A career matcher that stays on your device

Career quizzes usually live on someone else's server

The typical online career test asks you to create an account, answer a long questionnaire, and trust that whatever profile it builds about your interests, priorities, and constraints is handled carefully somewhere you cannot see. Novus Learn's Career Matcher takes the same starting point — a structured questionnaire about what you enjoy and need from work — and keeps the entire loop on your device. Your answers are scored in this browser, the result is stored in this browser, and there is still no Novus Learn account to attach any of it to.

The output is deliberately family-level rather than job-title-level. Twenty broad career families — from skilled trades and healthcare to public safety, finance, and information technology — are ranked against your responses, each with an explanation of which dimensions aligned and which did not. Families are a starting map: wide enough to surface directions you had not considered, honest enough not to pretend the questionnaire can pick your exact occupation.

This post walks through the mechanics — what the questionnaire measures, how the ranking is computed, what the engine refuses to read, and where your data does and does not go — so you can judge the tool on what it actually does.

Three questionnaire lengths, one versioned instrument

You choose how much time to give it. Quick Discovery draws sixty items — interests, work values, activities, and goals — and targets roughly eight to twelve minutes. The Full Career Profile extends to one hundred items over eighteen to twenty-five minutes, adding work-context, skill-confidence, and constraint questions. Career Change takes the full set and appends eight transition-specific items, one hundred and eight in all, about transferable skills and what retraining you would accept.

All three modes sample from a single authored bank, published under the instrument version novus-career-v2, and every item carries its own version and publication state. Shorter modes select an evenly distributed subset across the authored order rather than just taking the first items, so a Quick run still touches the full spread of interest scales instead of overweighting whichever constructs happen to come first.

Versioning is not decoration. Each saved result records the instrument, scoring, and weight versions that produced it, so a result from today remains interpretable — and honestly comparable — after the questionnaire evolves.

How twenty families are scored

The engine is deterministic: the same answers always produce the same ranking. Your responses become a set of dimension vectors — interests, values, activities, context, skills, goals — and each career family carries a hand-authored vector on the same dimensions. Similarity per dimension is computed with cosine similarity, then combined with fixed weights: interests count most at thirty percent, values twenty, and so on down a published list.

When you skip a dimension — a Quick run has no work-context items, for example — the engine does not guess. The remaining weights are renormalised over what you actually answered, and the result names the missing dimensions outright. Each family's explanation lists its strongest aligned dimensions, its trade-offs, and any constraints applied, alongside a confidence score built from completeness and consistency rather than enthusiasm.

The ranking is then grouped: your three strongest matches, practical options that fit stated constraints, development areas worth building toward, and a few unexpected directions with real signal. Sessions from under-18 users switch to a youth exploration mode that keeps recommendations broad and never attaches salary promises.

Seventeen inputs the engine refuses to read

A matcher that models you deserves scrutiny about what it is allowed to model. The engine keeps a hard-coded list of seventeen prohibited input keys, and anything arriving under one of them is stripped before scoring begins — with the stripped keys recorded on the result so the refusal is visible, not silent.

• Race, ethnicity, religion, sexual orientation, political affiliation, and union membership are never scoring inputs.
• Health and disability diagnoses are refused; the matcher does not build medical profiles.
• Birth date, photographs, and facial features are barred — an age band exists only to switch under-18 sessions into youth exploration mode.
• Gender and sex are not inputs, and family vectors carry no gendered weighting for them to interact with.
• IQ scores, Novus Reasoning Challenge results, and any other cognitive score are excluded by name: the matcher and the reasoning challenge deliberately do not talk to each other.

Tracking that stays in this browser

Finishing a questionnaire saves a result snapshot to IndexedDB on your device, alongside the session that produced it. From there the tracking tools stay local too: save a family and it starts in the exploring stage, moving through shortlisted, preparing, and applying — or paused — as your search progresses. Each saved family can carry a target country from the supported set (Canada, the United States, the United Kingdom, the EU, and Australia), a target occupation, and free-form notes.

A comparison view holds families side by side, and a roadmap keeps dated milestones per family, all in the same local database. Deleting a result removes both the snapshot and its source questionnaire session from the device in one action. None of this synchronises anywhere, which cuts both ways: nothing to breach, and nothing to recover if you clear site data. That is the same trade the rest of Novus Learn makes, stated plainly.

Job links go out; data does not

When a family page offers current openings, it builds outbound deep links into the public search pages of Indeed, Canada's Job Bank, USAJOBS, and the EU's EURES portal. Nothing is scraped, no private API is called, and no response is stored — the destination site renders its own live results under its own terms. The only thing that travels is the search query you chose, trimmed and length-capped before it becomes a URL.

Labour-market context, where shown, is kept separate from matching by design: market conditions never alter your exploration score, and a note says so whenever local openings look thin. The same restraint applies to pay. Where Novus Learn cannot cite verified wage data, it does not display a number — an absent figure is more honest than an invented one.

What the matcher is not

It is not a validated psychometric instrument. The questionnaire items and scoring have not yet completed the project's human-review packages, so treat the output as structured self-reflection with transparent mechanics — not as professional careers guidance, and not as a measurement of your abilities. It scores stated preferences against authored family profiles; it does not observe your actual work.

It is also not occupation-level prediction. Twenty families cannot tell you to become a specific kind of engineer, and the matcher does not pretend otherwise. Use it to widen the map, compare directions you are already weighing, and keep honest notes as you investigate — then verify anything that matters against real people doing the work and real listings in your region.