OutFigure Get in touch

Know what actually worked.

An AI tutor can get someone to the right answer today. OutFigure checks whether they can still get there next week, on their own, and how sure you can be that the tutor made the difference.

Ask about a pilot See how it’s measured
DemoTutor · sandbox project Synthetic demo data

Receipt for dt_int_88

Contrast case · Area of rectangles · delivered 2 September 2026, 08:32 UTC

Status evaluated every check is in, or its time ran out

How far MD4 later, on something new

How sure CS1 it came after, that’s all

What followed
  1. Before the lesson wrong
  2. Straight after wrong
  3. A week on, new item right
Evidence considered
  • 2 days before Before the lesson Starting point
  • 13 minutes after Right after Counted
  • 7 days after Later, same kind of question Counted
  • 7 days after Later, new question Counted
What this receipt can’t prove
  • Nobody knows how reliable the test was
  • It came after the lesson, and that’s all
  • One person’s change can’t be pinned on the lesson
Experiment · teaching strategy

A week later, each style against worked examples

  • Worked example52% the comparison
  • Socratic question59% +6.5 points (+2.1 to +11.1)
  • Contrast case65% +12.9 points (+8.4 to +17.3)

CS3 · a fair, random test · split held, p = 0.50

A drawing of the receipt page in the console: one made-up learner’s receipt and the experiment they were part of, worked out by the API’s engine on DemoTutor’s 5,000 made-up learners. See the whole example

The test that matters comes a week later.

Any tutor can make the next answer better. What counts is whether the learner can still do it a week on, with nobody helping.

We built a pretend tutor, DemoTutor, to show what that looks like. Its 5,000 invented learners each got one of three teaching styles at random. Worked examples won straight after the lesson, 79% against 70% and 71%. A week later they came last, 52% against 59% and 65%.

Invented data. We built those effects into the learners on purpose, so every estimate can be checked against the truth. They say nothing about real teaching.

DemoTutor · share answered correctlySynthetic demo data
STRAIGHT AFTER A WEEK LATER 79% 52% 70% 59% 71% 65%
  • Worked example
  • Socratic question
  • Contrast case

All six parts work today, on made-up data.

The last two only give advice: they tell your team what the evidence supports and never act on a learner. None of the six has run on a real company’s data yet.

  • Check your old logs

    works, on your computer

    Reads an export of your past logs on your own computer and tells you which questions they can answer. Nothing is uploaded.

  • Record what happened

    works

    Keeps a record of what your tutor did and what the learner did next. Nothing is overwritten, and nothing is counted twice.

  • Write a receipt

    works

    For each lesson it waits for the follow-up checks, then writes down what happened next and what that does and doesn’t prove. Later evidence makes a new version. The old one stays.

  • Test two versions fairly

    works

    Gives learners different versions of your tutor at random, fixes the plan before anyone is taught, and compares how they did later, with the margin of error shown.

  • Hold back a release

    works, as advice

    Reads a fair test’s result and says, for each new version of your tutor, release it, hold it back, or have a person decide. Your release process acts on it. OutFigure releases nothing.

  • Suggest the next step

    works, off until you turn it on

    For one learner and one skill: check them first, use the version a fair test backs, or run a test because nothing is proven yet. It goes to your system, never to a learner.

Every receipt answers two questions.

How far does the evidence reach: right after the lesson, a week later, or on something new? And how sure can anyone be that the lesson caused it? We score the two apart, because a result can last a long time and still be a coincidence.

How far it reaches

  1. MD0nothing we can use
  2. MD1only indirect signs
  3. MD2checked right after
  4. MD3checked a day or more later
  5. MD4later, on something newthe receipt

How sure of the cause

  1. CS0can’t say
  2. CS1it came after, that’s allone learner, at most
  3. CS2a pattern, other causes allowed for
  4. CS3a fair, random testthe experiment
  5. CS4a fair test, repeated

Each receipt also says what it can’t prove. A check that never came back is reported as missing, and missing is never counted as a fail. Late evidence gets a new version, and the old one stays.

Made for teams that teach with their own content.

Bring your own questions, IDs and skills. OutFigure doesn’t assume any curriculum.

  • AI tutoring

    Which of your tutor’s moves are followed by learning that sticks for a week.

  • Adaptive learning

    Whether a new lesson order changed how learners do later, on questions they hadn’t seen.

  • Language and test prep

    What learners still get right a week on, kept apart from the answers they had help with.

  • Learning research teams

    Fair, random comparisons, with the plan fixed before the first learner starts.

Start with the logs you already have.

Before you build anything, a small tool reads an export of your past logs on your own computer and tells you what they can show.

You get the receipts OutFigure would have written, and a straight answer on what your history can and can’t tell you. If it’s worth going further, your app starts sending us records. Learners see no change: receipts are written alongside your product, and your tutor keeps working exactly as it does now.

Ask about a pilot How it works

outcome-audit · synthetic export · 240 invented learners
  1. Clean recordssupported
  2. Checks that came backsupported
  3. Checks a week latersupported
  4. Checks on new questionssupported
  5. Comparing stylesassociation only
  6. Proof of causenot supported

No experiment was recorded in this export, so a causal comparison is not supported.

What we keep about a learner.

A pseudonym inside one project, and the answers recorded under it. That’s it.

  • Pseudonyms only

    No names, emails or birthdays. The same ID in two projects is two different people.

  • Nothing about how they behave

    No keystroke timing, hesitation or reading speed. We go by what the learner answered without help.

  • No AI model inside

    Records, receipts and experiments are worked out without a language model, and no learner data is sent to one.

A second reader for human marking.

For exams marked by people: every first mark set beside its second reading, and the gaps handed to a senior marker before results go out. No machine changes a mark.

How the second reader works A sample report

Pick one thing your tutor does.

Tell us what it does and which checks already come after it. A look at your past logs will show what your data can already say about it.

Ask about a pilot How it works

OutFigure grew out of Cognivia, a research project on what learning interactions can show. The method comes from there, and that research isn’t proof the method works. Every number on this site comes from invented data, run through the same engine as the API.