Integrity, layer by layer: how SkillCort protects what a score means

One rule before any feature: no signal is a verdict

Every capability below — behavioral, visual, acoustic, statistical — ends at the same place: a reviewer's screen, next to two buttons. Confirmed suspicious, or reviewed and fine. Each verdict is stored with the reviewer's identity and a timestamp, so an audit six months later shows not just what the system saw but who judged it and how.

This is not a legal disclaimer bolted onto an automated system. It is the architecture. The AI surfaces cannot change a score, block a submission, or fail a candidate — those paths simply do not exist in the product.

The browser layer: secure exam mode without downloads

Lockdown runs inside the candidate's browser — no separate lockdown application to install, no plugin. (Safe Exam Browser support remains available for institutions whose policy requires an external app.) The moment something happens, it is recorded: not on a polling interval, but on the event itself.

How strictly signals are enforced is a per-delivery choice: warn only, pause the exam, flag for review, or auto-submit at a violation limit. Proportionality is a setting you own, not an accident of the software.

• Fullscreen enforcement with a guided return path
• Tab, window and focus-loss detection — graced automatically for file dialogs and permission prompts, so ordinary exam actions never count against a candidate
• Second-display detection at entry and mid-exam
• Copy, paste, cut, right-click and print blocking; screenshot-key and DevTools detection
• Keyboard Lock (including Esc) in supported browsers
• A screen-share kill switch: stopping the share locks the exam until sharing resumes
• Virtual-machine fingerprint signals (advisory, never blocking)
• A session watermark with the candidate's identity and live timestamp on every frame

The session layer: three recording surfaces, eighteen AI observations

With the candidate's explicit consent — every capture is disclosed on the consent screen before the exam starts, and the candidate sees a live self-view of their own camera — a delivery can record webcam, screen and microphone. When a reviewer opens a session, AI scans all three surfaces and points at the moments worth opening. Each finding is a clickable timestamp that jumps the session replay to that exact second.

On the webcam, nine observations: no face in frame, more than one person, a phone or second device, looking away, earphones or a headset, actively consulting notes, smartwatch interaction, an obstructed camera, and an active screen in the background. On the candidate's screen, five: an AI-assistant window, a search engine in use, a messaging app, a remote-control tool, and unrelated windows over the exam. In the audio, four: speech during a silent exam, the question being read aloud (detected deterministically against the question text), dialogue patterns, and answer-dictation patterns.

The honesty constraints are part of the specification, not marketing. Coverage is always disclosed — a review that sampled 24 of 180 frames says so. Speaker identity is never claimed from audio; the reviewer listens to the actual recording. AI notes never quote personal message content. And the event log itself is hash-chained, so any after-the-fact edit, deletion or reordering is cryptographically visible.

The answer layer: the strongest signals need no camera

Under lockdown, question text carries an invisible marker. Copy a question out to an external AI tool and paste the generated answer back, and the round-trip is detected with certainty — not probability. It is the closest thing this field has to a smoking gun, and it works even when every camera is off.

Around it sit the behavioral forensics: large pastes into written answers, inhuman typing bursts that suggest injected text, and leave-then-paste sequences — a focus loss followed within seconds by pasted content — which are correlated into a single readable finding instead of a noise of raw events. An AI authenticity read of written answers adds a second opinion, labeled as the heuristic it is.

The cohort layer: cheating is social, so analysis is too

Collusion does not live inside one session; it lives between candidates. SkillCort compares every pair of submitted attempts on two fingerprints. First, shared wrong answers: candidates agreeing on right answers is expected — two candidates missing the same questions with the same wrong answers is not. Second, near-duplicate essays, measured with containment similarity, so copying an answer and appending an original sentence does not dilute the match.

A third list surfaces fast finishers — candidates far under the cohort's median time with at-or-above-median scores. Often that is a genuinely quick expert; occasionally it is leaked content. The panel says exactly that, because the point of every signal on this page is to start an investigation, not to end one. This layer is pure analysis: no AI calls, nothing stored, and — like everything else — nothing fed into scores.

What SkillCort will not do

Integrity tooling earns trust as much by its refusals as by its features. Four are permanent design decisions.

Everything above is published on the SkillCort Trust Center with a version number, and the statement changes in the same release as the behavior — when the product learns a new capability, the public disclosure ships with it, not after it.

• No automated verdicts — no layer can fail a candidate; decisions carry a reviewer's name and timestamp
• No automated face matching — identity photos are compared by human reviewers, deliberately kept out of the AI's hands
• No silent surveillance — every capture is organization-enabled per delivery and disclosed to the candidate before the exam
• No hidden reputation — no cross-client candidate memory, no secret scores