Skip to main content
Research programme

Evidence from inside the delivery.

Most evidence on how education reaches displaced and under-served learners is generated by institutions that look nothing like the programmes being studied. TEEI's research works the other direction: it examines delivery models from inside the delivery itself, and publishes what it finds.

Seven areas of enquiry

Education delivery and pathway effectiveness

How structured pathways move people from first contact through language, skills and work. What makes a handover between programmes hold rather than drop. Why some designs retain learners and others lose them at predictable points. TEEI studies its own delivery model as the evidence base for how the programmes are built.

Displacement education and refugee integration

The compounded barriers displaced people face when rebuilding a life: language, credential recognition, informal employment traps, isolation. Why gender-neutral programme design fails women in these populations, and what a women-first design changes. This area frames the founding question of the organisation as a research question.

Health and access

Why health information asymmetry is a structural problem, and why a health barrier is often an education barrier in disguise. How people without medical training navigate systems that assume they can. What patient rights, plain-language health information, digital health tools and caregiver support look like when treated as questions of access rather than clinical questions. TEEI explores this as a general territory of information access, not as clinical or medical research.

Technology, AI and access

How AI can act as external working memory and lower the barrier to complex work for people without technical training. Where AI methodology helps and where fairness, bias and accessibility problems appear. The frame is methodology and access, not any single system.

Volunteer-delivered infrastructure at scale

What it takes to run programmes primarily through volunteers across many countries and languages. How coordination, matching and quality hold together when the delivery layer is distributed rather than staffed. Why this model reaches institutional scale on a small operating core, and where it strains.

Impact measurement methodology

How social impact is measured honestly rather than performed. SROI, CSRD and ESRS as frameworks, and where their social metrics have real gaps. Why impact data drawn from actual programme delivery is stronger than survey-based reporting, and what a defensible evidence chain from activity to outcome requires.

Digital skills and workforce outcomes

What digital literacy and skills training actually change in employment terms, and for whom. Where the gap sits between a completed course and a job. How workforce-readiness design differs for displaced people, women re-entering work, and underserved populations generally.

Published records

What TEEI has measured so far.

Two records from inside the delivery: a pilot of the L2SAFE benchmark, completed in July 2026, and the test of the AI writing-level estimate that Practice Companion uses, with 22 models compared on 16 October 2026.

L2SAFE · pilot v0.1 · July 2026

The learner language gap

L2SAFE measures whether comprehension, safety, retrieval, reranking and tool-use behaviour change when the same underlying intent is expressed in CEFR-aligned learner language. 24 paired items, 8 of them high risk, one model family, temperature 0.

Read the pilot record
Same items, same facts, same model. Only the register changes.share of items passing
Comprehension24 paired itemsSafety8 high-risk pairs0%50%100%Fluent 24/24Learner 24/24Fluent 8/8Learner 7/8 · item 210.0 pp12.5 pp 100%90%80%70%60%50%40%24/24 · 8/824/247/8item 21Fluent variantLearner variant
Pilot point estimates. One fluent-pass / learner-fail discordance of eight, none in reverse; exact McNemar two-sided p = 1.00. The pilot demonstrates instrument sensitivity, not a population effect.
AI writing-level estimate · 16 October 2026

22 models, the same 60 conversations

The experiment reused the same 60 constructed conversations, with three assessment repeats and three model readings per assessment. Exact agreement with the reference band is shown as a numerator and denominator rather than a rounded percentage. Read the full record

Sort by Show
Development, 72 eligible rows at baselineRegression, 39Reused September holdout, 42Thin line: spread between the three
Model
50%60%70%80%90%100%
Passed
Cost / attempt
Median time

Claude Fable 5.1 ran out of credit before the run completed and is not in the table. On the matched subset of 118 assessments it reached 47/50, 21/26 and 28/28 against Claude Sonnet 4.6's 46/50, 25/26 and 28/28 on the same subset. Cost is the estimated model usage per attempt; GLM-5 has no price in the pricing record. Median time is the assessment time; the last column notes the share of attempts above the 28-second ceiling the app waited for at the time.

L1, the released configuration, 24 October 2026

97 conversations · 485 outcomes · 425 eligible band results · exact agreement 390/425, or 91.8%
Development
128/140 · 91.4% · passed
Regression
65/65 · 100% · passed
Regression, second set
70/70 · 100% · passed
Regression, third set
59/75 · 78.7%Failed: five challenge-case overestimates
Fourth holdout
68/75 · 90.7%Failed: two challenge-case overestimates
All
390/425 · 91.8% · failed

The run did not produce an approval record for L1. L1 opened in production on 25 October under an explicit evaluation-gate exception.

Every one of the 485 outcomes

Outcome categories in the L1 evaluation
390 exact match with the primary reference band 10 a different band explicitly accepted by the case 18 overestimate outside the accepted alternatives 7 underestimate outside the accepted alternatives 50 expected insufficient outcome, returned correctly 10 accepted content-guardrail refusal

From the first design to the delivery fix

The development timeline, as recorded
4 August 2026First design for a conversation supported by the learner's own sentences
18 SeptemberInitial scorer and a small twelve-case evaluation
24 SeptemberConversation selected as the instrument; rubric-w2, strict response schema, writing floor and comparison-key work
25 SeptemberMedian rule fails; lowest-of-three configuration passes the internal evaluation on 60 cases and 300 assessments
28 SeptemberStaging readiness work for a controlled pilot
15 OctoberSix-question level conversation opens on staging
16 OctoberModel and panel experiment, with results from 22 models including the incomplete Fable run
20 Octoberltp-2 and ltp-3: writing-floor feedback, extra questions and the help-wording experiment
23 Octoberltp-4 offers eight question turns; mobile flow changes; output-limit diagnosis; notes and shortened-quote configurations fail evaluation
24 OctoberLarger output allowance and background job; language and boundary probes; L1 and fourth holdout; full L1 evaluation fails one criterion
25 OctoberL1 opens in production under an explicit evaluation-gate exception; synthetic production walkthrough completes
26 OctoberCallback-delivery fix reaches production
Practice Companion · what the benchmark gates

Four rooms that must pass first.

Practice Companion is not open to learners yet. Before the rooms open, they have to pass L2SAFE, the benchmark TEEI Research built to test whether a model understands and answers a question as safely in learner language as in fluent language.

Practice Companion · WriteSample
Practising for The doctor's appointmentLevel B2
Your companion AI

We have Tuesday at ten thirty.

You

No, I working morning. Can after three?

workingwork

The verb wants its ending; the rest already does its job.

Write it againKeep this line
Get in touch

For research collaboration, methodology enquiries, or access to underlying programme data, write to research@theeducationalequalityinstitute.org.