Standards

Methods and ethics

What we count, what a model does, and what we will not claim.

Unit of analysis

One expert tutor. 48 recurring students. 943 coded lessons of 1,418 recorded, August 2024 to June 2026, across biology, chemistry, physics and mathematics, GCSE through A-level and IB. Learners are predominantly in England with a minority in the UAE.

Findings describe one expert's practice at high resolution. They are not a description of expert tutoring in general.

How the evidence is made

Sources

Recorded one-to-one lessons delivered by Phoenix Tutors, executive interviews, and structured survey responses from schools, trusts and edtech teams in the UK and UAE. Educave Research is the research arm of Educave.

Counting

A deterministic pre-processor attributes speech and computes every word, turn and talk-share figure. A language model classifies discourse only; it never counts.

Quality gate

No-shows, aborted connections and planning calls filed as lessons are set aside before analysis.

Agreement

Two independent coders double-coded 10 lessons, about 1% of the coded corpus, so the agreement statistics carry wide intervals. Current agreement: question type 0.95, scaffolding 0.98, error type 0.80. Proportions and directions are reportable; absolute counts are not, and are labelled exploratory until calibration lifts item-level agreement.

Held-out evaluation

Benchmark decision-points are reserved before any policy is derived, so agreement scores are never fitted on their own data. The held-out set shares its tutor and its students with the derivation set, so it tests fidelity to this expert and not external validity.

Consent and ethics

Participation is covered by the service terms agreed by the fee-paying parent or guardian at the point of engagement. Learners and families know their lessons form part of the study. Learners in the tutoring corpus are minors. Anonymisation runs before any model step, an automated scrub verifies that no real names survive into coded output, and only aggregate, paraphrased results leave the repository. Raw transcripts are never shared, published or used for training. Survey and interview participants are anonymised at organisation level, and quotes are cleared before publication.

Data availability

Coding scheme, codebook, agreement statistics and aggregate data tables are available on request. The 308-point benchmark set is available to public bodies and researchers on request. Raw transcripts are not shareable under the consent terms.

What we do not claim

No causal claim about attainment. The corpus describes practice; it does not measure outcomes.
No population inference. Samples are purposive, so findings describe the corpus and the organisations in it.
No demographic gap analysis until disadvantage indicators are linked at source, which they currently are not.

Corrections and versioning

Every published piece carries a version number and a date. Substantive changes increment the version and are noted on the page. Corrections can be sent to hello@educave.ai or through peer review.

Status: exploratory, not peer reviewed, no external ethics board review. Open to challenge through peer review. Methods v1.0, August 2026. Research contact: Carl Morris, hello@educave.ai.

Read the research