In build · no learners yet
Not a chat window waiting for a good question. A map of sixty individual skills, a running probability your child knows each one, and a gate they cannot talk their way past.
Two dips are answers the child got wrong after getting it right. The posterior falls, and the gate stays shut until the evidence comes back.
packages/learner-model/src/index.ts · updatePosterior() · masteryGate()
60 skills shipped67 prerequisite links Grades 6–9No trackers, anywhere
01 — the graph
Sixty knowledge components, joined by sixty-seven prerequisite links and six encompasses relations, where mastering the broader skill subsumes the narrower one. The scheduler can only select a skill whose every prerequisite has already passed the gate.
This is the real graph, not an illustration of one — the same file the product reads at runtime.
Nodes fill as you scroll: this is the order the scheduler is allowed to unlock them in.
data/kc-graph/prealgebra.json · 60 components · 73 edges
Every knowledge component the scheduler can select, in curriculum order, with its Common Core reference.
02 — the plan
due reviews, most-overdue first, surfaced the day their memory is predicted to fade
frontier lessons — only skills whose every prerequisite already passed the gate
quiz items from skills already mastered, to keep them honest
Then it interleaves them, because blocked practice feels better and works worse. Pause a topic from the parent dashboard and it is provably absent from tomorrow's plan.
apps/workers/src/orchestrator/plan.ts · assemblePlan()
03 — the living session
Every message in and out passes one package.
Child input reaches a model, and model output reaches a
child, only through packages/safety.
This is not a policy anyone has to remember — a CI test fails the build if any child-facing file imports a model client directly. Chat-template tokens and fake system directives are stripped deterministically before a classifier ever sees the text.
scripts/check-import-boundary.mjs · runs in pretest
Four levels, and the answer isn't in the first three. A nudge, then a targeted sub-question, then one worked step, then a full worked example on a different problem with the same bones.
The structural part matters more than the tone: at levels one through three the correct answer is stripped from the payload before the request is sent. The tutor cannot leak what it was never given — not to a clever prompt, not to a determined twelve-year-old.
packages/safety/src/tutor.ts · redactAnswer()
SymPy decides. Language models never grade. Answers go to a symbolic-algebra service, not to a chatbot.
It proves equivalence symbolically, then by a distinct numeric-refinement path,
then by sampling twenty points to a tolerance of 1e-6. So x+2 and
2+x are both simply correct. Input it cannot parse returns
ungradeable — not wrong — and the learner model records
nothing, so a typo never becomes a phantom gap.
services/mathcheck · POST /verify
Tone, latency and error streak, combined in code. A model reads the message for frustration; code makes the decision.
Code adds a point if response latency crosses 1.5 standard deviations above this session's own baseline, and another if the error streak hits three. Two points shortens the session. Three ends it warmly and flags it for the parent report. The model contributes an observation; it does not make the call.
apps/workers/src/orchestrator/frustration.ts
0.95 confidence and two successes 48 hours apart. Both conditions, always — and the second is the one that matters.
Ten lucky guesses in a single sitting can never unlock a skill, and neither can one good afternoon. Advancement is arithmetic over the learner model: the dialogue cannot override it, cannot be persuaded, and cannot be in a hurry.
packages/learner-model/src/index.ts · masteryGate()
04 — try it
advance if p(known) >= 0.95 AND successes >= 2 AND gap >= 48h
Awaiting evidence — no successes recorded.
This is the stated rule, not a live learner model. No network, nothing retained.
05 — verification
Nothing here depends on a model behaving well. Each stage is a filter that a template either survives or does not.
apps/workers/src/factory · runFactory() · promoteVerifiedToLive()
06 — the learner model
This is the thing that compounds, and the reason leaving hurts: a new tool restarts the diagnostic and the entire spacing history at zero.
Not a SmartScore. A mastery map across the whole graph, every session transcript in full, an append-only safety feed that requires acknowledgement, and a control that pauses a topic — which the next day's plan provably respects.
The promise a parent is buying is narrow and checkable: she can't subtract negative numbers, here is the plan, here is the proof it worked.
packages/db/src/dashboard.ts · withParentScope() · Postgres RLS backstop
A mastery posterior. A running probability the child knows this skill, updated after every answer, discounting lucky guesses and careless slips.
A memory stability. How long this skill will survive without review. Each success stretches the next gap; a lapse cuts it to 30% and brings it back soon.
A personal difficulty. This child's difficulty with this skill, not the population's.
A due date, a rep count, a lapse count, and the moment it was mastered.
packages/learner-model — pure functions, property-tested, 100% branch coverage
07 — evidence
A venture founded on an overstatement is fragile in exactly one way: a competent person reads the literature, finds the overstatement, and stops believing anything else you said.
In 1984 Benjamin Bloom reported that one-to-one tutoring under mastery conditions moved students about two standard deviations. That number is quoted constantly in edtech. Almost every quotation drops the title of the paper, which was “The 2 Sigma Problem” — the two sigma was the target, and the economics of delivering it were the problem Bloom was posing to the field.
What sat underneath it: two University of Chicago dissertations, instructional units of roughly three weeks on narrow content, small samples, experimenter-designed rather than standardized tests, tutoring that replaced classroom instruction rather than supplementing it — and a tutored group held to a 90% mastery bar against a comparison group held to 80%, which means some of the gap is a difference in the bar rather than in the delivery.
Every replication below. None of them reached it. We build on the replicated number, which is still worth building on.
docs/TWO-SIGMA-EVIDENCE.md — binding on all marketing copy
| Study | Scope | Effect |
|---|---|---|
| Cohen, Kulik & Kulik (1982) | 65 studies | 0.33 SD0.84 on locally-written tests · 0.27 on standardized |
| Nickow, Oreopoulos & Quan (2020) | 96 randomized PreK–12 studies | 0.37 SDnot one reached two sigma |
| Kraft, Schueler & Falken (2024) | 265 randomized trials | 0.42 SD0.55 under 100 students · 0.16 above 1,000 |
Kraft's own benchmark puts anything above 0.20 SD in the large band, so these are serious effects. The honest summary is: large effect, contested magnitude, shrinking with scale and with the rigour of the outcome measure. Two sigma is a ceiling observed once, under conditions nobody has reproduced. We do not claim it, and you should distrust anyone who does.
Bastani and colleagues (2024) randomized nearly a thousand Turkish high-school maths students. Unrestricted GPT-4 improved their performance while they were using it by 48%. Then the tool was taken away and everyone sat an exam. That group scored 17% worse than students who never had access at all. A tutor-configured version with safeguards came out statistically indistinguishable from the control group.
Read that carefully, because it is the whole design brief. Performance and learning moved in opposite directions, and the variable that separated them was whether the system handed over answers. The safeguards prevented harm; they did not, in that study, produce a gain. That is why the answer is physically absent from the model's context during an active attempt — and why we are not going to tell you it makes your child learn 48% faster.
Almost nothing about outcomes — and we would rather write that sentence than have you find it out. There are no learners yet, no cohort, no measured result. The held-out assessment banks are frozen and write-protected before the first item is placed in them, and the preregistration directory is append-only, because the temptation to quietly improve the instrument arrives precisely when the results disappoint.
What we will claim is everything on this page above this section: what the system does, what it verifies, what it withholds, and what it refuses to optimize. Daily actives, session length, streaks and self-reported confidence are logged as health signals and will never appear in a claim about learning. When there is a real number, it will arrive with its comparison condition, its instrument, its delay, and its attrition — or it will not arrive.
docs/TWO-SIGMA-EVIDENCE.md · docs/EFFICACY.md — binding on all marketing copy
08 — trailhead
$29/month
$290/year · second child −30% · 14-day trial
Adaptive Tutor is in build. The engine, the verifier, the safety membrane and the parent dashboard run today; the shipped curriculum currently covers 60 pre-algebra skills joined by 67 prerequisite links, and we are widening it before we take anyone's money.
Or write to hello@adaptivetutor.dev directly — the button above opens a mail client, which is not much use inside webmail.
