Skip to content

About

Open-source whiteboard AI tutor with animated lessons and interactive playback. Compare Claude, Gemini & DeepSeek across 288 generation runs and 64 blind-reviewed boards.

Topics

Resources

Stars

220 stars

Watchers

6 watching

Forks

Repository files navigation

Scrplay Vip Review Reliability Hands Free Scrplay Did Not Enough Become

That puts scrplay vip in the solid but not special category which is exactly where a lot of mixed platforms land when you strip away the marketing gloss the pros are clear enough easy browser play decent variety and Scrplay Vip Safety Look Payments Scrplay Delivers On User Plus Mobile a vip framework that may suit regular depositors.

Here s my detailed user review of scrplay vip focusing on bonus experience and overall reliability from a mobile first perspective this review examines everything from registration to payouts to help you decide Scrplay Review Rm66 Cashout Vip Look In Terms Of Mixed Based Everything if this platform meets your gaming needs.

Scrplay Review Scrplay Table Games Members Will Get Bonuses New

Scrplay vip review, if you want a sportsbook and casino platform with a vip style angle and a usable desktop experience scrplay vip is worth a closer look it is especially relevant if you Scrplay Trusted Online Rm66 Games Test Members Will Get Rm66 First Performance enjoy switching between sports betting slots and live dealer tables in one place.

Scrplay offers up to rm66 free credit for new members and maintains strong uptime during malaysian public holidays when traffic spikes Scrplay Vip Bonuses Balanced If You Want Rm66 Look Enforced slots and fishing remain the core draw community feedback highlights consistent payout behaviour on e wallet routes one account per ip policy is strictly enforced do not test with duplicate registrations.

Scrplay Vip Hands Casino In Terms Of Tables Sportsbook Closer

An in depth audit of scrplay s rm15 bonus in malaysia review wagering terms game Scrplay Review 4 Rm15 Table Full Claim Rm50 Free Closer Regular weightings kyc protocols and payout reliability, scrplay review 4 8 5 full hands on test, scrplay vip review bonuses games reliability for malaysian players.

Claim rm50 free credit no deposit needed plus top slots live casino and instant payouts register now, scrplay delivers on all three without compromise licensed operations tested games transparent bonuses Scrplay Vip Review Rm15 Roller High In Terms Of Exactly Payout This and a team that treats every player fairly that is what separates a truly trusted casino online malaysia from an average one.

Scrplay trusted online casino malaysia, scrplay did not win on the loudest rm number it won our second place Scrplay Vip Ranking Roller Malaysian Online If You And Review Because because when we asked messy promo questions agents answered like humans who had read the terms.

Scrplay Vip Table Malaysian Scrplay Scrplay Delivers Draw Requirement Malaysia

Scrplay vip review a balanced look at safety bonuses games payments, scrplay vip table games review a high roller s step by, in terms of uptime and performance scrplay vip remained rock solid during my sessions no crashes no lag spikes comparing it to other malaysia friendly platforms in 2026 it ranks high Scrplay Vip Ranking Bonus Scrplay Malaysian Members Will Casino Player on reliability along with having a strong focus on table games.

An open-source AI tutor that teaches on a whiteboard — and a benchmark of which LLM does it best.

It writes, draws and erases while it explains, and the student can interrupt.
One frozen prompt. Six models. 288 lessons drawn stroke by stroke. Blind-reviewed.

English · 简体中文

▶ Live gallery: replay every lesson in your browser

licensemodelsapi callsboards reviewedtotal cost

A model proves the Pythagorean theorem on a whiteboard, stroke by stroke, with narration

Gemini 4 Argon proving the Pythagorean theorem. The model only wrote JSON; the player turned it into handwriting. Played at 3× speed.

What is this?

An AI tutor (or AI teacher) that writes, draws and erases on a whiteboard while it talks — and lets the student cut in — looks like a model capability. It isn't. It is a small protocol: the model streams drawing instructions as JSON, a browser plays them as handwriting, and when the student interrupts, the app sends the board back to the model.

So any LLM can be the tutor. This repository is two things:

  • A reference implementation of that loop: the protocol, a whiteboard player, and a runner for Claude, Gemini and DeepSeek.
  • A benchmark of how well they teach: same prompt, same eight scenarios (science, maths, economics), six models, two thinking settings, three repeats. Every board is in this repo, and you can replay each one.

It is not a hosted tutoring app. The player replays recorded lessons; to generate new ones, run the benchmark runner with your own API key.

Leaderboard

Leaderboard: blind-review pass share, time to first stroke, cost per lesson
ModelBlind review: boards passedStrictFirst stroke (lesson, p50)Cost per lessonMain weakness
Gemini 4 Argon11 / 11118.0 s$0.0277slowest to start; price doubles on 2026-12-01
Claude Haiku 5.59 / 11912.9 s$0.0019thinks even on the low setting, so it starts late
Gemini 3.8 Flash8 / 1183.4 s$0.0080marks points off the curves, mislabels positions
Claude Sonnet 5.57 / 1152.8 s$0.0158cleanest layout, but the figure often contradicts the text
DeepSeek V4 Pro5 / 1043.8 s$0.0064overprinted text, covered diagrams (thinking off)
DeepSeek V4.1 Flash4 / 1042.5 s$0.00177 of 24 calls drew a blank board (thinking off)

Latency and cost are for the low-thinking setting. "Strict" also fails boards where a marked point sits 20–35 px off the true intersection. Eleven boards per model is a small sample: the gap between the top and the bottom is clear, the order in the middle is not. See Limitations.

Same question, six models

"Why does the Pythagorean theorem work?" — low-thinking setting, one board per model, exactly as drawn.

Gemini 4 Argon
Gemini 4 Argon · pass · two arrangements, exact geometry
Gemini 3.8 Flash
Gemini 3.8 Flash · pass · same proof, one misplaced highlight box
Claude Haiku 5.5
Claude Haiku 5.5 · pass · algebraic version
Claude Sonnet 5.5
Claude Sonnet 5.5 · pass · wrote the algebra, then erased it
DeepSeek V4 Pro
DeepSeek V4 Pro · fail · the c² square is not on the hypotenuse and has the wrong size
DeepSeek V4.1 Flash
DeepSeek V4.1 Flash · fail · no actual argument on the board

All 64 reviewed boards, with verdicts and notes, are in the live gallery; click a board to watch it being written.

What we found

  1. The words are right; the pictures are wrong. Of the 20 boards that failed, 12 failed because the figure contradicted the text: a season label on the wrong Earth, an equilibrium point drawn beside the intersection instead of on it. Written statements were almost never wrong.
  2. There is no single winner. Gemini 4 Argon drew the most accurate boards (11 / 11). Claude Sonnet 5.5 started drawing first (2.8 s) and had the cleanest layout. Claude Haiku 5.5 was 8–15× cheaper than either and still passed 9 / 11.
  3. A cheap, fast model with thinking off can simply not follow the protocol. DeepSeek V4.1 Flash returned instructions without the op field in 7 of 24 calls. They parse as JSON and draw nothing.
  4. Our linter was a bad judge. It flagged 41 of the 64 boards for "hard errors"; the blind review passed 26 of those. It punishes rich diagrams (labels near lines) and cannot see the main failure, a figure that disagrees with its own caption.
  5. More thinking mostly bought waiting. With thinking on, the first stroke took 30–120 s on Gemini 3.8 Flash and both DeepSeek models, and 23 calls ran into the 8,000-token output cap before finishing the board.
  6. Interruptions are handled, but messily. All 24 interruption boards answered the student. Eleven of them wrote over the existing drawing or wiped the board to start again.
season labels swapped
Labels swapped. The Earth leaning toward the Sun is labelled "North: winter". (Claude Sonnet 5.5)
equilibrium point off the intersection
Point off the crossing. E₂ sits on the supply line, above where D₂ meets S, so P₂ and Q₂ are read wrong. (Gemini 3.8 Flash)
overprinted titles
Overprint. Every step title written on the same spot. (DeepSeek V4 Pro)

How it works

flowchart LR Q([Student question]) --> M[LLM + frozen protocol] M -- "streams JSON ops" --> P[Player] P -- "handwriting, strokes, erasing" --> B[(Board)] I([Student interrupts]) --> S[Board state:<br>objects, positions, text] B --> S S -- "question + state" --> M 
Loading

The model never touches a canvas. It emits steps, each with a line of narration and a list of operations:

{ "title": "The tilt", "say": "Earth's axis is tilted 23.5 degrees and keeps pointing the same way.", "ops": [ { "op": "erase", "ids": ["guess", "check"] }, { "op": "write", "id": "key", "x": 60, "y": 128, "text": "Real reason: the axis is tilted 23.5°", "size": 36, "color": "blue" }, { "op": "ellipse", "id": "earth1", "cx": 280, "cy": 600, "rx": 34, "ry": 34, "color": "blue", "fill": "blue" }, { "op": "line", "id": "axis1", "x1": 255, "y1": 657, "x2": 305, "y2": 543 } ] }

Six operations (write, line, rect, ellipse, path, erase) on a 1600 × 900 board. The full prompt is protocol/PROTOCOL.md (the tested prompt, in Chinese) with an English translation. Because the first operation is usable as soon as it arrives, the student sees the first stroke after two to four seconds on the fastest models, long before the lesson has finished generating.

Interruptions

A student interrupts a supply-and-demand lesson; the model continues on the same board

When the student cuts in, the app stops the stream and sends the model the question together with what is on the board (every object's id, type, bounding box and text). The model replies with new steps only; it may erase or annotate existing objects by id.

How it was evaluated

flowchart LR R[288 runs] --> A[Automatic metrics:<br>parse, protocol, latency, cost] R --> L[Linter:<br>off-board, overlaps] R --> X[Render final boards] X --> Z[Random ids,<br>model names hidden] Z --> V[Blind review<br>with a fixed rubric] V --> U[Unblind,<br>aggregate] U --> H[Human audit sample] 
Loading
  • Sample. One board per model and scenario at low thinking (48), plus three scenarios at medium thinking (16). Of the three repeats, the one with the median linter score was taken, not the best one.
  • Blind. Boards were rendered to images under random ids. Scores were written to disk before the id mapping was opened (bench/blind_view.py never prints a model name).
  • Rubric. No scientific or mathematical error; the figure agrees with the text; layout 0–2; the lesson is complete. Interruptions add: the question is answered; the existing board is not overwritten or wiped. Pass = no error, figure agrees, layout ≥ 1.
  • Reviewer. Claude Fable 5.2, which was not one of the tested models. It checked geometry from pixel positions (is the dot on the intersection? which way does the axis lean?).
  • Human audit.audit.html lists 18 boards (every low-confidence or borderline call plus a random pass and fail per model) for a human to confirm or overturn.

Details and the exact reproduction steps are in docs/METHOD.md. The full write-up is docs/REPORT.md (中文).

Quickstart

Look at the results in the live gallery, or locally (no keys, no dependencies):

git clone https://github.com/molly554/whiteboard-ai-tutor.git cd whiteboard-ai-tutor python serve.py # then open http://localhost:8811/

The gallery opens; click any board to watch it being written. player/ also plays the eight scenarios and shows the linter output.

Run a model yourself:

pip install -r requirements.txt # one key covers Claude and Gemini via OpenRouter; DeepSeek uses its own keyexport OPENROUTER_API_KEY=... # Windows: powershell -File bench/set_key.ps1 -Name OPENROUTER_API_KEYexport DEEPSEEK_API_KEY=... python bench/run.py preflight --via openrouter # free: checks keys, model ids, pinned providers python bench/run.py smoke --via openrouter --models sonnet-5.5 # 4 calls, a few cents python bench/run.py main --confirm --via openrouter # the full matrix, about $4 python bench/aggregate.py # tables in results/summary.md

The runner pins every OpenRouter model to its first-party provider and disables fallback, stops at a per-model and a total budget cap, and appends each call to results/runs/*.jsonl so an interrupted run can be resumed. bench/mock_server.py imitates the three APIs locally if you want to test the runner for free.

Repository layout

protocol/ the frozen prompt (tested version in Chinese, plus an English translation) scenarios/ 8 scenarios: 5 lessons, 3 interruptions with their fixed prior board and board state player/ the whiteboard player, the linter and the frame-capture tool (one HTML file, no build step) bench/ runner, provider adapters, aggregation, blind-review helpers, chart and animation builders results/ runs/ every API call: timings, token usage, cost, raw output scripts/ the 263 outputs that parsed, ready to replay lint/ linter output per run review/ 64 rendered boards, the blind scores, the id mapping, per-board notes docs/ report, method, chart and animation assets index.html gallery of all reviewed boards (also the GitHub Pages entry point) audit.html the human-audit sheet 

Limitations

  • Small review sample. Eleven boards per model. 9 / 11, 8 / 11 and 7 / 11 are not distinguishable.
  • Final boards only. The review looked at each finished board, plus the state before the largest erase on six boards. Intermediate states and animation pacing were not reviewed.
  • One prompt. The protocol is frozen at one wording, written in Chinese. Another wording could reorder the models.
  • Eight scenarios, three repeats. Middle-school science, maths and economics; nothing about real student questions at scale.
  • Latency was measured from one desktop, and Claude and Gemini went through OpenRouter, which adds a hop. Compare models with each other, not with your own production numbers.
  • The medium-thinking results are capped. 23 calls hit the 8,000-token limit, so parse rates there measure truncation, not ability.
  • Handwriting depends on your system font. The player uses Ink Free, which ships with Windows. Elsewhere it falls back to a generic cursive font, so a replay can look different from the reviewed board images, which were rendered on Windows.
  • Prices move. Costs are list prices on 2026-11-11.

Ideas worth testing next

  • Higher-level operations (mark_intersection(a, b), label(id, text)) so that code, not the model, computes positions. This targets the most common failure directly.
  • A second pass that shows the model a screenshot of its own board.
  • A vision-model judge compared against human verdicts, to replace the linter as a gate.
  • Speech for the narration, and true mid-stroke interruption.

License

MIT. Model outputs in results/ are included for research and comparison.

About

Open-source whiteboard AI tutor with animated lessons and interactive playback. Compare Claude, Gemini & DeepSeek across 288 generation runs and 64 blind-reviewed boards.

Topics

Resources

Stars

220 stars

Watchers

6 watching

Forks

Releases

Packages

Contributors

Languages