AI-assisted marking
Every script is marked against the scheme with a fixed seed and no temperature, so the same answer earns the same mark twice.
Server-side · deterministic
Edexcel IGCSE 4MA1Examiner-level marking
Mark smarter.Trust every mark.
AI-assisted mathematics marking, built around examiner-level consistency, transparent corrections and human oversight.
- Correction runs on the server
- A human decides every exception
- One workspace, one boundary
igcse 4ma1 · algebra ii
Homework 6
34 / 48 submitted
- Q1Expand and simplify2/2Approved
- Q2Factorise fully3/3Approved
- Q3Linear inequality2/4Needs review
- Q4Rearrange the formula3/3Approved
- Q5Simultaneous equations4/4Override
- Q6Algebraic fractionsQueued
Correction job
3f2a…c19
Class median
78%
A marking console you can read at a glance.
A designed preview of the interface the rebuild is working toward: assignment state, submission progress, question-level marks and the exceptions that need a person. Every figure below is synthetic demonstration data.
workspace / assignments / homework 6
Demonstration dataIGCSE 4MA1 · Algebra II — Homework 6
48 submissions · 10B · Set 2 · Scheme v3 · locked
MarkingNeeds review
Submitted
34/48
Marked
29/34
Needs review
2
2 assigned to youMedian score
78%
Score distribution out of 30 marks across 29 marked scripts: 0–4 marks, 1 scripts; 5–9 marks, 2 scripts; 10–14 marks, 5 scripts; 15–19 marks, 9 scripts; 20–24 marks, 8 scripts; 25–30 marks, 4 scripts.
Review queue
- Q3D. HaddadMethod marks present, final answer missingNeeds review
- Q7F. MbekiHandwriting unclear on the final lineNeeds review
Marking activity
- 09:47Job 3f2a finished · 7 of 7 questions reconciledApproved
- 09:46Q4 override 3 → 4 recorded with a reasonOverride
- 09:44Q3 parked for review · answer matched two methodsNeeds review
- 09:43SUB-0417 queued · scheme v3Queued
Correction jobs
staged progress · polling- SUB-0418A. OkonkwoQ5 of 7 · verifyingMarking
- SUB-0419B. FerreiraQueued · position 2Queued
- SUB-0420C. NakamuraQueued · position 3Queued
- SUB-0416D. HaddadQ3 · mark below method totalNeeds review
- SUB-0412E. LindqvistApproved 09:41 · J. AdeyemiApproved
Queued scripts show a skeleton until the server has something to report. A missing question becomes Needs review — never a silent zero.
Question-level results
SUB-0416 · D. Haddad| Q | Topic | Max | Awarded | Class mean | State |
|---|---|---|---|---|---|
| 1a | Expand and simplify | 2 | 2 | 1.8 | Approved |
| 1b | Factorise fully | 3 | 3 | 2.4 | Approved |
| 2 | Linear inequality | 4 | 2 | 2.9 | Needs review |
| 3 | Rearrange the formula | 3 | 3 | 2.2 | Approved |
| 4 | Simultaneous equations | 5 | 4 | 3.6 | Override |
| 5 | Quadratic by factorising | 4 | 4 | 3.1 | Approved |
| 6 | Algebraic fractions | 5 | 0.0 | Queued | |
| 7 | Proof by completing the square | 4 | 3 | 2.5 | Approved |
An override keeps the original mark beside the replacement and records who changed it, when, and why. Totals are recomputed from the question set.
Synthetic demonstration data. No live workspace is shown.
Six steps from a stack of scripts to a published grade.
This is the intended workflow of the rebuild, drawn from the product blueprint. Each step names what it produces, because a marking system should always be able to say what state a script is in.
- 01
Create the assignment
Name the paper, choose the class, attach the mark scheme. The scheme is read into a fixed question list with its official maximum marks.
Assignment · scheme blueprint
- 02
Upload the submissions
Add a class set of scans, or pull them from Google Classroom. Each file is validated, then bound to one student and one assignment.
Submissions in R2
- 03
Let the pipeline mark
A queued job prepares the script, marks it against the blueprint and reconciles every question. Nothing is marked in the browser.
Provisional marks
- 04
Review what is uncertain
Anything ambiguous arrives as Needs review, never as a zero. Change a mark, record the reason, and keep the original beside it.
Reviewed results
- 05
Approve the results
Approval recomputes totals from the question set and writes an audit entry naming who approved what, and when.
Approved marks
- 06
Report, then publish
Generate the class report, then optionally attach it and publish draft grades to Classroom as an independent step.
Report · Classroom
Built for the judgement, not just the throughput.
Eight capabilities define the product. None of them is claimed as live on this page: Aetrix V5.5 is a pre-release build, and the marking pipeline is specified and in active development.
Mark-scheme blueprint
The question list and its maximum marks are fixed once per assignment, then enforced when marks are written.
Locked per assignment
Human review and overrides
A person decides every mark that matters. Overrides keep the original value, the reason and the reviewer.
Audited · reversible
Assignment management
One place for the paper, the scheme, the class and the state of the marking, from draft through to published.
Lifecycle · owned
Class and student records
Rosters and per-student history stay inside the workspace. Nothing is shared across a tenant boundary.
Workspace-scoped
Correction-job monitoring
Each job reports its stage, its attempts and the reason it parked. A failure is a state with a next action, not a spinner.
Queues · idempotent
Reports and analytics
Question-level analysis, distributions and trends, generated from the approved marks rather than inferred from them.
Report · class level
Google Classroom
Courses, coursework and submissions can be read in; reports and draft grades can be written back, each as its own operation.
Integration · optional
Trust is a property of the design, not a badge on the page.
These are the intended principles of the rebuild. They describe commitments the architecture is being built to hold — not certifications, and not a claim that every control is already deployed.
What this page does not claim
- No SOC 2, ISO 27001 or GDPR certification is claimed. None has been obtained.
- No pricing is published, because plans are not finalised.
- No account can be created from this site yet.
- 01
Workspace isolation
Every query is scoped to a single workspace in the typed data layer, and the database rejects a cross-tenant reference rather than trusting the handler to remember.
- 02
Correction stays on the server
Prompts, mark schemes and provider calls never reach client JavaScript. The browser asks for a result; it never performs the marking.
- 03
No provider credentials in the browser
Provider keys are encrypted at rest and decrypted only inside the Worker for the duration of one outbound call.
- 04
Overrides are auditable
A changed mark records the reviewer, the time, the previous value and the stated reason. Approvals are version-checked, so a second save cannot silently overwrite the first.
- 05
Questionable results are reviewed explicitly
A missing question becomes Needs review, never zero. A mark cannot exceed the blueprint maximum, and that limit is enforced by the database as well as the handler.
Marking that can answer for itself.
Every mark traceable to a question, a scheme and a person. That is the standard the rebuild is being held to.