, stays on this page)

Question 6 of 7

Media-streaming round, a week later, voice only, on a line that flattens stress and drops word endings. After your player fix the rebuffer rate fell to 15 per 1,000 plays; before the fix it was 50 per 1,000 plays, as you told the interviewer earlier. She asks: “And after the fix?” Each answer below adds a second form.

Under the lesson's listener model, which answer leaves her only one rate to write down?

Question 7 of 7

Media-streaming round, a week later. The player's watch history lives in a PostgreSQL database, and you are about to name it for the third time in a minute.

Which name do you use?

Go deeper

Pro tip

Before an interview, read three numbers from your own work aloud into a voice memo, then play it back and write down only what you hear. A value you can't write back with its unit is one you'd lose on a call.

Checked against the tools and docs as of 2026-10-15.

Predict first