Home / Validity / Match

DecipherU Match technical manual

Written for an outside reviewer. It follows the headings of the EFPA test review model and the 2014 Standards for Educational and Psychological Testing. Where nothing has been measured, it says so.

1. Purpose, intended uses and non-uses

Match helps a person decide which licenses and certifications to look at first. It compares their answers with a written profile of each one and shows where the answers line up and where they do not. It is for self-discovery and exam practice.

It cannot tell you whether you will pass an exam, get a job or like the work. Fit says how your answers compare with the O*NET occupation profile behind each license, not with answers from people who hold the license. It is a soft signal. In large studies of hiring, interest scores relate to job performance with a correlation of about .24, and the spread across studies is wide (Sackett et al., 2022). That figure is about job performance, not about how well a person fits a license.

The fit bands are relative. Strong fit means near the top of the licenses compared. It does not mean a high level of fit in absolute terms. We will set absolute levels after the pilot.

Non-uses

Match is not a hiring, admission or licensing tool. Employers, schools and agencies may not use it to screen, rank or decide about anyone. There is no employer view and no feed of results, and we do not sell or share results.

Match is also not a medical, clinical or psychological assessment, and it does not measure intelligence, personality disorders or any health condition. It collects no health or disability data.

2. Constructs and items

Every item was written for Match. No item is copied from a commercial test or from the O*NET Interest Profiler. The question screens total 79, about 14 minutes, computed from the per-screen seconds in the instrument. The quick form has 36 screens, about 8 minutes.

Reasoning items per ability: quantitative 3, verbal 3, pattern 3, spatial 0, clerical 3. Answer keys are held on the server and are not published here. The reasoning block enters the fit formula as one part.

Item development: items were drafted from the constructs above, not from another test. Cognitive interviews on wording and reading level are planned and have not been run (see section 6). Match is available in English only; Spanish and Portuguese need a test adaptation process, not translation alone.

3. Scoring and the engine

Each block is scored into a profile. Each license or certification has a profile built from O*NET and BLS data (section 8). For every profile the engine computes five distances between the person and the license, each scaled to 0 to 1 (0 is a perfect match), then combines them:

fit_raw = 1 - (0.45 x interests + 0.3 x abilities + 0.1 x work style + 0.1 x values + 0.05 x kind of work) / (sum of the weights of the blocks that were measured)

The weights sum to 1. They are placeholders chosen before any data existed and will be re-estimated in the pilot. When a block was not measured (no reasoning items answered, no value answers), its weight is dropped and the rest are rescaled the same way for every license.

From raw fit to the percentile and the band

• Some licenses ask little of anyone and score high for almost every profile. To correct for this, each license's raw fit is placed against how that same license scores for 1,000 seeded, uninformative answer sets (seed 20261011). Licenses are ranked on that position. The raw fit is still reported.
• Bands are relative to the licenses compared: strong at percentile 80 or higher, good at 55 or higher, otherwise possible. They are not absolute levels of fit. Absolute levels wait for pilot data.
• A result is shown as a band with a range from a weight-sensitivity re-run, not as a single number.

Gates

A gate either excludes a license or adds a plain-sentence flag. The only exclusion is a minimum age that a licensing board publishes, and only when the fact carries a source. Education, time, money, exam language and a record question add flags and never remove a license. A missing answer never excludes and never flags.

Choosing the cards

• At most 5 cards, and no more than 2 that map to the same occupation.
• An anchor card (the person's own pick) only when it is in the strong band.
• Surprise cards are drawn from the person's own best absolute matches: within 0.1 of their best raw fit and among their top 6 by raw fit, with a floor on viability. They are ordered by surprise, which is fit percentile minus how expected the license is, and a license the person ruled out gets 20 extra points in the ordering only. If that leaves open slots, a second tier with looser limits (within 0.18 of the best raw fit, top 12) fills them.
• "Least aligned" shows only when the raw fit is at or below 0.4 and the answers were not rushed. It never says the person could not do the work.
• Viability (pay, openings, growth, time and cost) is used to rank and display, never inside fit.
• If the interest answers vary less than 0.8 (standard deviation), or other answers carry too little information, Match returns no cards and stores nothing.

The source of every constant above is packages/shared/src/match/engine. This page imports them, and a test fails if the page and the code disagree.

4. Samples and norms

There are none yet. No pilot has run, so there is no norm group and no sample to describe. The percentile in a result compares the person with the other licenses and certifications in the list, and with the 1,000 synthetic answer sets above. It never compares the person with other people.

The pilot plan (sample sizes and what each stage checks) is on the method page.

5. Reliability

No reliability estimate exists yet. We have computed no internal consistency, no test-retest correlation and no standard error of measurement. What does exist is a consistency check inside the instrument: 2 of the kind-of-work pairs repeat an earlier pair with the sides swapped, and a very fast answer pattern marks a result as low confidence. Neither is a reliability coefficient.

6. Validity evidence

None yet for content, internal structure, relations to other measures, relations to outcomes, or consequences of use. The following is what has been done and what has not.

Done: second-coder check of the license profiles (2026-10-12)

A second coder independently coded the occupation codes, job zones, kind-of-work tags and work values for the 34 license families, then compared. The launch gate in that note is Cohen's kappa of at least 0.80 on every field. Verdict: the gate is not met. 8 of 12 gated fields fail it.

• n is 34 license families, so every kappa has a wide interval.
• Both coders are the same model family (Claude), so shared blind spots are not caught.
• The pass on occupation codes and job zone is a floor on quality, not proof. A credentialed human reviewer has not yet checked the Klimov tags.

The occupation mapping, which the build spec gates on, passes. The kind-of-work tags and the work values do not, so the weights on those parts are the least trustworthy. Fixing the tags, or dropping those parts of the weight, is a condition for making Match public.

Planned, not run

1. Cognitive interviews with 15 to 20 people on wording and reading level. Pass: no item misread by more than 2 people.
2. Pilot A with 300 or more completers, to check structure, reliability and timing, with an established interest inventory as a second measure. We use such an inventory only with its publisher's written permission. Match itself uses only items we wrote, so it needs no one's permission to run.
3. Pilot B with about 100 retesters at 2 to 3 weeks, to check that results hold steady.
4. A swap and acceptance study with 40 to 60 people: can you pick your own report over a random other one at least 75 percent of the time, and does no statement get accepted by more than 85 percent of takers.
5. A predictive study with at least 190 completers per outcome, asking whether fit relates to practice score, completion and the exam result at 6 to 12 months, beyond what prior practice score and education already tell us, and whether it works equally across groups.
6. A known-groups check with existing license and certification holders: do they land closest to their own credential?
7. Calibration before any pass percentage appears: at least 200 passes and 200 fails per exam family. Until then Match shows no pass percentage at all.

Beta, not validated. The weights are placeholders, the mapping from licenses and certifications to occupations was done by one coder and has not been double-checked, and many time, fee and record-policy facts are not checked yet. This beta is also our first study.

7. Fairness and accessibility

What runs

• No protected attributes as input. The engine has no input for name, sex, gender, race, religion, disability or health. The situation questions are: state, budgetUsd, months, hoursPerWeek, age, record, education, languages. Age is used only for the sourced minimum-age gate, state only to pick state-specific facts, and the record question only adds a "check with the board" note.
• Hubness test. A test runs random answer patterns through the engine and fails if any one license comes first for more than a set share of them. It exists because a license that fits everyone would otherwise crowd out the rest.
• Insufficient-information stop. Straight-lined, blank or very fast answer patterns return no cards instead of a guess.
• Counterfactual on preference. Where a card has one, the ledger re-runs the fit with one preference shifted and says how the rank would change. This is an explanation aid, not a fairness test.
• Minors. Under 13 stores nothing; 13 to 17 results are kept only in the browser tab.
• Accessibility. One question per screen, no time limit, keyboard operation, labels on every control, reduced motion and forced-colors support. Automated checks (axe-core) and keyboard checks ran on 2026-10-12 against WCAG 2.2 AA.

What does not run yet

• No analysis of group differences in results, because we collect no sex, race or similar data and no pilot exists. A pilot would have to ask for it with consent.
• No differential item functioning analysis on the reasoning items and no check of how scores relate to outcomes by group.
• No recorded screen-reader test (VoiceOver, NVDA).
• No external psychometric review.

8. Data sources and licenses

Data release

12 U.S.C. 5104, SAFE Act license requirements+AICPA CPA exam FAQ+ANAB food handler certificates+ARRT application fees+ARRT ethics review preapplication+ASWB exam registration+AWS Certified AI Practitioner page+AWS Certified Machine Learning Engineer - Associate page+AWS Certified Security - Specialty page+Arizona Board of Barbers fees+Arizona Board of Cosmetology exam information+Arizona Board of Cosmetology fees+BACB BCBA Handbook+BACB RBT Handbook+BLS Employment Projections Table 1.2, 2025-2035+BLS OEWS May 2025 national estimates+California DMV instruction permits+California DRE fees+California DRE salesperson requirements+California Department of Insurance, life agent requirements+CompTIA CySA+ certification page+CompTIA PenTest+ certification page+CompTIA SecAI+ certification page+CompTIA Security+ V7 page+CompTIA Security+ certification page+Credentia North Carolina exam fees+EC-Council CEH page+EC-Council voucher prices and eligibility+EPA Section 608 technician certification requirements+ETS Praxis Core combined test+FINRA Securities Industry Essentials exam+FINRA Series 66+FINRA Series 7+FMCSA age requirement for interstate commerce+FMCSA entry-level driver training+FSMTB MBLEx FAQs+FSMTB MBLEx requirements and process+Florida DBPR real estate requirements+Florida Department of Financial Services, agent license fees+Florida Department of Financial Services, insurance license qualifications+Florida FLHSMV teen licensing+GIAC GCTI page+GIAC GSEC page+GIAC pricing+IAPP AIGP exam store page+IAPP AIGP page+ISACA AAIA exam content outline+ISACA AAIA page+ISACA CISM page+ISC2 CC exam fee statement+ISC2 CCSP exam outline+ISC2 CCSP page+ISC2 CISSP exam outline+ISC2 CISSP page+ISC2 Certified in Cybersecurity page+ISC2 annual maintenance fees+Illinois IDFPR cosmetology+Illinois IDFPR esthetics+Illinois IDFPR nail technology+Kansas Board of Barbering+LeadingAge, CMS nurse aide training requirements+Microsoft AI-103 exam page+Microsoft AI-901 exam page+Microsoft Azure AI Fundamentals page+Microsoft SC-500 study guide+Microsoft exam FAQ (pricing)+NASAA Series 65 exam content outline+NASAA exam FAQs+NASBA CPA exam, New York+NASBA CPA exam, Washington+NBCC NCE registration form+NBRC Certified Respiratory Therapist+NBRC Registered Respiratory Therapist+NBSTSA CST eligibility+NCSBN NCLEX fees and payment+NMLS SAFE MLO testing FAQ+NREMT EMR and EMT certification examinations+NVIDIA NCA-GENL page+New York DMV learner permit+O*NET 30.2 Database: work values (dropped from 31.0)+O*NET 31.0 Database: interests, abilities, work styles, job zones, work context+PTCB Certified Pharmacy Technician+PTCB guidebook, general application information+ServSafe Manager exam, online proctor+TREC become a sales agent+TREC fee schedule+Texas DPS learner license+Texas DSHS food handler FAQ+Texas Department of Insurance, general lines life, accident and health application+Texas TDLR electrician candidate information bulletin+Texas TDLR journeyman electrician application@2026-10-12: the distinct source names behind the profiles and the latest retrieval date, the same string stored with every result.

Profiles

54 in total: 34 licenses, 20 certifications, 0 other credentials.

• O*NET. Occupation data: O*NET 31.0 Database, USDOL/ETA, CC BY 4.0. We have modified and combined it. USDOL/ETA has not approved or endorsed these changes.
• Bureau of Labor Statistics. Employment Projections and Occupational Employment and Wage Statistics. BLS data is public domain. Wages are national OEWS, growth and openings are BLS projections.
• Licensing boards and exam owners. Time, fee, minimum-age and record-policy facts come from the board or body that issues the credential, with the date retrieved. A value we have not checked is shown as "not checked yet", never as zero.
• Items. Built on ideas from Holland's interest theory, O*NET, the Theory of Work Adjustment and Klimov's work classification. No item is copied from any commercial test.

Sources behind the profiles

9. Limitations, in plain words

• The weights are placeholders. No data has tested whether this mix is better than another.
• Fit compares your answers with an occupation profile, not with people who hold the license. Interest scores in the occupation data are estimated by a model with expert review.
• Interest scores relate to job performance with a correlation of about .24 in large studies, with a wide spread (Sackett et al., 2022). That is about job performance, not about fit with a license.
• The fewest reasoning items for any one ability is 0, which is too few to treat as a measurement of that ability. The ability part still carries 30% of the fit formula.
• The kind-of-work tags and work values failed the second-coder check.
• Bands are relative. A strong band means near the top of the list compared, not a high level of fit.
• Rules, fees and times change. Each fact carries the date it was read.
• English only, and United States licenses and certifications only. People aged 13 to 17 can take it; their results are not saved.

10. Version history and audit contact

Every result stores the three versions that produced it. A change to any of them adds a row here.

Audit requests, data questions and corrections: editorial@decipheru.com. Past fixes are in the corrections log. How Match works for a taker is on the method page.