Trust Score

The Integrity Budget: ProctorSafe's Trust Score

A 0–100 score that aggregates every integrity signal into a single reviewable metric — computed by a deterministic, auditable algorithm, not a black-box model.

proctoring trust scoreexam integrity scoreonline exam risk scoreintegrity budget

0–100 Integrity Budget

Starts at 100. Deductions are severity-weighted (CRITICAL, MAJOR, MINOR) and normalised by session duration.

Per-exam-type scoring profiles

Different assessments carry different risk profiles. Configure separate scoring weights for a timed coding test versus a multi-hour essay exam — and compare scores across profiles for the same session.

Transparent, not black-box

Every deduction is tied to a specific event with a timestamp. Reviewers can see exactly why a score is what it is.

Why a trust score matters

Reviewing every proctoring session manually is not scalable. For a cohort of 1,000 candidates, that means 1,000 event timelines — most of them clean. The trust score exists to answer one question quickly: which sessions deserve a human reviewer's attention?

A well-designed trust score is not a verdict. It is a triage tool — one that should be transparent, auditable, and contestable.

How the Integrity Budget works

Starting value and deductions

Every session begins with a score of 100. As integrity events are detected, points are deducted according to a configurable severity table:

Event type	Default severity	Example deduction
Remote desktop detected	CRITICAL	35 pts
Virtual camera confirmed	CRITICAL	35 pts
Second speaker suspected	MAJOR	20 pts
Sustained face absence (>30s)	MAJOR	15 pts
AI domain access (ChatGPT, Claude…)	MAJOR	20 pts
Screen sharing detected	MAJOR	20 pts
Tab switch (sustained)	MINOR	5 pts
Brief gaze-away	MINOR	2 pts

How the score is computed

The trust score engine runs server-side as a deterministic, rule-based algorithm. This means:

No black-box model

the score is not derived by a machine learning model that cannot be inspected; it is computed by a transparent algorithm with configurable parameters

Reproducible

given the same event stream and severity configuration, the score will always be the same

Auditable

every deduction maps to a specific event in the session timeline, with a timestamp and a reason code

Scoring profiles for different exam types

Not all assessments carry the same risk profile. A 15-minute formative quiz and a 4-hour professional certification exam may warrant very different triage thresholds and severity weights. ProctorSafe supports multiple named scoring profiles per tenant. Each profile defines its own severity weights, deduction amounts, and normalisation parameters. You can:

Apply different profiles to different exam types at configuration time
View scores computed under all active profiles side by side for a single session
Compare how a borderline session would be triaged under your standard policy versus a stricter high-stakes policy

This makes it straightforward to run pilots with new scoring parameters without committing to a single global policy.

Reviewer integration

The trust score feeds two reviewer workflows:

Bulk triage

The sessions dashboard sorts and filters by trust score. Reviewers can:

Auto-clear sessions above a configurable threshold (e.g. score ≥ 90)
Queue sessions in the 60–90 range for spot-checks
Flag sessions below 60 for full investigation

Session detail

Inside a session, reviewers see the score broken down by contribution. Each deduction is linked to the triggering event with a timestamp. Reviewers can step through the event timeline and understand exactly why the score is what it is — not just that it is low. Every reviewer decision (clear, spot-check, refer) is logged in an immutable audit trail with a timestamp and mandatory comment.

AI-assisted triage

In addition to the trust score, ProctorSafe runs an LLM analysis (powered by Mistral AI, hosted in the EU) over the session's event timeline. The analysis produces:

A risk level (LOW / MEDIUM / HIGH)
A plain-language summary of the session
Highlighted events that most influenced the assessment
Recommended actions (clear, spot-check, investigate, void)

The LLM operates on metadata and events only — no raw video or audio. The trust score and AI summary are shown together in the reviewer interface, giving reviewers both a quantitative signal and a qualitative narrative.

Get started

Try the interactive demo to see trust scores in action
Contact us for a walkthrough with your specific assessment type
Read our signal intelligence article

Related guides

Dive deeper with guides from our articles library.

Technical architecture

Solving the Latency Gap: Why On-Device Analysis is Crucial for High-Stakes Exams

Why network dependency breaks fairness and how on-device monitoring keeps detection consistent under real-world connectivity.

Read article Pedagogy & ethics

Building a Culture of Integrity: From Surveillance to Signal Intelligence

A modern integrity philosophy: extract only what matters, preserve dignity, and keep humans accountable for outcomes.

Read article Pedagogy & ethics

Ethics in AI Proctoring: Bias, Fairness, and the Human-in-the-Loop

How to design and operate proctoring that avoids biased outcomes, documents limitations, and keeps humans accountable.

Read article

Frequently Asked Questions

Common questions about privacy, compliance, and integration.

What is the ProctorSafe trust score?

The Integrity Budget is a 0–100 score that starts at 100 and decreases as integrity violations are detected. Each violation type has a severity (CRITICAL, MAJOR, or MINOR) that determines the deduction. The score is normalised by session duration so that a brief violation early in a long exam weighs differently than the same violation in a short test.

How is the trust score computed?

The score is computed server-side by a deterministic algorithm. It processes each integrity event, applies severity-weighted deductions, groups temporally proximate incidents into single violations, and normalises the result against session duration. Because the algorithm is rule-based rather than model-based, every deduction is fully explainable.

What severity levels exist?

Three: CRITICAL (e.g. remote desktop detected, virtual camera confirmed), MAJOR (e.g. second speaker suspected, sustained face absence), and MINOR (e.g. brief tab switch, short gaze-away). Each maps to a configurable point deduction.

Can reviewers see why a score is low?

Yes. The reviewer dashboard shows a ranked event timeline alongside the trust score breakdown. Each deduction is linked to the specific event that caused it, with timestamp and severity.

Can the trust score be configured per tenant?

Yes. Severity weights and deduction amounts are configurable per tenant. An institution can increase the weight of AI tool detection while reducing the weight of brief tab switches to match their specific threat model.

Can we use different scoring profiles for different exam types?

Yes. You can define multiple scoring profiles — for example, one for high-stakes certification exams and another for routine formative assessments. ProctorSafe can compute and display scores under all active profiles for the same session, making it easy to compare how a session would be triaged under different policies.

Ready to Get Started?

Start a free trial and run real sessions against your own application, try the interactive demo in your browser, or contact us for a walkthrough tailored to your program.

Start Free Trial Try the Demo Contact Us

Questions? We're here to help.

[email protected]About Privacy Policy

Solutions

Compare

Articles

Guides on privacy-first proctoring, compliance, and exam integrity.

View all articles →