Kevin Liu

Relay

A mock team chat app for testing AI agents, with every run graded.

AgentData

Open relay.kevinliu.studioSource

The numbers

  • 70commitsgit rev-list --count f50e9e097
  • 1GitHub stargh api repos/Kevin-Liu-01/Relay --jq .stargazers_count (2026-10-30)
  • 70check filesgh api repos/Kevin-Liu-01/Relay/git/trees/f50e9e097?recursive=1 | grep -cE <glob>
  • 8.75commits per active day70 commits over 8 distinct commit dates, 2026-09-30 to 2026-10-17

Trait

Hidden Grader A hidden check reads only the final state, so its hits ignore barriers and raised Proof.

Relay@f50e9e097:README.md:63-64; README.md:220; scripts/check-workflow-graders.mjs

Moves

  1. Live Browser hits for 100, 95% of the time. Pick a model and a task, then watch every action it takes in a real browser.

    Relay@f50e9e097:README.md:70-72

  2. Head To Head hits for 45 twice, 95% of the time. Two models run the same task from the same starting data in fresh workspaces, side by side, each graded separately.

    Relay@f50e9e097:README.md:94-97; src/live/duel.jsx

  3. Action Menu hits for 80. An AI model picks one action from a menu of visible actions, and the panel shows the probability it gave each.

    Relay@f50e9e097:README.md:165-169; docs/system-one.md:13-17,26

  4. Watch Replay sharpens its next hit. Any recorded run plays back in the real interface, step by step, with its dialogs, typed text and targets.

    Relay@f50e9e097:README.md:82-92

Numbers pinned at f50e9e097, 2026-10-17.

# every page, one line each
$ curl -s https://www.kevinliu.studio/llms.txt
# every card in one file
$ curl -s https://www.kevinliu.studio/llms-full.txt
# this page as Markdown
$ curl -s -H "Accept: text/markdown" https://www.kevinliu.studio/cards/relay

HTTP 200 · text/markdown · 1,809 bytes
# Relay

A mock team chat app for testing AI agents, with every run graded.

Card 16 of 26 in Kevin Liu's collection. 2026. Agent and Data.

- Live: https://relay.kevinliu.studio/
- Source: https://github.com/Kevin-Liu-01/Relay
- Inks: 2
- Health 295, attack 103, defence 122, speed 117

## The numbers

- **70 commits**. Receipt: `git rev-list --count f50e9e097`
- **1 GitHub star**. Receipt: `gh api repos/Kevin-Liu-01/Relay --jq .stargazers_count (2026-10-30)`
- **70 check files**. Receipt: `gh api repos/Kevin-Liu-01/Relay/git/trees/f50e9e097?recursive=1 | grep -cE <glob>`
- **8.75 commits per active day**. Receipt: `70 commits over 8 distinct commit dates, 2026-09-30 to 2026-10-17`

## Trait

**Hidden Grader**. A hidden check reads only the final state, so its hits ignore barriers and raised Proof. Receipt: `Relay@f50e9e097:README.md:63-64; README.md:220; scripts/check-workflow-graders.mjs`

## Moves

- Live Browser hits for 100, 95% of the time. Pick a model and a task, then watch every action it takes in a real browser. Receipt: `Relay@f50e9e097:README.md:70-72`
- Head To Head hits for 45 twice, 95% of the time. Two models run the same task from the same starting data in fresh workspaces, side by side, each graded separately. Receipt: `Relay@f50e9e097:README.md:94-97; src/live/duel.jsx`
- Action Menu hits for 80. An AI model picks one action from a menu of visible actions, and the panel shows the probability it gave each. Receipt: `Relay@f50e9e097:README.md:165-169; docs/system-one.md:13-17,26`
- Watch Replay sharpens its next hit. Any recorded run plays back in the real interface, step by step, with its dialogs, typed text and targets. Receipt: `Relay@f50e9e097:README.md:82-92`

Facts pinned at commit f50e9e097 on 2026-10-17.

Back to every card: https://www.kevinliu.studio/