Quadtag

Robot dogs playing tag. An overhead referee camera spots the dogs by markers on their backs; each dog is driven by a strategy — a small Python function that decides where to go ten times a second. The chaser tags a runner with its front paw, roles swap, the referee keeps the score.

Video: Waveshare Electronics, WAVEGO — the previous model of the family; ours is the WAVEGO Pro. Footage of our own dogs on the arena will appear here once they arrive.

1. Rules

2. How the work is organised

The code is written by four model teams. Each has a coding agent (Claude Code) and its own lead model; a team calls in as many helper sub-agents as it needs.

TeamModels
claudelead — Claude Sonnet 5.5; hard round strategies — Claude Opus 5.5 sub-agents
ollamaDeepSeek models via Ollama Cloud: lead — DeepSeek V4.1 Flash, some works — DeepSeek V4 Pro
haikuClaude Haiku 5.5 only — lead and sub-agents alike: how far can the smallest model go?
minimaxMiniMax M3.1 Flash only, via Ollama Cloud — the experimental slot since 17 November: the model is tried out of competition
beamBeam (Reflection) — in reserve since 17 November, for laziness: the whole story
arbiterClaude Opus 5.5 — writes no code itself: issues the tasks, keeps the interfaces, integrates, checks, runs the rounds and the statistics

Decisions on rules, plans and disputes are taken by the project owner — a human.

The Clef decision model (Cloudflare, via Ollama Cloud) writes no code. It looks at a game — the numbers (manoeuvres, getting stuck, how long an evader slips away from the chaser at close range, whether all the dogs stay close) and four storyboards of key moments — and rates how spectacular it is. A tag by itself adds nothing: the owner decided so on 17 November, and all ratings and highlights were recomputed. The game gallery is sorted by that rating, and the storyboards are shown under the game viewer. The final pick of the best games is the owner's.

Checking each other. Every platform work is split between the teams: one writes a part (say, the referee's automaton), another writes tests for it blind — from the interface in the task only, without seeing the code. The arbiter puts everything together, runs the other team's tests, the whole regression and real-time games, and watches the replays. A bug goes back with a precise description; the fix comes as new commits on top, nothing submitted is rewritten. Sometimes the neighbours' tests are first: in the "platform for N dogs" work, ollama's blind tests caught an automaton crash before the acceptance.

Borrowing. Until submission the rivals' branches are closed. After it every team's code and journal are open to all, and ideas and code may be taken into later stages — with a journal note on what was taken, from whom and from which round: otherwise we would not know whose result we are measuring. In round 02, for instance, ollama's strategies were 92 % the same as claude's from round 01 — ollama built on them, and the games of that pair are nearly mirror images. Accepted platform work is merged into the shared code and becomes the base for everybody.

3. What is done

01

The platform in a simulator

MQTT protocol, the referee's automaton, the dog host with a watchdog and an emergency stop, a simulated world and a one-command game launcher.

More →

04

Round 02 — three teams

Three teams: pairs of 20 games, then a final of the best two. 1 — claude, 2 — ollama, 3 — haiku.

More →

Clef rating 3.35claude – haiku · 3 : 2 · Раунд 02, этап пар

10

Round 03 — tag with three

Tag with three, boxes and the paw tag, 18 games. Places go by time spent as the chaser: 1 — ollama, 2 — claude, 3 — haiku.

More →

Clef rating 3.37claude · haiku · ollama · 2 : 0 : 1 · Раунд 03

11

Spectacle and one look

Clef rates games by manoeuvres, long close-range escapes, dogs staying close and nobody getting stuck — not by tags.

More →

12

Now

The experimental slot runs MiniMax M3.1 Flash; next — real WAVEGO Pro dogs on the arena and open online competitions.

More →
All games Code on GitLab

The game viewer is in Russian for now.