How to read a Wizzad record

What a record is

A record is a digitally signed document of one student’s study activity in Wizzad: what was asked, how the answers were checked, and what was recorded — counted at the time and fixed when the record was issued. The figures do not update. The student chose to make it and chose to show it to you; they can withdraw the link at any time, and a withdrawn link shows nothing.

Wizzad checked control of the student’s email address at sign-up and nothing else about who they are. A name on a record was entered by the student; the record says so on its cover and in Note 7.

How it is laid out

The cover names the student, the period covered, the record number and the date of issue. Part I is the record: Item 1, a summary in one paragraph; Item 2, a statement of results by defense — one column per defense and an “All” column; Item 3, each defended piece with its own statement of results and, where answers were spoken, the recordings; Item 4, practice by topic and, where the record carries one, a statement of estimates: the figure on the student’s own map for each topic at issue, beside what Wizzad’s model expected of the answers after a gap and what came back, and, for the defenses that asked about the topic, what the model expected of those questions and how many connected. Part II is the Notes, 1 to 9, which say what each count is and is not. Part III is the statement on the record’s integrity, how to check it, the technical details, and a key to terms.

Every number in it is a count beside its total, except the estimates in Item 4, which the record names as Wizzad’s own. A dash means none. Nothing is compared with other students.

Dates are the student’s own: each is the day it was where the student was when it happened, and the cover names that time zone beside the date of issue — “New York time”, say. A sitting, a file or a run dated in another zone names its own. A record issued before zones were kept, or with none known, gives UTC days and says “UTC”. The moment itself is kept in UTC, and the zone is covered by the signature.

The summary and the practice lines

“N practice questions across N topics on N separate days; N of those were checked against an answer key rather than self-rated.” Practice questions are answered on the student’s own device against the material they were studying. “Checked against an answer key” says how the result was produced — marked against the key held with the material; it is not a statement that Wizzad checked it. The rest were rated by the student.

“Of the N answers that came after a gap, when a topic had been left to fade, N were correct (N%).” The line that carries weight. Wizzad schedules review; when it asked about a topic that had been left long enough to fade, did the student still have it? Below five such answers there is no rate — too few to give one.

“N practice items named the topic they belong to; N were matched by similarity.” How each answer was tied to a topic. A named tie is exact; a similarity match is a best fit above a floor. If you distrust matching, ask for the link with the answers included and recount without them.

Practice is not an examination. It is volume and recall, dated.

The estimates

“Estimate at issue: 72.” Wizzad keeps, for each topic, an estimate from 0 to 100 of how well the student holds it — the same figure the student sees on their own map, shown as it stood on the day of issue. It is Wizzad’s figure, moved by the student’s own answers — up when a topic came back, down when it did not, by less when the answer was self-rated — and it eases as time passes without an answer. Where the map gives no figure yet, the record shows none.

“Expected after a gap: 4.1 of 5 · came back: 2.” Before each practice answer given after a gap, Wizzad’s model put a chance on the topic coming back; added up, those chances are what it expected. Beside it is what happened. A chance is on record only for answers given since Wizzad began keeping it, and the line under the statement says how many answers that leaves out; it also names the topics furthest from what the model expected, as numbers.

“Asked in a defense: 1 of 2 · expected in defenses: 0.7 of 2.” The defenses’ own count of questions about the topic and how many connected, then what the model expected of them. For each history question in a defense, the model’s chance that the topic would connect is kept as the model held the topic when the question was answered — from the answers it had taken in by then — and set against what happened. A question asked where the map gave no estimate yet, or before the chance was kept, carries none, and the record counts it apart. A second line under the statement sets the two against each other across the topics, the same way. An answer given in a defense is never counted as practice.

A difference is information, not a finding: a model can be wrong in either direction, and it says nothing about why. Read it as the record’s own check on Wizzad’s model — where it was right, and by how much it was not.

Defended work

A defense is a timed session about one piece of the student’s own work. The questions were built from the piece itself; each points to a passage the piece contains, or it was thrown away. The student answered one question at a time, 90 seconds each, 15 minutes in all — or more where they chose 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 — with the piece out of sight — the only text shown was the short excerpts the questions were about. Every attempt is kept and shown in order; a later attempt never replaces an earlier one.

“N of N factual questions answered correctly.” Questions with a set answer — a number, a term, a choice, a detail changed in an excerpt for the student to put right — compared automatically with the student’s answer, with no AI judging it. Most set answers are taken from the piece; a what-if question’s was worked out by the AI from the piece’s reasoning when the question was written.

“N of N explanations graded as full.” Reasoning given in the student’s own words, graded by an AI grader against a marking guide written from the piece’s own reasoning, and graded again by one or two other AI graders, each blind to the others’ grades. Where their grades differ they confer once — each is shown the others’ scores and reasons and asked to reconsider its own — and the grade of record is the middle one of three, or the first grader’s of two. “Three AI graders agreed on N of N (M of N before conferring)” says how often every grader that read an answer gave it the same grade, after conferring and before; an answer only one grader could read counts in neither number, which is why the second N can be smaller than the number of explanations. On a record made before the graders conferred, the line reads “a second AI grader agreed on N of N”. No human grader is involved.

“N of N changes to the piece accounted for.” For a piece written in Wizzad, its saved states are kept, and the student was asked why they cut, rewrote, moved or added particular passages — quoted from the state they lived in. The final piece does not contain the answer; what was cut, or what a rewritten paragraph used to say, is in the saved states alone. A reason that fits the change that actually happened counts as accounted for. “N of N things studied that the piece names connected”: topics the piece names that the student had studied in Wizzad, and whether they could say how each bears on the piece. A piece made elsewhere has neither line, and the record says so rather than inventing them.

“N answers late, N skipped.” A late answer arrived after its window and is not counted as answered; a skipped one had no answer. A question left blank is counted as unanswered, never as wrong.

“Identified by passkey” means the student confirmed presence with a passkey on their device when the session began; “identified by signed-in account” means only the usual sign-in. Neither is an identity check. “Left the session window N times · pasted text N times” were counted during the session and change no result; they are reported for you to weigh.

“N of N answered aloud.” When the student chose to answer aloud, each question was read to them and the clock started after the reading; the answer was recorded, written down automatically, and graded as the same words typed would be. Each spoken answer is on record, and Item 3 includes a speaking sample — clips chosen by question type and order, never by how well they were answered — so you hear the answers as they were spoken. Wizzad does not check whose voice it is; a sitting with a host says whether a photo ID was checked. The pause before speaking is kept as a number in the signed record that Wizzad draws no conclusion from; the record page does not display it, and no answer is timed out for thinking.

Where the record says answers were given “to the AI examiner”, the student chose to be examined by an AI voice and face. It read the questions built from the piece and listened; it did not grade, help, or add a question of its own, and the words it heard were graded exactly as typed ones. The student was told it was an AI before it spoke.

“The student recorded the session (camera and screen, N min).” The student chose to record the session, and the recording is on record: its fingerprint is inside the signed record, so a copy shown to you can be checked against it on the check page. Wizzad does not analyse recordings — no face, gaze or room check, ever — and a recording cannot show what was outside the frame.

A student may ask a person to review a sitting afterwards. A review that changed a result is shown beside the sitting, signed by Wizzad and naming the reviewer; how a review works is set out on the page “How a Wizzad defense works, and what it does not claim”.

Pieces written in Wizzad, and pieces made elsewhere

“Written in Wizzad over N sittings on N days — N saves between two dates. Of everything added, N% was typed, N% pasted, N% written with Wizzad’s assistant. The version defended is one of the saved states.” For a piece written in Wizzad, each saved state is kept while the piece is open, with how much was added and by what route. The last sentence says whether the version the student defended is among those states; if it is not, the piece was changed somewhere Wizzad did not see, or before this history began.

This history describes how the piece changed while open in Wizzad. It does not show who was at the keyboard, or what was written somewhere else first. Read it beside the defense, not as proof on its own.

“N of N end-of-session questions answered, each within a minute of being asked.” When a writing sitting ended, Wizzad put one question to the student about the newest text — what a passage they had just written is there to do — to be answered within a minute, in their own words, and kept with the piece. It is not graded. A question not answered is counted as asked, and nothing more.

“The student brought this piece into Wizzad as a Word file. The file is on record.” A piece made elsewhere came in as a file or as pasted text. The file, exactly as it arrived, was fingerprinted into the signed record; if you hold a copy, the check page tells you whether it is the same file, without uploading it. Wizzad has no history of how such a piece was written and claims none.

The editor also keeps counts of how the typing ran — the pauses, the bursts, how much was removed. They are shown to the student on their own screen and are not put on the record, until Wizzad’s own trial has produced ranges to read them against.

What it does not establish

It does not say whether AI was used to make any piece of work. It says whether the student could answer for the work, timed, without it in front of them, and whether what they studied stayed with them.

It does not say who the student is beyond control of an email address, and it does not say who sat a session; a passkey confirmation raises the bar without settling it.

It does not draw anything from a recording, and it does not say what a pause before an answer means.

It is not a grade, a ranking or a comparison with other students. It is counts, dates and methods — and, where it carries them, Wizzad’s own estimates, named as such and set beside what happened. Wizzad does not add a verdict, so that you can apply your own.

How to discount what you distrust

If you discount answers graded by an AI grader: read only the factual-question counts in each defense, and the checked-against-an-answer-key count in the summary — but note that the changes-to-the-piece count rests on facts only the author could know, whatever graded the reason.

If you discount similarity matching: ask for the link with the answers included, drop every item matched by similarity, and recount; the record says how many there are.

If you discount self-rated practice: the summary says how many answers were checked against an answer key; treat the rest as the student’s own account.

If you want the strongest signal only: the factual-question counts, the after-a-gap rate, the speaking sample, and the dates.

How to check it yourself

Every record is digitally signed with a key whose public half Wizzad publishes. The check page confirms the signature and shows the key it used; it takes a few seconds and needs no account. If you hold a copy of a piece the student brought in, or of a recording, add it there: it is read on your own device and only its fingerprint is compared.

You can also confirm the record has not been edited since it was issued using any Ed25519 library, with no account and without asking Wizzad anything: fetch the record’s JSON, fetch the key, serialise the record’s payload with object keys sorted at every depth, and check the signature over those bytes. The technical details at the end of every record spell this out. The whole point of the signature is that you do not have to take Wizzad’s word for it.

A task set by a reader and finished is also issued as an Open Badges 3.0 credential — a W3C Verifiable Credential with a Data Integrity proof, signed with the same key under Wizzad’s did:web identifier. A system that verifies such credentials checks it without this page; it carries the same counts and the same sentences as the record.