Every child now does the whole EGRA and Core EGMA battery: 18 tasks, each given the way the toolkits give it, with the same items, practice and 60-second timing. Tap through child 1 as a coach does. The phone shows the bot's own messages. Beside it, the table shows the booklet page in front of the child (click it to see it full size), the exact words to say, and the items with their answers from the coach card. On a wide screen a third column explains why each task works the way it does. The strip above the phone jumps to any task.
Reading follows the May 2026 Rawalpindi instrument, our baseline of 2,233 children, so scores stay comparable with it. Urdu made-up words are added where English has them. Maths is Core EGMA. Nothing is newly written. Every item comes from the May study, where its list survives in full, or else from an official EGRA/EGMA source, cited per item in the bank. Reading is one form for Grades 3 and 5, as in the May study; Grade 5 maths gets harder pairs and sequences from item 7.
| Task | The child | Items | Time | Stop rule | Items from | Scoring |
|---|---|---|---|---|---|---|
| Urdu · Listening | Hears a short story read twice, answers 6 questions | 6 | untimed | — | Rawalpindi May 2026 study | AI + coach review |
| Urdu · Letters | Names the letters on the page | 100 | 60 s | nothing right in the first 10 → stop | Rawalpindi May 2026 study | AI, provisional |
| Urdu · Made-up words | Reads made-up words | 50 | 60 s | nothing right in the first 5 → stop | Sindh EGRA 2024 (Urdu) | AI, provisional |
| Urdu · Familiar words | Reads familiar words | 50 | 60 s | nothing right in the first 5 → stop | Rawalpindi May 2026 study | AI |
| Urdu · Story reading | Reads the story, then answers the questions reached | 60 words + 6 Qs | 60 s | nothing right in line 1 → stop | Rawalpindi May 2026 study | AI + coach review |
| English · Listening | Hears a short story read twice, answers 6 questions | 6 | untimed | — | Rawalpindi May 2026 study | AI, provisional |
| English · Letters | Names the letters on the page | 100 | 60 s | nothing right in the first 10 → stop | Rawalpindi May 2026 study | AI + coach review |
| English · Made-up words | Reads made-up words | 50 | 60 s | nothing right in the first 5 → stop | EGRA Toolkit 2009 (sample grid) | AI, provisional |
| English · Familiar words | Reads familiar words | 50 | 60 s | nothing right in the first 5 → stop | EGRA Toolkit 2009 (sample grid) | AI + coach review |
| English · Story reading | Reads the story, then answers the questions reached | 60 words + 6 Qs | 60 s | nothing right in line 1 → stop | Rawalpindi May 2026 study | AI + coach review |
| Maths · Number identification | Says each number | 20 | 60 s | — | Core EGMA | AI + coach review |
| Maths · Which is bigger | Says the bigger number of each pair (G5 harder from item 7) | 10 | untimed | 4 wrong in a row → stop | Core EGMA · G5 items 7–10: RTI Rwanda P6 items | AI + coach review |
| Maths · Missing number | Says the missing number (G5 harder from item 7) | 10 | untimed | 4 wrong in a row → stop | Core EGMA · G5 items 7–10: RTI Rwanda P6 items | AI + coach review |
| Maths · Addition, quick | Quick additions | 20 | 60 s | — | Core EGMA | AI |
| Maths · Subtraction, quick | Quick subtractions | 20 | 60 s | — | Core EGMA | AI |
| Maths · Addition, harder | Harder additions (paper allowed) | 5 | untimed | 4 wrong in a row → stop | Core EGMA | AI + coach review |
| Maths · Subtraction, harder | Harder subtractions (paper allowed) | 5 | untimed | 4 wrong in a row → stop | Core EGMA | AI + coach review |
| Maths · Word problems | Answers problems the coach reads aloud | 6 | untimed | 4 wrong in a row → stop | Rawalpindi May 2026 study | AI + coach review |
Scoring, from the AI test on 185 May recordings: AI scores alone; AI + coach review means the coach confirms only what the AI could not settle; AI, provisional means the recording is kept at full fidelity but the AI is not yet good enough, so the score is marked ≈ until the pilot recordings are checked.
The real bot on the NIETE sandbox, with v3 switched on, played by synthetic children: scripted strong, middling and weak Grade 3 voices (computer-generated, with classroom noise and a phone band; no real child). Every voice note went through WhatsApp's webhook, was stored, scored by the AI, and the end-of-visit review was answered from each child's script. 5 children, 90 task rows, 90 scored, 0 failed, 90 checked; 89 voice notes; the review had 29 items on 2 pages. Synthetic notes arrive faster than real speech, so the 16.7 minutes it took say nothing about a real visit.
| Task | strong | middling | weak | strong | middling |
|---|---|---|---|---|---|
| Urdu listening | 5 / 5 | 4 / 4 | 2 / 2 | 5 / 5 | 4 / 4 |
| Urdu letters | 63 / 62 | 38 / 30 | 10 / 7 | 64 / 62 | 37 / 30 |
| Urdu made-up words | 28 / 27 | 13 / 11 | 5 / 2 | 27 / 27 | 14 / 11 |
| Urdu words | 44 / 44 | 25 / 21 | 4 / 0 | 44 / 44 | 25 / 21 |
| Urdu story | 49 / 50 | 30 / 28 | 12 / 5 | 50 / 50 | 30 / 28 |
| English listening | 5 / 5 | 3 / 3 | 0 / 0 | 5 / 5 | 3 / 3 |
| English letters | 50 / 58 | 34 / 29 | 8 / 7 | 51 / 58 | 37 / 29 |
| English made-up words | 24 / 29 | 8 / 13 | 2 / 2 | 22 / 29 | 9 / 13 |
| English words | 41 / 41 | 19 / 21 | 8 / 2 | 41 / 41 | 25 / 21 |
| English story | 55 / 54 | 23 / 23 | 4 / 2 | 55 / 54 | 23 / 23 |
| Number identification | 20 / 20 | 16 / 15 | 3 / 3 | 20 / 20 | 16 / 15 |
| Which is bigger | 9 / 9 | 6 / 6 | 2 / 2 | 8 / 9 | 6 / 6 |
| Missing number | 10 / 9 | 6 / 6 | 5 / 3 | 10 / 9 | 6 / 6 |
| Quick + | 13 / 14 | 7 / 6 | 1 / 0 | 14 / 14 | 6 / 6 |
| Quick − | 5 / 15 | 8 / 7 | 5 / 5 | 5 / 15 | 7 / 7 |
| Harder + | 5 / 5 | 4 / 4 | skip / skip | 5 / 5 | 4 / 4 |
| Harder − | 4 / 4 | 3 / 3 | 0 / 0 | 4 / 4 | 3 / 3 |
| Word problems | 6 / 6 | 4 / 4 | 3 / 3 | 6 / 6 | 4 / 4 |
Each cell: the score saved after the coach's review / what the synthetic child was scripted to get. Over 89 scores, 46 match exactly and 78 fall within tolerance (±5 for reading counts, ±1 otherwise). Synthetic voices are not children: this proves the whole pipeline works and finds defects; accuracy on real children is measured on the May recordings (lanes R8, L37) and must be re-measured on the pilot.
Outside tolerance (11 of 89): Urdu letters, middling 38 against 30; Urdu letters, middling 37 against 30; Urdu story, weak 12 against 5; English letters, strong 50 against 58; English letters, strong 51 against 58; English letters, middling 37 against 29; English made-up words, strong 22 against 29; English words, weak 8 against 2; Missing number, weak 5 against 3; Quick −, strong 5 against 15; Quick −, strong 5 against 15.
The full battery takes about 24 minutes per child, so about two hours for five children. Two separate analyses of the May data (lanes O1 and O2) looked for every way down to 30 minutes that keeps the items, timing and stop rules intact. They reached the same answer independently.
● comparable with May and EGRA/EGMA · ● comparable with a stated error (≤0.1 SD bias, or a linked scale) · ● not comparable
| Option | Min / child | Min / 5-child visit | School information kept | Comparable |
|---|---|---|---|---|
| A0 Full v3 battery (18 tasks) | 24.0 | 120.0 | 100% | yes |
| A1 45-s windows on the 11 timed tasks | 22.5 | 112.6 | 100% | with stated error |
| A2 30-s windows on Urdu words + story only | 23.5 | 117.6 | 100% | with stated error |
| A3 30-s windows on all 11 timed tasks | 20.6 | 103.0 | 99% | no |
| A4 Routing: story skipped if words line 1 all wrong; English nonwords skipped if words <5 cwpm | 23.6 | 118.0 | 99% | with stated error |
| A5 Short comprehension forms (IRT): 5 Urdu / 4-5 English listening Qs | 23.6 | 117.9 | 99% | with stated error |
| A6 Skip letters if familiar words >=40 cwpm (imputed) | 23.0 | 114.9 | 95% | no |
| B1 = A1+A2+A4+A5 (all per-child levers that keep comparability) | 21.4 | 107.1 | 99% | with stated error |
| M1 Matrix: each child 2 of 3 blocks (UE/EM/UM), linked | 14.1 | 70.7 | 70% | with stated error |
| M2 Matrix: one block per child + anchors (Urdu story, L1 add/sub), linked | 9.3 | 46.5 | 55% | with stated error |
| M3 Matrix: one block per child, no anchors, linked | 6.8 | 33.8 | 45% | with stated error |
| M4 English on 2 of 5 children, Urdu+maths on all, linked | 17.5 | 87.7 | 84% | with stated error |
| M5 Matrix: one half-block per child (Urdu, English, maths-A, maths-B, Urdu/English alternating), linked | 5.5 | 27.3 | 40% | with stated error |
| R-U One subject per visit, all 5 children: Urdu (rotating) | 5.9 | 29.6 | 28% | yes |
| R-E One subject per visit, all 5 children: English (rotating) | 6.5 | 32.3 | 28% | yes |
| R-M One subject per visit, all 5 children: maths (rotating) | 9.0 | 45.2 | 44% | yes |
| R-M2 Maths split over 2 visits, all 5 children (half of maths) | 4.5 | 22.6 | 22% | yes |
Rows on a green ground fit in 30 minutes. "School information kept" is the share of the full battery's precision on the school mean, averaged over the 18 tasks. "With stated error" means a bias of 1/10 SD or less, or a score on a linked scale.
One A4 PDF, one copy per coach, printed double-sided and laminated: four child booklets (one task per page, as EGRA stimulus books are) and four coach cards. Nothing on a child page shows a number, an answer or a question. Click any page to see it full size.
Grades 3 and 5 · one task per page; a practice row only on the made-up words page
Grades 3 and 5
Core EGMA, green
Core EGMA with harder items 7–10, blue
every line to say, the listening story, questions and answers
the same in English
every answer, the word problems to read aloud
every answer, the word problems to read aloud
Yes, all three, for every child. Listening: the coach reads a short story aloud and asks 6 questions (the May study's own, read twice as May did). Letters: 100 letters, 60 seconds, the child names each one; nothing right in the first row stops the task. Made-up words: 50, 60 seconds, the stop rule on the first 5. They test decoding, because a made-up word can't be guessed from its shape. Urdu uses letter names (as Pakistan's Urdu EGRA and the May study do).
Six skills in eight tasks, 96 items, three of them timed. Number identification (20, 60 s), which is bigger (10), missing number (10), quick additions and quick subtractions (20 each, 60 s each, separately), harder additions and subtractions (5 each, paper allowed), word problems (6, read aloud). v3 gives all eight, with Core EGMA's own items for Grade 3.
No, EGMA does not allow it. It is oral and one-to-one: "no student writing is required. Answers to all items are orally transmitted." Paper is only an optional aid while a child answers aloud. Children waiting must not hear the testing. RTI's written maths tests are separate instruments whose scores are never pooled with EGMA's.
Where EGMA times it: number identification, quick additions and quick subtractions, 60 seconds each from the begin line. The rest are untimed and stop after 4 wrong in a row.
Ten pairs now. Grade 3 gets Core EGMA's ten. Grade 5 shares the first six (so the grades stay comparable) and gets four harder ones from RTI's official Rwanda Grade 6 items. On the old four pairs, 62% of Grade 5 got everything right; nothing could be learned from them.
Done. At "stop" the coach notes where the child's finger is. The message lists each question with the words the child must have read for it, and the score is out of the questions asked.
About 24 minutes per child, two hours for five. The path to 30 minutes is above: every task on every visit, on different children, at 27 minutes.
Nothing is written new. The May 2026 study where its list survives in full; otherwise RTI's EGRA Toolkit sample grids, the Government of Sindh 2024 Urdu EGRA, Core EGMA, and RTI's Rwanda Grade 6 items. Each item cites its page in the bank, and the grids were checked cell by cell against the source pages.
No roll numbers. Full name, the class the way the school writes it, and the class teacher; one button sends each teacher only their own room's children.
No. The recording is the score sheet; the AI counts. The coach marks only the few items the recording could not settle, at the end of the visit.
No. EGRA gives every child in a round the same form. One set per term (Set A now); Set B next term, after the pilot equates it.