The seven-layer defense stack
Maestro does not police cheating. It is engineered so that real mastery is the only thing cheaper to have than to fake. Here is the full architecture, layer by layer.
The problem, stated honestly
Any on-screen task whose complete answer is visible can be outsourced to an AI tool in seconds: screenshot the item, paste it into a chatbot, type back the answer. Drag-and-drop sorting, matching, and ordering activities are solved this way trivially.
This matters twice over. Faked evidence corrupts grades — but the deeper damage is to adaptation itself. A faked Evidence Check writes a false “secure” belief into the learner model, and the system then routes the learner past the support they needed. A cheated check does not just cheat the institution; it sabotages the learner's own preparation and poisons the engine.
Maestro deliberately does not respond with lockdown browsers, screenshot blockers, or AI-detection tools. That arms race is being lost — every production detection method has known bypasses — and surveillance contradicts the platform's ethos. The strategy instead: design the system so honest evidence is easy and rational to give, dishonest evidence is expensive and self-defeating to sustain, and evidence is only ever believed when it stays coherent over time.
The authoring-side layers
Built into the content in Studio, before any learner sees it.
1 · Incentive design
Evidence Checks are never graded directly, give no correctness feedback in-flow, and always offer a penalty-free “I'm not sure” with its own diagnostic band. Outsourcing one check buys nothing — and the platform tells learners openly why.
2 · Task design
Every official check demands at least one input no outside tool has: the learner's own reasoning in their own words, their own confidence judgment, or their own history on the platform. Questions built from a learner's earlier work cannot be answered by a tool that has never seen it.
3 · Per-learner variation
Every check is drawn from a pre-generated, SME-reviewed variant pool. Shared screenshots match nobody else's item, no answer bank can form, and re-checks are a genuine re-test of the concept — not a memory test of the item.
4 · Template classification
Every template carries an AI-resistance classification. Screenshot-solvable templates physically cannot carry an official Evidence Check — the validator hard-fails them before publish. A rule the system cannot break, not a guideline.
The runtime layers
Executed while the learner works.
5 · Behavioural signals
Renderers capture how an answer was produced — latency against the learner's own baseline, focus-loss, paste-versus-typed, revisions. These signals only weight evidence and trigger support. They are never shown as flags and never accuse.
6 · Longitudinal triangulation
A single result is one witness, never proof. Mastery is a belief that consolidates only when independent witnesses agree over weeks — practice, checks, confidence calibration, re-encounters, and the controlled-conditions milestone.
7 · Dialogic verification
When evidence does not cohere, the AI Companion opens a short conversation: “walk me through how you approached this.” For an honest learner it is genuine reflection; for outsourced answers it is where the absence of understanding becomes visible. It only ever uses SME-approved probe questions.
Why the layers reinforce each other
No single layer is the defense. Variation stops answers travelling between learners, but does nothing about a learner using AI on their own item — that is the job of task design and template classification. Behavioural signals add context but never decide anything alone. Triangulation makes a one-off fake worthless, and the dialogic layer confirms what the evidence chain suspects.
To defeat the stack, a learner must produce fake evidence that stays coherent for weeks — timing, confidence calibration, reasoning quality, every re-encounter, and a matching controlled-conditions milestone. At that point, the only reliable way to fake mastery is to have it.
The honest limitation, stated openly
A learner who outsources everything, all semester, produces internally coherent fake evidence. That learner is caught by two anchors — the controlled-conditions milestone and the dialogic conversation. Those two are not optional extras; they are the ground truth the whole chain is calibrated against.