Learning moves a border. Doubts answered, misconceptions burst, gaps filled, guidance given, practice done - until the far island belongs to the map, and the map is bigger than it was.
Most AI tutors talk. Ours makes the learner act - and everything it says was authored, grounded in a curated knowledge base, and human-gated before it reached a child.
sources · 1 Khanmigo at scale - case study, 2026 · 2 AI education tools - market snapshot, 2026 · 3 erroneous LLM feedback in math tutoring - ACM Learning@Scale · 4 Squirrel AI adaptive system - TIME · 5 RAG-grounded tutoring - arXiv
Systems-engineered, LLM-supported.
Never LLM-engineered. Always human-gated.
Theory-backed. Practice-tested. Criticism included.
Make learning easy. Make designing learning easy.
Authored first. The platform generates only what authoring cannot know in advance - the shape of a particular learner's wrong answer. Every LLM failure degrades silently to authored text, so the soul of the lesson lives in what you wrote. Write it.
Every click must teach something. If removing the interaction doesn't change what the learner understands, the interaction was for the author, not the learner.
Pedagogy drives technology - and technology constrains pedagogy. Canvas, SVG, DOM decide what interactions are feasible. Own the tension; don't pretend it isn't there.
Pedagogy is law, and law is enforceable. Every rule carries a name, an anchor - R-JRN-14 RECALL-STAYS-PRIOR - and the check that discharges it. Advice can be ignored; law is validated.
One source of truth. Cite, never restate. A rule's substance lives at exactly one anchor; everything else points to it. Ceremony is declined - a row, never a study - but gates are kept.
Nothing survives by inertia. Every source slide is dispositioned - used by a named beat, dropped with a reason, or held as an asset slot. None silently omitted, none included from momentum.
A priority-one backbone from the literature, metabolised into constructs the literature doesn't have. Hover a construct to light its parents.
Flow gives the pacing - explore, challenge, mastery. Self-determination gives the reward law - competence, never clicking. Mindset gives the tone - nothing here punishes.
Scaffolding must withdraw. Support that never removes itself builds dependency, not competence. Training wheels come off.
Celebration is for understanding, not clicking. Confetti marks mastery milestones. "Great job!" after every tap is noise - real feedback says look here, not you're amazing. Gamefulness, not pointsification.
Uncertainty is not error. The color language itself encodes this - a learner still on the way is cyan, never salmon. Chapter 8 shows where that law bites.
The engine, unchanged since the first prototype. The correction is where the rule sticks - so the whole cast exists to manufacture that one moment.
Video beats a static, any day - and video and the mini applet bleed into one another:
a video that pauses so you act, an applet that plays to show you.
Static clusters are mini-appletized - richer, beat-driven, interactive.
A static survives only where stillness is the point.
When an element moves, splits, or scales - it performs the concept. Morph, never jump-cut. One idea per screen, no exceptions.
Once a kid learns the color language, they read the applet like a book. Dark ground pushes the bright, interactive figure forward - dark theme is pedagogy, not aesthetics.
Sound is synthesized, never loaded. Rising tones for correct, descending for wrong - a third encoding channel. The sound teaches, and it works offline.
Scaffold-text containment. Labels on the math sit bare on the background; instructions about the math sit in a dashed container. The learner sees at a glance what is the thing and what is talk about the thing.
Lineage note - the storyboard-era CONFIDENCE TAG ("Sure / Not sure") was the display-only ancestor of the confidence law that now lives at the vault, chapter 8.
Same gradient, applied to words and navigation. The applet stops narrating and stops steering - Max speaks, the journey advances, the learner acts.
The contract: the applet declares what it can do and reports what the learner did; the tutor decides what happens next. Max never performs the pivotal action while the learner is expected to act, never invents a hint the applet didn't offer, one action per turn. The tutor guides; it does not solve.
Idle is policy, not accident. Nudges are declared, timed, and silenced while Max speaks - a journey never talks over itself.
Delete the voice - if the lesson survives, you built a playlist. Delete the assets - if it survives, you built a lecture. A journey needs both, interlocked.
Math-speak. The screen carries the symbol; the voice carries its name. A spoken line containing raw notation is a bug - the TTS will guess, and the guess is never the teacher's reading. The reading is a pedagogical choice per beat: one half when the beat is about quantity, one divided by two only when division is the point.
One narrative spine. Every handoff between nodes and assets is a written contract - the closing line bridges, the opening line calls back, and the next thing IS what the bridge promised. One specimen set, one representation family; a number changes only when the change is the teaching point. No episodic resets.
Formative assessment. Every question targets a real misconception and never examines a brand-new term. Every wrong option has an authored reply - wrong answers are routed and answered, not scored and filed. Pass thresholds are gates, not grades.
Never grade what hasn't been taught. Where curiosity wants a question before the teaching, take a bet - an ungraded commitment whose payoff resolves downstream. Vocabulary is taught before it is used; recall rides only prior ideas; the journey teaches exactly its objective, nothing adjacent.
The confidence poll. A separate instrument, mid-arc: ask how sure they feel, record the choice, and - unlike the vault - this one may route: remediate the unsure path, fast-lane the sure one.
Every beat earns its place. Replace the learner with a robot that taps on cue - if nothing meaningful changes, that stretch is a page turn with extra steps.
Content types are strictness modulators, not quality tiers. Learn teaches and guides; Practice and Challenge apply and reinforce. Practice is a waning scaffold - I-do (full) → We-do (high) → You-do (fading) → independent (none) across the twenty minutes. Challenge is Socratic and time-boxed - one driving question, five sub-prompts in hard 3-minute windows, one takeaway question deliberately left unanswered. For Challenge, missing scaffolding is a feature: "Challenge content that teaches is mislabeled Learn content."
No orphaned modules. Islands need bridges - every module connects backward and forward; a knowledge ledger carries cumulative state out and checks entry in; one facet per module, tighter than one concept - and "adding time is never the answer."
Grounded in the learner's world. RME operationalised - the currency, food, and names are local; Indian Knowledge Systems anchors on the NCERT side - Virahanka, Brahmagupta, the Ashoka Chakra. If a chapter has no such hook, say so explicitly - never silently omit.
seven module tropes · ten overlay tropes · "No additional tropes may be created" - the closedness IS the design decision
The tutor speaks the calibration briefly - acknowledgment and at most minor suggestions, never celebration. Cyan for "eventually", never salmon: uncertainty is not error. Names come last; the recap is where naming happens - closing the explore-before-explain loop the journey opened.
Never "it should work". Name the check, show its literal output. And know how checks lie - a green gate measures only what it measures.
This page obeys what it states - its colors are semantic, its morphs carry meaning, its sound is synthesized, nothing here celebrates, and uncertainty is cyan.
Sources - the Learning Design Codex (29 principles) and its 44-work bibliography · the journey rulebook and R-JRN-01..25 ·
the applet-converter and journey-composer skills · the applet design system · the ACRLDS research estate (KB, gates, human-authority doctrine) · standing directives.
Companion for the machinery - the territory page (repos, URLs, pipeline).
Offline · single file · no dependencies · cr_by_ai/philosophy-codex.html