It tests itself before a child sees it
Every generation renders headlessly and asserts its own interactive invariants. If the pedagogy is wrong, it retries with the error trace. Nothing untested ships.
OpenAI Build Week · Education
A sentence becomes a self-tested manipulative.
A teacher types the misconception she fights every year. Codex writes an interactive, runs its own test, retries until it passes, and publishes a phone-friendly manipulative her students open on any device.
- ~2 min
- sentence → live sim
- 0
- API keys to try it
- 4
- verification gates
- 6
- invariant types
Live · watch it test itself
Math & physics manipulatives
Demo mode — pre-seeded build trace, no live Codex/Supabase call.
The moat
A chat can write code. It can't test it.
Chalkbox executes what it generates, reads the failure, fixes it, and only publishes what passes. The pedagogy itself is a machine-checked assertion — so a sim that renders but teaches the wrong thing never reaches a student.
- G101
Safety + standard gate
Luna triages the prompt: subject, grade band, curriculum code — or rejects it.
accept · math · 6-8 · CCSS.6.NS.A.1 - G202
Generate + static check
Sol writes a single-file React manipulative; AST validation enforces the allowlist.
import allowlist · no network APIs - G303
Render + interactive invariants
It mounts headlessly, then the pedagogy is asserted as a test. Fail → retry with the trace.
assert: smaller divisor, bigger quotient - G404
Output safety + publish
A final safety pass, then it ships under a strict no-network CSP.
CSP connect-src 'none' · null origin
retry-with-traceon any gate failure, within a bounded token budget — never an infinite loop.
Why teachers can trust it
The one artifact a teacher could never make herself.
AI-written code, safely sandboxed
Sims run in a null-origin iframe with a strict CSP (connect-src 'none'), an import allowlist, and AST validation. No network, no surprises.
Standards-aligned by construction
Math and physics, tagged to real Common Core / NGSS codes — the gate rejects anything off-curriculum or off-grade.
A link, not an app store
Students open a zero-chrome manipulative on any phone. No login, no install — just the thing their class is stuck on tonight.
Proof, not promises
Watch it catch its own mistake — and fix it.
This is the real self-debug loop on the flagship fraction-division sim: two rejected attempts, then a verified one. The static validator and the interactive-invariant runner it exercises are real and unit-tested — you can run them yourself.
- 50
- unit + E2E tests green
- 6-stage
- CI: quality → deploy
- 15
- curriculum-tagged sims
- MIT
- open source
chalkbox generate · fraction-division
attempt 1codex reached for live data — fetch() rejected by G2
attempt 2renders, but quotient SHRINKS as divisor shrinks — G3 fails
attempt 3quotient grows as divisor shrinks — G3 invariants green
published → sandboxed share link
The gallery
Browse all 15 Every card is a real standard and the exact sentence that made it.
CCSS 6.NS.A.1
Fraction-Division Bars
“Show why dividing by a fraction makes the answer bigger, not smaller.”
Grade 6Drag the divisor smaller → the quotient (pieces that fit) grows
CCSS 7.NS.A.1
Integer Number Line
“Let them feel that subtracting a negative moves you to the right.”
Grade 7Drag a hop vector along a number line; sign flips direction
CCSS 6.RP.A.3
Double Number Line
“Show that 3 apples for $2 is the same rate as 9 for $6.”
Grade 6Stretch a linked pair of number lines; ratio stays fixed
CCSS 8.EE.B.5
Slope Explorer
“Let them see that a steeper line means a bigger rate of change.”
Grade 8Drag a line's second point; rise/run and slope update live
CCSS 4.NF.A.1
Equivalent-Fraction Wall
“Show that 1/2 and 2/4 cover the exact same amount.”
Grade 4Stack fraction-strip rows; equal widths snap-align
CCSS 8.G.B.7
Pythagorean Squares
“Let them see a² + b² literally fill up c².”
Grade 8Drag legs of a right triangle; the three squares' areas tally
Questions
Straight answers.
Is it safe to run AI-written code in front of students?
Every sim runs inside a null-origin iframe with a strict Content-Security-Policy (connect-src 'none'), an import allowlist, and AST validation before it ever mounts. It can't reach the network, read cookies, or touch the parent page.
How is this different from asking ChatGPT for a simulation?
A chat completion produces plausible-looking code. Chalkbox executes it, asserts the pedagogy as an interactive invariant (e.g. 'dragging the divisor smaller must make the quotient bigger'), reads the failure, and retries with the trace. It only publishes a sim that passed its own test.
Do students need an account or an app?
No. They open a zero-chrome share link on any phone or laptop — no login, no install, no app store.
What subjects and grades does it cover?
Math and physics today, tagged to real Common Core (CCSS) and NGSS standards. The safety gate rejects prompts that are off-curriculum or off-grade-band.
Do I need an API key to try it?
No. The site runs in demo mode with zero keys — the seeded gallery is fully browsable and the Create flow replays the flagship build end-to-end.
Is it production-ready?
It's a runnable demo built for OpenAI Build Week. The harness it exercises — the static validator and the interactive-invariant runner — is real and unit-tested. The live engine that generates a brand-new verified sim from any prompt is the next milestone.
No login · no keys · ~2 minutesType the misconception.
Type the misconception.
Watch it test itself.
The manipulative your class is stuck on tonight — built, verified, and shareable before the bell.