Useful before beautiful
Everything on this page comes from published, randomised, effect-sized research — including the studies that argue against the easy version of AI education. AnTeraju's design rules are derived finding by finding, and the Malaysian numbers say why it matters here specifically.
What the research actually says
Six findings, cited honestly — the two that support us, the three that constrain us, and the one that makes us refuse our own easiest feature.
Design beats model
In a randomised crossover trial, a research-built AI tutor more than doubled learning versus an excellent active-learning class — in less time (49 vs 75 minutes median). The active ingredient was pedagogy, not the model: reveal one step at a time, never the full solution, make the student attempt first, keep answers short, teach only from pre-verified solutions.
Kestin et al., Scientific Reports · Harvard RCT 2025
Unguarded AI backfires
GPT-4 help raised performance while students worked with it — and lowered their later exam scores by 0.28 to 0.41 standard deviations. Ease during practice is not learning. This is the single most important result in AI education, and it is a warning label, not a feature list.
Bastani et al. · field experiment, ~970 students
The boring techniques compound
Retrieval practice g=0.37, interleaving g=0.40, spaced practice g=0.20 (distributed over massed: d=0.54). None of these doubles anyone's IQ; all of them work in real classrooms, measured in honest tenths. Education software should be judged in these units — the units the ledger can actually report.
real-classroom meta-analyses, 2025–26
Struggle first, then teach
Solving a problem before being taught it beats direct instruction on conceptual understanding and transfer — provided instruction follows the struggle. Failure is data, not shame. A tutor that rescues too early is not kind; it is deleting the learning event.
Sinha & Kapur meta-analysis · productive failure
“When will I use this?” is an intervention
Connecting course content to students' own lives measurably raises achievement and interest — with the largest effects for exactly the students with the lowest expectations. Relevance is not engagement theatre; it is a lever, and it pulls hardest where the need is greatest.
Hulleman & Harackiewicz · utility-value field experiments
Cognitive debt is real
EEG studies show LLM-assisted work produces weaker neural connectivity and poorer recall of what was written — and measurable deskilling appears within minutes of unassisted problem-solving being replaced. The question is never whether to use AI in learning, but which hard parts the human must keep.
MIT Media Lab · “Your Brain on ChatGPT” 2025 + follow-ups
So AnTeraju is built this way
Each rule names the study it obeys. These aren't aspirations — they are refusals of the easy design, made concrete.
The attempt-first reveal ladder
AnTeraju's tutor never sends a full solution. Hints are rungs — one step at a time — and the first rung only opens after the student has attempted. Straight from the Harvard design, straight against the Bastani failure mode.
The answer bank, not the improviser
Tutor minds choose which hint to give; they never invent the maths. Every solution AnTeraju teaches is pre-resolved and syllabus-verified before a model touches it — the same anti-hallucination move that made PS2 Pal work, generalised as provenance.
Struggle windows are a feature
Per-topic friction the learner controls: honest effort before hints unlock, because productive failure is where transfer is built. A tutor that rescues at the first frown optimises for comfort and grades worse on the exam that matters.
Green requires a transfer probe
A topic earns green only after the student handles it in a context they have never seen — new story, new framing, real numbers. Schools teach in one context and examine in another; the mastery gate closes that gap on purpose.
Life-context by default
Any worked example can be re-skinned into the student's actual life: the family phone plan, a payslip with EPF and SOCSO lines, durian-season cash flow, a kiosk inventory sheet. Not decoration — the intervention with the biggest effect on the students who need the most.
Publish or perish the features
The ledger turns every helper into an experiment with a control. Anything that does not move verified mastery inside a term gets demoted to grey on our own status page — rebranded, never. Claims are numbers or they are nothing.
The Malaysian numbers
Useful for real human life means useful against the problems the country's own statistics admit it has.
The OECD's 2026 Economic Survey says Malaysia's problem is understanding, not access — another content portal changes nothing. That is why AnTeraju's gates measure demonstration, not exposure.
MOHE's own 2024 numbers, with skill-related underemployment above 32%. A transcript proves attendance; a mastery ledger proves capability. The difference is exactly what the mismatch is made of.
DOSM 2024. AnTeraju Bakat maps verified mastery against the jobs that actually exist — not the aspirational ladder everyone designs for and nobody fills.
DOSM 2023, while a third of the workforce is underemployed. The mismatch is a signalling failure as much as a training failure — verifiable mastery is the handshake both sides are missing.
A Tuesday, honestly described
A Form 4 student in Ipoh states an intention — pass Add Maths, help at the family kiosk. Tuesday night the tutor sets a problem she hasn't seen, in kiosk numbers, and lets her struggle for eleven minutes because the research says that's where transfer is built. She gets one rung of the hint ladder, not the answer. Wednesday the ledger records what she proved, not what she watched; the same fact shows her mother one honest sentence in Tamil and her teacher which three desks need Thursday's reteach. Nobody was entertained. Something was learned.
The frontier ideas on the AI Deep Stack page all owe these rules a demonstration — every card there is an experiment this page can fail.
See what the deep stack promises — and what it must proveReady to build something intelligent?
Let's engineer your next product — from idea to launch.