Technology

Chalk-1 — our own teaching model

Chalk-1 — our own teaching model

Most AI tutors are prompts on top of someone else’s chatbot. brightboard runs on Chalk-1 — a speech-native model we train ourselves to think, speak, listen and drive a live whiteboard, all at the same time.

3,000+

hours of real tutoring behind the training corpus

30k+

real tutoring turns — growing with every lesson

<2%

of steps ever corrected by the live checking layer

~12p

per student-hour of self-hosted inference

The model

Why we train our own model

Why we train our own model

No chatbot can do this

Teaching on a live board means speaking, listening and writing at the same time. That is a trained capability, not a prompt — no general chatbot offers it.

Aligned to the exam board

Chalk-1 is trained for GCSE teaching against your exact exam board, in brightboard’s own whiteboard language — not general-purpose chat.

Verified before it teaches

A live checking layer verifies every step before it reaches the board. Fewer than 2% of steps are ever corrected.

Economics that scale

Self-hosted inference runs at roughly 12p per student-hour — the difference between a demo and an always-on tutor every family can afford.

Training

The training flywheel

The training flywheel

Chalk-1 is not trained once — every new version goes through the same loop, and every brightboard lesson strengthens the corpus. A dataset no wrapper can copy.

01

Learn from real teaching

Chalk-1 trains on thousands of hours of real tutoring — and on every lesson it teaches on brightboard.

02

Filter relentlessly

Every candidate training example passes hard automated quality checks. Anything that fails never reaches the model.

03

Prove it before it ships

No version is deployed until it clears a fixed evaluation suite it has never seen — teaching complete lessons, scored blind.

Evals

Measured against the frontier

Measured against the frontier

Every version of Chalk-1 teaches complete lessons that are scored blind, on a 5-point rubric, across five GCSE maths modules — with a frontier model as the bar. We publish the trajectory with every release.

3.15

3.60

4.30

4.35

Chalk-1 v1

Chalk-1 v2

Chalk-1 v2.3 — current

Opus 5 (benchmark)

Chalk-1 v2.3 scores 4.30 against Opus 5’s 4.35 — within 0.05 of the frontier benchmark, at a fraction of the cost to run.

What’s next

Chalk-2 is already in training

Chalk-2 is already in training

Every GCSE and A-level maths and science board, then English and the humanities — benchmarked head-to-head against human tutors.