We built the engine first, and the graphics second.

1. Reading: one order, no guessing

The sequence follows the ordering principles used by structured-literacy programmes: continuous sounds (m, s, f) before stops (t, p) so that blending is teachable from day one; visually and auditorily confusable pairs separated (b/d, m/n, e/i); one vowel introduced at a time; and every new letter-sound used immediately in real words rather than drilled in isolation.

The stages run: first consonants and short a → the remaining short vowels → consonant digraphs and the doubling rule → blends → magic e → open syllables → r-controlled vowels → vowel teams → diphthongs → soft c and g, -dge and -tch, silent letters → consonant-le and syllable division → prefixes, suffixes and -tion. That is 49 rungs, kindergarten through grade three, with heart words woven in alongside.

What we refuse to do. No picture cues on decoding items. No "what would make sense here?". Distractors are deliberately one grapheme away from the answer — hat / hit / hut / hot — so a child who guesses from the first letter gets it wrong and gets corrected. That is uncomfortable by design; guessing that is allowed to succeed is the single most expensive habit in early reading.

2. Blending is shown, not described

When a word is modelled, the app highlights each grapheme as it is spoken, pauses, and then says the whole word. Letter teams light up as one unit, so sh, igh and ai are never sounded out as separate letters. The silent e is marked as silent rather than pronounced. This is a small thing that a great many apps get wrong.

3. Mathematics: concrete, then picture, then symbol — and then fade

Every math idea appears first as something you can move, then as a drawing, then as notation, with the representations linked rather than swapped. Ten-frames, number bonds, arrays, bar models and number lines are reused for years instead of replaced with a new toy each lesson, because the sameness is what carries meaning forward: the array a seven-year-old counts is the same object as the area model a ten-year-old uses for 24 × 36.

Fractions are taught primarily on a number line, not as pizza slices. Pie-only instruction reliably produces the belief that a quarter is bigger than a third because four is bigger than three, that a third plus a third is two sixths, and that five thirds is not a real number. The number line dissolves all three.

Fact fluency is built the humane way. A fact only enters the timed pool once the child can already derive it — making ten, doubles and near-doubles, ×10 minus one group. There is no leaderboard, no visible clock, no red, and an untimed retry is always available with no penalty and no visual difference.

4. Mastery means three things, not one

A rung is climbed when the child is right in at least 9 of their last 10 attempts, and their median response time is under the target for that skill, and they have held both on two separate days. Accuracy alone is how a program can tell you a child "mastered" something they will have lost by Thursday.

The difficulty controller aims at roughly an 80% success rate — the setpoint that Rosenshine's classroom research, Vygotsky's zone of proximal development and flow theory all independently point at — with a little deliberate variance so it is not monotonous.

5. Review is scheduled, not random

Every skill and every individual word or fact sits in one of five boxes with review intervals of 1, 3, 7, 16 and 35 days. A correct and fast answer promotes it; any error sends it straight back to box one; correct-but-slow holds it in place. We cap the longest interval at about a month because a child's retention horizon is a school year, not a decade, and we re-seed at term boundaries. The child sees this as a simple, explainable thing: "that word moved up a box."

6. A wrong answer triggers a routine, not a buzzer

The correction sequence is the one direct-instruction teachers use: model (say what is true and why), lead (the child must select the correct answer before moving on), test (the same item returns later in the same session), and re-firm (it comes back in the next session regardless of the review schedule). We never re-serve the identical item immediately, and we never let a child click past an error.

7. Eleven named misconceptions, watched for by name

Wrong answers are not all equally informative. We keep a bank of specific, named bugs — each with diagnostic items whose distractors are that bug's output — so we can tell the difference between a slip and a broken idea:

• sub smaller from larger — Takes the smaller digit from the larger one whatever the position.
• equals means answer — Reads = as "write the answer here", so 3 + 4 = _ + 2 gets 7.
• count all — Recounts from one instead of counting on.
• whole number bias — Thinks 1/4 > 1/3 because 4 > 3.
• add denominators — Adds tops and bottoms: 1/3 + 1/3 = 2/6.
• decimal longer bigger — Thinks 0.125 > 0.5 because it has more digits.
• mult always bigger — Believes multiplying always increases.
• vowel guess — Guesses the vowel from the first and last consonant.
• digraph split — Sounds out sh as /s/ /h/.
• silent e said — Pronounces the final e.
• first letter guess — Says a word that starts with the right letter but is not the word (three-cueing behaviour).

Choose the bug's answer twice and the app does not simply mark it wrong again — it branches to the rung that repairs the underlying idea. Grown-ups see the flag in Field Notes in plain language.

8. Motivation that cannot be decoupled from learning

The progress artefacts are the curriculum. The sound wall fills in as you master spellings. The bookshelf unlocks books you can genuinely decode. The ladder is the skill graph. There is no separate currency to farm, because a currency that can be farmed will be farmed instead of learning.

Deliberately absent: leaderboards, leagues, class rankings, public accuracy, variable-ratio rewards, mystery chests, purchasable streak protection, guilt notifications, and any prompt to buy anything shown to a child. The weekly target is a rolling "four days this week", not a consecutive-day streak, so nothing can ever be lost.

9. Sessions are built to end

Ten to fifteen minutes for ages five to seven, fifteen to twenty-five for eight to eleven. There is a real completion screen, a suggestion that takes the child away from the screen, and a question for a grown-up to ask at dinner. Continuing requires deliberate effort; stopping is the default.

10. Accessibility is dials, not a "mode"

The reading panel exposes text size, line spacing, letter spacing, word spacing, page tint and letter shapes individually, because the evidence supports the spacing and size settings, not any particular font. OpenDyslexic is offered as a preference and explicitly not sold as a benefit — controlled trials have not found one. Everything is keyboard-operable, targets are at least 44 pixels, nothing is signalled by colour alone, and reduced-motion is honoured.