Clash Jev

No policy is trained. OpenCV plus a small network the author labelled by hand names the troops, and that measurements-only state is what Jev receives. The author is explicit that nothing is trained to play: no reinforcement learning and no simulator. The claim under test is whether Jev's general reasoning transfers to board states the author did not pre-select.

The README reports roughly 150 to 200 Jev requests and about $0.004 per match across the 20-odd matches the author ran. A play costs three requests (strategy, card, square) and declining to play costs one. When Jev picks save_elixir or hold_elixir_for_threat the bot waits for elixir to regenerate rather than re-asking, which is what keeps the count bounded.

A simple baseline policy makes that one decision, and the log gives it its own source so no analysis of Jev includes a move Jev did not make. There is also a hard request cap per run (250 by default, about one match) and every run reports how many requests it made. Code acts only after Jev chooses, so an unaffordable card means the bot taps nothing and logs the pick.

Automating play is against Supercell's terms of service and the README says so, calling it a research project not affiliated with Supercell. Running it needs macOS or Linux with Python 3.12+, adb, and a real Android device with USB debugging on; emulators work but the game's anti-tamper check closes it on some and the author does not work around that. clash-jev snap reads one frame and makes no Jev request, which is the safe way to check calibration first.

The device streams its screen as H.264 over adb, OpenCV reads the hand, elixir and towers, and a small network the author labelled by hand names the troops. Once a second those readings become one JSON game state holding your hand, elixir, next card, troops, tower health and the clock.

Jev picks a strategy from the full list of 12, then one of the four cards in hand, then a legal square, each in its own request because each step depends on the previous. One question over every combination would hold hundreds of near-identical options, so three short lists also show which step went wrong.

Every decision is appended to runs/<timestamp>.jsonl with the state, questions, answers, probabilities, latency and a small hand image, alongside an optional 480-pixel 10-fps video sharing the same clock. Published replays let you pause any decision and re-ask Jev, though the state and briefing cannot change because the experiment is the same moment with a different question.