How a Wizzad defense works, and what it does not claim

What a sitting is

The questions are built from the piece itself, and each points to a passage the piece contains. The student answers one question at a time — 90 seconds each, 15 minutes in all, or 1.5 times that or twice that where the student chooses more time before the first question, as a student with extended time in tests would; every sitting states the time it allowed, and none says why — and the piece is not shown, apart from the short excerpts a question is about. A student sitting alone may sit a piece once a day. Every sitting that counts is kept in order, and a later one never replaces an earlier one.

When the sitting ends its results are counted and signed. That signed record is never edited afterwards — not by the student, and not by Wizzad.

Practice is kept for the student alone. It is never shown on a link and never counted; a sitting that counts after practice says how many practice sittings on the piece came before it.

Every sitting on a piece, practice included, is given questions the student has not been asked on it before. The questions already put are listed to the model that writes the new ones, which is told not to ask them again in any words; one that comes back in the same words about the same passage is refused. Each sitting states, signed, how many of its questions had been asked before.

Where an AI is used

Most of the questions are written by an AI model from the text of the piece. A question that does not point to a passage the piece contains is thrown away before anyone sees it. Questions about how a piece written in Wizzad changed are not written by a model: they are made by a fixed template from the piece’s saved history, and each quotes the passage it is about.

Explanations — reasoning in the student’s own words — are graded by an AI grader against a marking guide written from the piece’s own reasoning — the points a full answer contains, each marked present or absent, so the grade is the share present — and graded again by a second AI model from another company, and by a third where Wizzad holds a key for one, each blind to the others’ grades. Where their grades differ, each is shown the others’ scores and reasons once and asked to reconsider its own; the grade of record is then the middle one of three, or the first grader’s of two. No grader is shown the piece: the marking guide is all they grade against. The record names every grader and says how often they agreed, before and after conferring.

When a student chooses to answer aloud, each question is read to them in an AI voice. That audio carries a machine-readable mark that it is synthetic (the IPTC digital source type “trainedAlgorithmicMedia”). What the student says is written down by speech-to-text software and graded exactly as the same words typed would be.

A student may choose to be examined by an AI voice and face. It reads the questions built from the piece and listens. It does not grade, help, or add a question of its own.

The student is told before the first question. The words on the screen are: “These questions were made from your piece, most of them by an AI. Facts and figures are checked against the piece itself; a what-if question is checked against an answer the AI worked out from your piece; what you explain in your own words is graded by an AI against a marking guide, and again by a second AI when one is available.”

Where no AI is used

Facts, figures, choices and altered details are compared automatically with a set answer taken from the piece, and no AI judges the student’s answer. A what-if question — what would follow if one thing in the piece were different — is marked the same way, against an answer the AI worked out from the piece’s own reasoning when it wrote the question; nothing checks that answer against the piece.

The counts on a record are arithmetic over the answers. No model writes or adjusts them.

The signature on a record is ordinary public-key cryptography, and anyone can check it without Wizzad.

What a person does

No person grades a sitting as it is sat.

A sitting can be sat with a host in the room. A host is admitted by their school — a school a person at Wizzad has checked and approved — or by Wizzad itself, and either way Wizzad has confirmed the host’s school email address. A host their school admitted signs under the school’s name, and the record of each sitting they host names the school and says how it came to be one on Wizzad. The host checks the student’s photo ID in person before the sitting begins, watches the whole of it, and then gives their own signed word on what they saw, with anything they noticed. A host with a room of students may give one word for every sitting in it, saying in it what they noticed in each; each student’s record carries that word with its own sitting and nothing of any other student’s. Wizzad does not see or keep the ID.

A piece’s standing compares what its supervised sittings found with its sittings sat alone. Where that comparison would show a status asking for more evidence, a person at Wizzad looks at it first, within 3 working days of the sitting or the host’s word it waits on; until they have, neither the student nor any reader is shown that status, and the piece reads “Awaiting review”, which is no finding. The student is told the day it is due, and told when it is done. The person may then release that status, or let the supervised result stand; where a standing rests on their word, the record says so.

A student can ask a person to review any sitting that counts. That is described below.

Two kinds of sitting, and what each can show

A sitting with no host is sat wherever the student chooses. Before it the student is asked to answer on their own, in these words: “This sitting is for the record, so answer on your own: no person, no notes and no AI helping while you sit. Nothing in this sitting checks that — a reader takes it on your word.” Wizzad does not watch the room, and such a sitting cannot rule out help from a person, from notes or from an AI. A record says so.

A sitting with a host in the room carries the host’s signed word. It is the only kind that moves a piece’s standing. On a record made since Standard v1.11 (5 October 2026), a piece reads “Supervised sitting on record” only once the supervised sittings that count earned at least half their question credit between them, always with who hosted the latest and when, and the score beside it. Since v1.12, they must also hold at least 8 graded questions between them before anything is read from them, and half is also reached if the sittings whose host noted no technical fault reach it on their own, with at least that many questions. On a record made before, any supervised sitting that counted read so, whatever it scored, and the record page says so beside each standing where the rules since would read it otherwise. A difference between the remote and supervised results that no person at Wizzad has looked at yet reads “Awaiting review”, never a finding; one a person released reads “More evidence requested” or “Rests on supervised sittings”.

A reader of a record may ask for a supervised sitting and host it themselves. Such a sitting answers that reader’s request and is shown as what it is; it never counts toward a standing, because anyone may ask from an address they hold.

What Wizzad never does

A defense does not say whether AI was used to write a piece, and no AI-writing detector is run on any piece that is defended. A defense shows whether the student could answer for the work, not how the work was made.

It infers nothing about a student’s feelings, stress, confidence or honesty — from a voice, a face, a pause or anything else. The pause before a spoken answer is kept as a number in the signed record; the record page does not display it, and no conclusion is drawn from it.

It does not analyse a recording. A student may choose to record a sitting; the recording is kept as a file with its fingerprint in the record, and no face, gaze or room check is ever run on it.

It keeps no face or voice template. A passkey is checked by the student’s own device; Wizzad receives a signature and nothing of the face, fingerprint or PIN that unlocked it.

It does not check who a student is. A name on a record is as the student entered it. A signed-in account or a passkey ties a sitting to an account, which is not an identity check; only a host’s check of a photo ID in person is, and the record says which a sitting had.

It gives no grade, rank or comparison with other students, and no verdict on the student. A record is counts, dates and methods; each piece’s standing is a status read from those counts by published rules, among them a line of half the question credit in supervised sittings.

What has not yet been shown

Wizzad has not yet measured how closely its AI graders agree with expert human graders on real sittings. Until it has, read an explanation’s grade as what it is: two or three AI graders’ reading against a marking guide, with their agreement stated. The counts of facts do not depend on any grader. To measure it without reading any student’s work, Wizzad keeps a fixed set of answers written for the purpose: a person it names grades them without seeing any AI grade, and the graders are run on the same answers.

The rule that compares supervised sittings with sittings sat alone allows for the gain anyone gets from sitting a second time, by an amount taken from published retest studies, and then asks for more evidence when the gap is more than 2.3 standard errors (3.1 after two supervised sittings). Wizzad chose those thresholds; neither they nor the allowance have yet been set from its own data. That is why a person looks before any such status is shown.

Speech-to-text software makes more mistakes for some accents and voices than for others. Answering aloud is a choice: a student may type every answer instead, a typed sitting counts exactly as a spoken one does, and the student is told so beside the choice.

One sitting is a small sample. A record shows every sitting on a piece it covers, never only the best, so that no single sitting is read as the whole.

The line of half the question credit, under which a piece’s supervised sittings read “Not yet confirmed”, is Wizzad’s choice. It has not been set from data and makes no allowance for measurement error: on one ten-question sitting, a student who would earn six in ten over many sittings falls under it about one time in six. The numbers are always shown beside the status, and the results of further supervised sittings are added to them.

When something fails

If the model that writes questions cannot answer, a second model from a different provider is used, and the question set records which one wrote it. If neither can, the student is told the fault is on Wizzad’s side and no sitting begins.

If a grader cannot answer, the others’ grades stand; an answer only one grader could read is left out of the agreement count. If none can, the answer is left ungraded: it is counted neither for nor against the student, and the record says how many were left so. A grader that does not answer the conferring round in time keeps the grade it gave blind.

Asking a person to review a sitting

A student who thinks something in a sitting was asked, heard or graded wrongly can ask for a person to look at it, from the sitting itself in their Defend tab: once per sitting, in their own words. They are told at once the day a reply is due — 3 working days — and the reply names the person who gave it.

The reviewer reads the questions, the answers and the grades, and decides one of three things: the result stands; one or more answers are regraded and the results are counted again; or the sitting is set aside because of a fault in how it was asked, heard or arranged. A sitting set aside stays on the student’s record, marked, counts toward nothing, and the student may sit the piece again at once.

The sealed sitting is not edited. A decision that changes something is its own note, signed by Wizzad and naming the reviewer, and a reader is shown it beside the sitting (Note 14 of a record). What the student wrote and what the reviewer replied stay between the two of them.

Checking all of this without Wizzad

What a record’s fields mean, how each is counted and how each signature is checked are written down in a public document, the Defended Work Standard. An open-source checker — a command line and a browser page that use none of the app’s code, published by Wizzad — checks a record, and what it carries, against that document.

A student chooses to make a record and chooses who sees it, and can withdraw a link at any time. A withdrawn link shows nothing.