How we measure ability

A certificate is only worth the method behind it. This page is that method, in full — how questions are written and checked, how difficulty is chosen, and how we tell a learner who is working from a learner who has found a shortcut.

Both you and every question carry a rating

We use the Elo system, the same idea used to rank chess players. You hold a rating in each course you study; every question holds one too. When you answer, both ratings move — you gain more for a question rated above you than one rated below, and a question that most people get wrong drifts upward on its own.

A new learner starts at 1,800 and a new question at 1,250. Neither number means anything on its own — only the gap between the two decides how likely you are to get a question right, which is what makes the scale self-calibrating rather than something we set by hand.

Ratings are held per course, not per person. Being strong at one subject tells us nothing about another, and treating it as if it did would serve you questions you have no business being asked.

Your questions are chosen from your rating, and nothing else

Each weekday we select 20 questions whose ratings put your expected accuracy at about 96%. In Elo terms that puts the ideal question roughly 552 points below your current rating, sampled from a band 150 points either side of it so the day is varied rather than 20 questions of identical difficulty.

As your rating rises, the questions get harder. That is the whole mechanism: improvement is measured by what you can now be asked, not by a score you can reach by repetition.

Difficulty changes what is asked, never what is paid. Every correct answer credits the same rate whatever the question was rated. The code that selects your questions cannot read balances, sponsor funds, or how much you have earned — it is given your rating and the question ratings and nothing else, and there is a test that fails if anyone ever wires more into it.

New ratings move fast, then settle

A rating nobody has tested yet is a guess, so it is allowed to move quickly — a K-factor of 48 — and then slows to 16 once it has 30 answers behind it. This applies to you and to questions alike: a new question finds its true difficulty within its first few dozen answers instead of being stuck at whatever the author assumed.

While a question is still calibrating it is served but excluded from scoring decisions that depend on a trustworthy rating, so an uncalibrated question cannot distort what you are measured against.

How questions are written, and double-checked

Questions are drafted against a written syllabus for the course, then put through checks before anything reaches a learner. A wrong answer key on this platform pays the wrong person and certifies the wrong thing, so no single check is trusted:

1. Blind re-answer. A second, independent pass is shown the question and its options but never the intended answer, and has to answer it. Disagreement rejects the draft outright. This is what catches a confidently wrong key, which is the failure that matters most.
2. Structural checks. Exactly one correct option. No duplicated option text. No option that gives the answer away. A body that is not already in the bank.
3. Duplicate detection, in two layers. An exact fingerprint of the question text with spacing, case and punctuation normalised away catches a repeat and its trivial re-wordings; a separate similarity check catches a genuine rewrite of a question already banked. You should never be paid twice for the same piece of knowledge.
4. Human review. Drafts land in a review queue, never directly in the live bank. A person approves, edits or rejects before anything is served.

Some subjects also use parameterised generators — where the computation genuinely is the skill being tested, a template that produces fresh numbers each time is more honest than a fixed question everyone eventually memorises.

Answer sharing does not work here

The options under every question are shuffled independently for each learner. There is no stable “option B” to pass on. Two people sitting side by side with the same question are looking at the same five answers in two different orders, so a shared answer key carries no information at all.

Sets are drawn per learner from your own rating band, so two people rarely have the same day of questions in the first place, and a question you have already seen is not served to you again within a cycle.

Timing is measured on our side

The clock on a question starts when the server sends it, not when your browser says it did. Anything a browser reports about how long you took is treated as information, never as the basis for a payment — a number the client controls is a number worth whatever the client wants it to be worth once money depends on it.

Submit stays disabled for the first 3 seconds of every question, on every cycle. Answering faster than a question can be read is not learning, and clicking through 20 of them would otherwise be the quickest way to a day’s earnings.

Taking the jump: why going faster pays less

20 questions a weekday is the default, and it is a gate rather than a target — you cannot pull tomorrow’s questions forward. That is the whole design. Spacing is what moves knowledge into long-term memory; a cycle finished in a weekend is cramming, and cramming does not survive contact with the job you are studying for.

But some people genuinely want to move faster, and refusing outright would only teach them that the platform is arbitrary. So there is intelligent jump: you may take the gate off and work ahead. What you cannot do is take the gate off and be paid as though you had not.

On a jumped question, what it pays depends on how long you take:

• Answer within 30 seconds and it pays in full.
• Past that it halves immediately, then runs down in a straight line.
• At 60 seconds it pays nothing at all — correct or not.

The cliff at 30 seconds is deliberate rather than a smooth curve from the first second. A decay that started immediately would punish reading the question, which is exactly the behaviour we want. A flat grace period followed by a drop separates thinking about it from going to look it up somewhere.

So the timer is not a speed test. It is what makes jumping cost something. Working through the cycle at its own pace has no timer attached to your earnings at all — the countdown on an ordinary day is advisory, and an answer after it still pays. Pull the days forward and the same clock starts deciding what you are paid.

It is one decision for the whole cycle, not a choice you re-make each morning, and it is recorded on your cycle the same way your rates are — so the terms you agreed to cannot be rewritten underneath you, in either direction. Ability is measured identically either way: jumping changes what an answer pays, never which questions you are asked or how your rating moves.

Jumping moves the start, never the deadline. Every day in your cycle belongs to its own date and ends at your midnight, whether you opened it early or not. Open a day ahead of time, answer half of it and come back after its date has passed, and the rest is gone — exactly as it would be for a day you never opened at all. There is no backlog to work through and nothing rolls over. The one thing that is always still there is the day you are actually standing in.

We look at results that do not look like learning

Because your questions are chosen to sit near 96% accuracy, we know what an honest cycle looks like statistically before it happens. A cycle far above that band over 400 questions is not a compliment, it is an improbability, and it is measured as one — how many standard deviations above the expected result it sits.

We review the same data from other angles too: accuracy against what your rating predicted, the spread of time spent per question, and hint usage. None of this is shown to the learner it describes, and none of it is an automatic penalty — it marks a cycle for a person to look at. What happens then is set out in the terms.

We are open about the limit of this. An objective, multiple-choice test taken on your own device cannot be made impossible to cheat, and we do not claim otherwise. Our main defence is not detection — it is that the difference between an honest cycle and a perfect one is deliberately small, so getting help is not worth the trouble it takes.

What there is to read, and what it is for

The daily task is retrieval practice: a question, your answer, then the explanation. Being made to pull something out of memory is what fixes it there — better than reading the same thing again, which feels like learning and mostly is not.

But retrieval sharpens what you half-know and cannot hand you what you have never met. So a course is written in named parts, and where those parts have material, buying the course is what gets you it — to read before you sit the test, during it, or long after. A question you get wrong tells you which part it came from and links you to it. It links you after you answer, never beside a live question: reading it up while answering turns the practice into lookup, and the practice is the part that works.

Some courses have material for only part of what they cover, and some have none yet. What a course covers, and the name of every part of it, is listed on its page whether or not you have bought it — so you can see what you would be examined on before paying. Whether you read any of it afterwards is your business: we do not track it, and it changes nothing about what you earn.

What a certificate certifies

A certificate is issued for a completed cycle where you answered at least 90% of the questions available to you and your rating in that course was higher at the end than at the start. Participation plus measurable improvement — not a percentile, and not a place on a leaderboard, and nothing at all about whether you read the material. Reading is not something we can measure, and a certificate that implied otherwise would be claiming something we cannot check.

Every cycle gets a verification code, and anyone holding a certificate can check it at /verify without an account. A certificate nobody can check is a picture.

What “Intelligent Jump” means

An intelligent jump is our term for commitment-based motivational learning: you pay a fee up front for a course, that fee buys your access to its question bank and daily task, and answering correctly returns a share of it as you go. The money is not a prize for being clever. It is there because free courses are easy to abandon, and something at stake is what gets people back to the work on the days they would rather not.

The design follows from that. The daily rhythm exists so learning is spaced rather than crammed. The adaptive band exists so the work stays gettable instead of demoralising. The narrow gap between an honest cycle and a perfect one exists so the motivation stays motivational and never becomes a reason to cheat.