Reach120 Intelligence
Four modelled components, the methods behind each one named, and the limits of each one published. Everything below describes code that runs in the product today.
The smartest way to get prepared for TOEFL 2026. That is a strong sentence, so here is what is underneath it — the method, and what was actually fitted, for each of the four.
Each error type is a latent skill with its own learn, slip and guess rates. Every scored response is an observation, and your mastery estimate for that error type moves accordingly — so a mistake you make once and never repeat separates itself from the one that keeps costing you points.
What is fitted:
16
error types modelled
984
sequences in the fit
An error type is fitted on its own only once enough sequences exist to estimate it; the rest are pooled into one fallback skill rather than fitted on too little data.
The features are your last three scores, an exponentially weighted average, how many sessions you have done, how long since the last one, which task family you are about to attempt, and your recent error rate. It estimates the band your next session is heading for — not a test-day result.
What is fitted:
1,125
sessions
98
learners
5-fold
grouped CV
Cross-validation is grouped by user, so no learner appears in both train and test.
Each task family carries a posterior over how much score movement it tends to produce. The router draws from those posteriors and adds a bonus for the skills your own mastery estimates say are weakest, so two learners on the same band can be sent to different next tasks.
What is fitted:
3
Writing task arms
Each arm accumulates settled outcomes continuously, so no count of them is published as a fixed figure. Reading and Speaking are routed by deterministic per-learner rules instead — see the limits below.
The 0-5 rubric estimate on an open response is a judgment an AI model makes. Everything after it is not: mastery values, readiness figures and the next-task choice are all computed in code. The language model is handed those finished facts and asked only to write them up, and its text is checked against them before you ever see it.
What is fitted:
A narration containing a number, a pattern label or a route that is not in the computed facts is discarded rather than shown, and the writer returns nothing at all if the check fails.
Everything we know to be a limitation is listed here rather than left for you to find.
A real gap over both baselines — and still about half a band of typical error. Good enough to point your study at a section; not good enough to plan a test date around.
Cross-validated mean absolute error is 0.40 on the 0-5 rubric scale, against 0.75 for "assume your previous score" and 0.64 for "assume your average". That is a real gap, and it is still about half a band of typical error. It is good enough to point your study at a section. It is not good enough to plan a test date around.
Some of the model's advantage over "assume your previous score" comes from knowing which task family you are about to attempt, because task families score differently. That is legitimate for "what is your next Academic Discussion heading for", but it means the model is partly a calibrated per-task prior rather than a purely personal one.
The word-count feature contributes nothing: word counts are missing on almost every historical rubric session, so the model effectively ignores it. It is listed here rather than removed because the backfill is the fix, and until then the feature list overstates what is being used.
Punctuation, plural forms, word choice and capitalization have too few and too short sequences to pin down. Read their fitted transition rates as "this error rarely comes back" rather than as a calibrated speed of learning. The guess-rate ceiling also binds on most skills, which flattens differences between them.
Thompson sampling runs in Writing and Listening. Reading and Speaking are routed by deterministic per-learner rules over your own weakest skills — still personal, but not adaptive in the same sense. Writing carries much the most settled evidence of the two adaptive sections, so that is where the routing is best supported.
The posteriors are population-level: one shared set of arm estimates, not a model fitted to you. Your own contribution is the mastery bonus and the draw, which is what makes two learners diverge. Nobody should read this as "a model trained on your data alone".
Open responses are judged by an AI model against criteria we wrote from publicly available task descriptions. They are not ETS rubrics, the model is not trained on officially scored responses, and a judgment can be wrong. Reading and Listening are the exception: those are graded against the answer key.
Both fitted artefacts date from July 10, 2026 and are refitted as more practice accumulates, so these figures move. They are published as of the last refit rather than as a standing property of the product.
Any tool can tell you what it does when it has an answer. These are the places Reach120 has no answer, and what it shows you instead of one.
Your mastery estimates are built from your own scored responses, so before you have written any there is nothing to build them from. A brand-new account shows an empty list of repeated mistakes rather than the mistakes a typical learner makes. That is the honest state, and it is the state the cold-start audit checks for.
Each modelled error type gets its own learn, slip and guess rates only once enough sequences exist to estimate them. The ones below that bar are pooled into a single shared fallback instead of being handed parameters fitted on almost nothing, which would look identical on screen and mean nothing.
The filter behind the band forecast needs a minimum number of scored sessions before it will estimate a level and a direction at all. Below that it returns nothing, and the section reports insufficient history instead of drawing a trajectory through one or two points. Typical error is about half a band even when it does fit, so it is a study signal and never a predicted TOEFL score.
This is the part that is not a matter of care. Insufficient history is a state in the API contract, and the contract rejects any response in that state that carries a band, a score, a readiness label or an activity date. A bug that tried to fill the gap with a plausible number would fail validation rather than reach you — the refusal is enforced by the schema, not by remembering to be careful.
Adaptive routing that updates from observed outcomes runs in Writing and Listening. Reading and Speaking are routed by deterministic rules over your own weakest skills — still chosen from your evidence, but not adaptive in the same sense. Every recommendation carries the kind of decision that produced it, so the two are distinguishable in the response itself rather than blended into one confident-sounding suggestion.
The arm estimates behind the router are population-level — one shared set, not a model fitted to you alone. What is yours is the weighting toward your own weakest skills and the draw itself, which is what makes two learners on the same band diverge. Nobody should read the routing as a model trained on their data alone, and we do not describe it that way.
When a language model is used to put your figures into sentences, it is handed the finished values and allowed to narrate them and nothing else. Its output is checked back against those values first: a sentence containing a number, a pattern, a route or a claim about your own practice history that the engines did not produce fails the check, and the write-up is discarded rather than corrected or shown. When that happens you get no prose, which is the intended outcome.
None of this makes the product more accurate. It makes it legible: when a figure is missing here, the reason is that the evidence for it is missing, and that is a different thing from the product being broken.
The same rule, applied to every figure we publishAll four TOEFL 2026 sections in one account, writing to one learner record. Each card shows the section itself, names the task families in the bank, and states where the result comes from.
Task families
What you get back
Graded against the answer key. No model judges a Reading item.
Open Reading
Task families
What you get back
Graded against the answer key, per question.
Open Listening
Task families
What you get back
Transcribed automatically, then judged by an AI model against our criteria.
Open Speaking
Task families
What you get back
Build a Sentence is matched against the accepted arrangement and is not judged. Email and Academic Discussion are judged by an AI model against our criteria.
Open WritingEvery score-shaped figure in Reach120 carries the same label: Reach120 practice band estimate — not an official ETS score. This is what sits behind that sentence.
Length targets are ours, not ETS rules. Any range Reach120 shows is an optional pacing target for practice, ETS publishes no word-count requirement for Write an Email, and no response is ever marked down for going past a Reach120 target.
Read how scoring works in fullReading, then Listening, then Speaking, then Writing — back to back, on the clock, in one session that you cannot cherry-pick your way through. The listed blueprint runs 1 h 23 min – 1 h 29 min.
Listed blueprint times, shown as estimates — ETS states the 2026 specification is subject to minor revision until launch.
You cannot skip ahead or return to a finished section. The run is what produces a comparable score across all four rather than four good days stitched together.
Each section returns its own result on the same practice-estimate scale, and the run writes to the same learner record as your ordinary practice — so the model sees you under test conditions too.
The run is stored server-side as it goes, so a closed tab or a dead battery does not cost you the whole mock.
Missed items and generated vocabulary cards are aged by the same spaced-repetition scheduler, so a word you half-know and an error you keep repeating come back at the right moment instead of whenever you happen to remember them.
Mistake review and vocabulary review share a single FSRS implementation targeting 90% recall, so a card and a past error age at the same rate and can be counted in one due number.
Generate a deck for any topic — a field you keep reading about, a task type you keep losing points on — with each entry carrying its meaning, its use and a worked example.
Generation is what is capped, because generation calls a model. Once a deck exists it is yours to review for as long as you like, on any account. Free accounts can generate 1 deck.
Readiness per section, the trend under it, the skills that are moving and the ones that are not — presented as an estimate with its own uncertainty rather than as a number that flatters you.
An unpractised section is shown as unmeasured, never counted as a zero. A readiness figure built by averaging in silence would be the easiest number on the page to fake.
Practice estimates are labelled as estimates and carry a link to how they were produced. Failed reads render as named unavailability rather than as a confident zero.
Typical error is about half a band, so it is a study signal — not a predicted TOEFL score. With too little history it returns nothing at all rather than an invented figure.
Reach120 does not replace you — it does your marking. The workspace exists so class time goes to teaching instead of correcting grammar between lessons.
A private invite link, a short class code they type themselves, or a QR code you put on the screen in front of them. No account provisioning, no spreadsheet of names to upload.
Publish a task to the whole class, watch completion move as work comes in, and see what is still outstanding without asking anyone.
Open the full feedback report behind any submission — what the student actually wrote or said, the per-dimension breakdown, and the fixes.
Put the class code, a task, or a piece of work on a projector at a size a room can read, without exposing anything a student has not agreed to share.
A student sees the exact access list before joining and can leave whenever they want. It is not a policy sitting on top of a broader permission — it is what the endpoints will return.
Students buy their own pass and own their own practice record. Nothing about a class transfers ownership of a learner’s work to the person teaching them.
Every scored writing and speaking attempt can be downloaded as a typeset report — a purpose-designed document rather than a screenshot of the app — which is the artefact a learner takes to a teacher, a parent or an application file.
The band estimate with its label, the per-dimension breakdown, the specific fixes, and your own response — so the reader can check the judgment instead of taking it.
The document is rendered in your browser from the feedback you already have. Your response is not sent anywhere new to produce it.
Speaking attempts get their own report shape, and both carry the same estimate labelling as the app, so nothing becomes an official-looking score once it leaves the screen.
A real report, not a mockup.
A model that remembers your mistakes is a model that stores them. Here is exactly what that means, including the parts that are inconvenient to say.
Your responses, the scores and feedback they produced, the error types found in them, and the review cards scheduled from those errors. That record is what makes the model personal — there is no version of it that both remembers your mistakes and keeps none of them.
Cancelling leaves your access running to the end of the period you paid for. Deleting the account is self-service and removes your practice record with it — you do not have to email anyone to leave.
Joining a class shows you the access list before you accept, not after. You can leave a class whenever you want, and your work stops being visible when you do.
These are not policy promises layered over a broader permission — they are the boundary of what the class endpoints will return at all.
Reach120 is independent. Nothing here is published, licensed, endorsed or reviewed by ETS, and no question in the bank is a real test question. TOEFL is a registered trademark of its owner, used here only to say which test we prepare you for.
Every number Reach120 shows is a practice estimate produced by our own methods. It is a way to track your own movement over time. It does not predict, convert to, or stand in for an official result.
No tool can promise you a particular band, and we will not pretend otherwise. What practice produces depends on the practice you do — anyone selling you the other version of that sentence is selling you something they cannot deliver.
Reach120 does the marking. It does not do the teaching, it cannot see why you are stuck, and it cannot decide what you should do about your life. If you work with a teacher, this is the thing that fills the days between lessons.
There is no lecture series, no grammar syllabus from scratch, and no beginner track. Reach120 assumes you can already write and speak English and are trying to score better on a specific test.
Open responses are judged by a model against criteria we wrote. It is fast, consistent and available at midnight, and it is not a person reading your work. Where the result matters, read the feedback rather than the number.
Reach120 practice estimates are study aids for tracking your own progress. They are not official ETS scores and do not predict one.
Reading and Listening are free on any account, so you can check the scoring against your own judgment before paying for anything.
Reach120 is an independent practice tool. It is not affiliated with, endorsed by, or approved by ETS, and it does not provide official TOEFL® test scores.
TOEFL® is a registered trademark of ETS. This product is not endorsed or approved by ETS.
