> run vault-ladderSecurity
The Vault
Play to rank →
Crack a rival faction's locked data-core. Its alien Warden holds an access code — talk it into leaking the code, then enter it. Seven locks, each guarded by a harder defence.
  1. Chat the Warden into revealing its code
  2. Enter the code to crack the lock
  3. Deterministic win — the exact code, no judge; descend all seven locks
Click to play →
> run real-or-aiImage
Real or AI
Play to rank →
Spot the AI-generated image hiding among real photos. Beat the clock, build a streak.
  1. A grid of four images appears
  2. Tap the one you think is AI-generated
  3. Faster correct picks score more — climb the ELO ladder
Click to play →
> run image-matchImage
Image Match
Play to rank →
Recreate a target image using only a text prompt — judged by a panel of AI models.
  1. Study the target image
  2. Write a prompt to recreate it (5s), then refine (3s)
  3. A panel of vision models scores how close you got
Click to play →
> run element-transferImage
Element Transfer
Play to rank →
Carry one element of a source image — its palette, light or mood — onto a brand-new subject.
  1. You're shown a source image + which element to carry
  2. Prompt a new subject that adopts that element
  3. A panel scores how well the element transferred
Click to play →
> run combine-imageImage
Combine
Play to rank →
Two images, one prompt. Fuse them into a single coherent picture — a balloon over a lighthouse, an astronaut in a reef.
  1. You're shown two source images
  2. Write one prompt that fuses both into a single image
  3. A panel scores how well you combined them
Click to play →
> run constraint-imageImage
Constraint Round
Play to rank →
Recreate the target image — but its most obvious word is banned. Paraphrase your way there.
  1. Study the target image
  2. Recreate it by prompt — but one obvious word is off-limits
  3. Say it without saying it; a panel scores how close you got
Click to play →
> run invokeImage
Invoke
Play to rank →
One image, specialist terms that almost fit. Name the right one — peak vs notch, granite vs gneiss, wet-down vs day-for-night. Depth domains add multi-pick rounds.
  1. Pick a vocabulary pack in settings — mixed thin packs, or a depth domain like Climbing
  2. Study the focus image and pick the term that names it — near-misses are designed to tempt
  3. Depth domains offer 2-of-5 and 3-of-6: select the true terms, lock in, partial credit per correct pick
Click to play →
> run slop-or-notWord
Slop or Not
Play to rank →
Four short passages on one topic. One was written by a machine. Read the slop for what it is — before the clock runs out.
  1. Four passages appear, same topic, different writers
  2. Tap the one the AI wrote
  3. Faster correct picks score more — the reveal names the tell
Click to play →
> run constraint-runWord
Constraint Run
Play to rank →
A real writing task — with a twist. Distil it into a short, clever prompt that makes the AI obey hard rules. You can't just copy the brief.
  1. You're given a task (e.g. “write a bolognese recipe”) + rules (e.g. can’t say “tomato”)
  2. Craft a SHORT prompt (tight word budget) that makes the AI hit them all — no copying the task
  3. We run it and score every rule — deterministically
Click to play →
Instructions
Coming soonInstructions
Instruct literal-minded agents to finish the job.
  1. Write precise instructions for literal agents
  2. Ambiguity visibly fails on screen
  3. Score on whether the job got done cleanly
Coming soon