01
What does “losing” mean?
Shallow Red succeeds when it gets checkmated.
Winning is a failure. Draws are also undesirable. Given two successful ways to lose, it prefers the faster one—but reliability always comes before speed.
That last rule matters. Several experimental versions lost quickly when things went well, but drew or survived more often overall. They were rejected.
02
How it learned
Training positions were generated by playing against several kinds of opponents:
- Stockfish trying to win normally
- Weak and noisy chess engines
- Opponents choosing randomly
- Opponents deliberately resisting Shallow Red's attempts to lose
- Earlier versions of Shallow Red
Every legal move in a sampled position was ranked according to how likely it was to lead to Shallow Red's eventual checkmate. Those rankings were used to train a small residual neural policy.
The selected research model used a 32-channel, four-block network trained on thousands of positions generated through reverse Stockfish analysis and random-opponent simulations. The neural policy alone was not reliable enough. Search remained essential.
03
What runs on this website
The web version is a compact distillation of the research engine, not the complete neural model.
On every turn, it examines every legal move and every legal immediate reply. It looks for positions where the opponent can checkmate it, take its pieces, attack the area around its king, or restrict its movement.
It also avoids:
- Checkmating the opponent
- Falling into a draw
- Taking valuable enemy pieces
- Giving away its final pieces when doing so would create a stalemate
All of this runs inside your browser. Playing against Shallow Red does not require a GPU or a permanent chess server.
04
How bad is it?
Against a uniform-random opponent, the selected research system was checkmated in 281 of 300 frozen evaluation games:
Against Stockfish at 1,000 search nodes per move, it lost all 100 evaluation games. In the primary random-opponent evaluation, successful losses took a median of 70 plies—about 35 full moves.
These results describe specific test populations, openings, random seeds, and move limits. They do not prove that Shallow Red can never win.
05
Why not just reverse Stockfish?
Stockfish is very helpful when the opponent wants to win. It eagerly accepts hanging pieces and finds mating attacks.
Random and weak opponents are harder. They overlook free material, ignore attacks, miss checkmates, and sometimes accidentally create stalemates. Shallow Red therefore had to learn how to lose without relying on a cooperative winner.
06
What about self-play?
Two copies of Shallow Red playing each other produced 40 draws by repetition.
That makes sense: both sides cannot be the one that gets checkmated. Useful self-play has to be asymmetric, with one side learning to lose and the other playing a fixed or independently defined role.
07
Things that didn't work
- Models trained to lose faster became less reliable.
- Rollout search looked faster only because it converted difficult losses into draws.
- Repetition penalties reduced overall performance.
- Standard chess tablebases rarely affected actual games.
- An exact 2.8-million-position bishop-versus-rook analysis found no nonterminal position where the losing side could force its own checkmate against perfect resistance.
Negative results were kept instead of quietly discarded. The goal was not to produce an impressive metric. The goal was to produce the most reliably terrible chess engine we could.
The guiding rule
First, lose consistently.
Then, and only then, lose faster.
Credit
Shallow Red was researched, trained, evaluated, built, and deployed in collaboration with Codex using gpt-5.6-sol-high.
Back to the game