MaestroMaestro Overview — contents
The Concept
Academic Integrity
Integrity is carried by the design of the platform, not by watching the learner. This topic covers how the system stays honest against cheating, against gaming the adaptivity or Atlas, and against its own misreadings — with no proctoring.
Maestro is architected around integrity by design, not surveillance: no cameras, no lockdown, no proctoring. Honesty is made the cheapest strategy; answers don't travel between learners; kept work is locked; every milestone ends in a spoken defense of the learner's own work; Atlas cannot be steered into giving answers; and behavioural signals only ever weight evidence and trigger support — they never accuse.
Integrity by design, not surveillance
Traditional integrity relies on watching the learner during assessment: proctoring cameras, browser lockdown, identity checks at exam time, timed windows, prohibited AI use. Maestro deliberately does not work this way. The platform is self-paced with no proctoring, and integrity is carried through the design itself. This isn't a gap to be plugged later — it is the integrity strategy: make real understanding cheaper to have than to fake.
Lesson checks are never graded, so there is little to gain by faking them. The grade comes from milestone work that is personal to the learner and then defended aloud. That spoken conversation about their own locked work is the ground truth the rest of the system is calibrated against.
The defense layers
No single mechanism carries integrity; the layers reinforce each other. Some are built into content in Studio before any learner sees it; the rest run in Journey while the learner works.
| Layer | How it protects integrity |
|---|---|
| Honesty economics | Lesson checks carry no grade; "I'm not sure" is always there and costs nothing; Revision needs one honest try, not a pass. There is nothing to win by guessing or outsourcing. |
| Your own thinking | Every check asks for reasoning in the learner's own words and a confidence rating. Revision's "Look again" quotes the learner's own earlier answer back to them — something an outside tool doesn't have. |
| Per-learner variants | Fresh checks and Revision draw unseen variants from a reviewed pool; milestone briefs come from the learner's chosen world, with per-learner details. A shared answer fits nobody else. |
| Locked kept work | A check's first answer is kept and never overwritten. Milestone work keeps a version history and a daily work log, and becomes locked but readable at submission. Each learner stays on the frozen course version they took. |
| Spoken defense at every milestone | After submitting, the learner answers questions about their own work by voice. It is part of the weighted report. |
| Belief over time | One clean result never settles mastery; an idea must survive a later return. Outsourced understanding tends to collapse on the next encounter. |
| Signals support, never accuse | Process signals weight evidence and can prompt support. They are never shown as cheating flags and are never proof on their own. |
Honesty economics
The cheapest path through Maestro is the honest one, by design:
- No grade on lesson checks. A check produces one of five plain results and shapes the next step. Faking a "Solid" only routes the learner past support they needed.
- "I'm not sure" is rewarded, not punished. It records no mastery judgement at all and opens a warm "let's make sure together" path.
- No correctness shown in-flow. The result appears only after submission, on the result card.
- Light gates. Before a milestone, each required Revision item needs one honest try; passing isn't needed. "Welcome back" accepts any honest answer.
- Openness. The platform tells learners plainly how this works.
Per-learner variants
Each lesson check has a pool of variants — same idea, same rubric, different story — every one reviewed and approved in Studio. A learner's fresh checks, Revision questions and look-backs always draw a variant they haven't seen, so screenshots and answer banks are worthless and a re-check is a real re-test. At milestone level, each learner chooses a world for the brief; worlds are approved with an equivalence check, and some tasks carry per-learner material (for example, a text to audit) so no two submissions answer quite the same thing.
Locked, kept work
The record is the learner's own and it doesn't move. A check's first answer is "kept for you" and read back exactly as written; later attempts are listed beside it, not on top of it. Milestone work autosaves with a version history and a daily work log, so the process is visible, not only the product. At submission the work becomes locked but readable — it is what the defense is about. A released course version is frozen, and each learner stays pinned to the version they took, so a later change to the course never rewrites what they were assessed on.
Spoken defense at every milestone
Every milestone ends in a spoken defense, always after submission. The page splits: the learner's locked work on the left, the examiner's questions on the right. The conversation is live voice; the learner can cut in and the examiner stops to listen. The questions are about this learner's work, so knowing it is the only preparation that helps. Code decides every question; the voice model only speaks the exact question text and transcribes the answer. The transcript is stored — no audio is kept — and the defense is a criterion in the weighted report, marked from the transcript. It is not an interrogation and is never framed as an integrity check — for an honest learner it is a chance to show their thinking.
Resisting Atlas steering
Could a learner steer Atlas, the learning companion, into giving them answers? The attack is architecturally constrained, because Atlas is not an open assistant with answers to extract:
| Constraint | What it enforces |
|---|---|
| It holds no answers | Atlas's grounding is an allowlist of the learner's own course facts in words. Check answer keys, rubrics and milestone work are not in it, and it has no tools. |
| Hard rules | Never reveal, confirm or hint at a check or milestone answer, before or after submission; never answer a Revision item. The rules are versioned in code and outrank anything the learner types. |
| The help ladder | Help starts light — point to the part, ask back, bring back the learner's own work — and a worked example is always on a different problem. |
| Quiet zones | Atlas is absent while a check is open and inside assessments; the server refuses chat there, not just the screen. |
| Checked replies | Every reply is linted before the learner sees it; a reply that breaks a rule is refused and re-asked. |
Attempts to change Atlas's rules get a brief, calm decline and a way back to the learning. More in Atlas.
Resisting mastery gaming
The deeper protection is against a learner producing plausible responses without genuine understanding. Several mechanisms — described in the lesson's check — do the work together: reasoning judged against a hidden rubric, confidence calibration, declared tempting errors that surface specific mix-ups, bounded fresh tries, and a belief that only settles after an idea survives a later return. A fluent wrong answer still shows itself in the reasoning, the confidence mismatch, or the mix-up that recurs.
Behavioural signals
Checks capture light process signals already present in the interaction — time per step, focus changes, pasting and its length, revisions — and feed them in as supportive evidence, never as verdict. They are read against the learner's own baseline, not universal thresholds. A large paste, an unusually fast answer or repeated focus loss lowers the weight that one piece of evidence carries; a single small paste means little.
No cameras, no microphone monitoring outside the defense the learner starts, no screen recording, no browser lockdown. Signals can only reduce how much weight a piece of evidence carries — never to zero — and they never reach the learner as an accusation. The same signal that might suggest outsourcing also describes an anxious learner; the response to both is support. Anything staff see informs human judgement; nothing is automated into a finding, and raw signals never leave Maestro.
The check runners capture process signals with each submission; the engine's integrity weigher turns them into a bounded weight (with a floor) applied in the belief update. Milestone packs are frozen per build, and each task carries an AI posture (off, retrieval only, sparring partner, critic, assisted and disclosed) shown to the learner as a neutral label. See Learning Signals for the inventory and Lessons, Checks & Milestones for the build view.
Identity via LTI
Identity verification is not a separate concern in this model. LTI 1.3 launches deliver verified identity from the institution's LMS at course entry, tied to student registration. Identity is established at launch and persists through the session — there is no assessment moment where the learner's identity is uncertain. (See Integration.)