Transparency

Assessment Methodology

How b4skills measures English proficiency — the psychometrics, AI, and fairness safeguards behind every score.

Computerised Adaptive Testing (CAT)

The adaptive engine selects questions using the current ability estimate (θ), item information, skill coverage and exposure constraints. Objective items use IRT parameters for discrimination, difficulty and guessing. EAP estimation updates ability as scored evidence becomes available.

Testing stops according to each product’s precision targets, section coverage and item limits. A test can reach its item limit without achieving its target SEM. Reported precision must be read alongside the result.

CEFR Alignment

b4skills reports levels aligned to the CEFR using configured θ thresholds and CEFR-based task descriptors. Alignment is an intended interpretation of the assessment, not certification or endorsement by the Council of Europe.

Active item status does not by itself establish empirical calibration. Parameters and level thresholds may be provisional. Independent standard-setting and external validity evidence must be demonstrated separately; current evidence is shown on the Research page.

AI Scoring & Human Review

Full assessments support AI-assisted scoring of writing and speaking against CEFR-based rubrics. Assessment scoring uses OpenAI. Speaking audio may be transcribed before evaluation.

Responses requiring review are held for human assessment rather than treated as final scored evidence. We do not claim a published human–AI agreement coefficient. The free Quick Check does not evaluate open writing or speaking responses; its level is a provisional estimate from objectively scored responses.

Fairness & Bias Monitoring

The platform includes tools for item analysis, exposure control, person-fit checks and Mantel-Haenszel DIF screening. These tools can support item review when sufficient response and demographic data are available.

Their presence does not establish that a study has been completed or that an assessment is free of bias. Monitoring signals require contextual review and should not be treated as proof of misconduct.

Question Types & Constructs

Reading and listening tasks use selected-response or structured formats to assess comprehension. Grammar and vocabulary tasks assess language use. Writing and speaking tasks collect constructed responses where supported by the assessment configuration.

Question counts, task formats and scored skill coverage vary by product. The free Quick Check includes six skill areas, but open writing and speaking responses are not evaluated or included in its provisional level.

AI Scoring Rubric Dimensions

AI-assisted writing and speaking scoring uses CEFR-based rubric dimensions such as task achievement, grammatical range and accuracy, lexical resource, coherence, and speaking fluency where supported.

Dimensions, weights, providers and review thresholds depend on the scoring configuration. Model agreement is not a substitute for independent validation against human ratings; no numerical agreement claim is made here.

External Score Concordance

No independently validated conversion between b4skills scores and IELTS, TOEFL, Cambridge or Pearson scores is claimed. Sharing a CEFR label does not make scores interchangeable or establish equivalent admission requirements.

Any external score references are contextual only. The receiving institution decides which assessments and evidence it accepts. Current paired external-test observations, including an insufficient or empty sample, are disclosed on the Research page.

Score Validity & Certificates

b4skills issues its own assessment reports and certificates. Certificate status can be checked using the verification page; verification establishes the status of a b4skills record, not accreditation or acceptance by another institution.

Reliability and validity are separate questions. The Research page distinguishes descriptive IRT marginal reliability from Cronbach alpha and does not infer reliability from item discrimination.

Independent validation remains in progress. Decisions with substantial consequences require additional evidence, appropriate human review and an institution’s own standard-setting process.

Frequently Asked Questions