REMY (Rapid Evaluation, Maximum Yield) is a chess engine written from scratch, one step at a time, that keeps a new idea only when it proves itself in games.
The goal: beat Stockfish at 2 minutes plus 1 second per move on an ordinary desktop computer. On the CCRL Blitz list REMY's ratings use (computed October 17, 2026), the highest-rated Stockfish, Stockfish 19, is rated 3793 on one processor core and 3804 on eight. Why, and the milestones on the way.
Current version: REMY 9, promoted 2026-11-06.
Narrows its search around the score it expects, judges captures before trying them, and runs the same search faster, so it sees further in the same time.
Estimated rating 2213 (likely range 2182 to 2244)
Estimated CCRL Blitz-scale rating (REMY measurement), not a CCRL rating
First milestone, 1500, passed by REMY 7. Second milestone, 2000, passed by REMY 9. What this means · Next milestone: 2500 on the CCRL Blitz scale.
What made REMY 9: it starts from REMY 8.1, so it judges positions with the final learned square values. It adds two search ideas, aspiration windows and static exchange evaluation. With aspiration windows, REMY first looks for the best move within a narrow band around the score it expects, and widens the band only when the answer falls outside it. Static exchange evaluation works out in advance whether a series of captures on one square wins or loses material, so losing captures are tried last and hopeless ones are left out when REMY settles the captures at the end of a line. In its promotion match against REMY 8.1, REMY 9 won 47 games, lost 24 and drew 51: about 66 rating points better. It also carries two rounds of speed work, which make the same search about 2.3 times faster. The speed work changes no move: at any given depth REMY 9 chooses exactly the same moves as before, only sooner, so on the same clock it can look further ahead. The timed checks of the speed work were exploratory.
REMY's learning of its square values finished after three rounds. Josh registered the third round's values as the final ones on 4 November 2026. On games it had never seen, the final values predicted results clearly better than counting material alone.
Endgames in REMY 8's Lichess games: a Stockfish check found that endgames were rarely where it lost. In 96 of its 165 losses Stockfish already judged it clearly behind in the opening, and in 63 more in the middlegame; only 4 turned in the endgame. From endgames it reached two pawns or more ahead, it won 29 of 30 and drew the other. Twice it let a won position slip into a draw by repeating moves. Since 9 November 2026 the bot plays REMY 9.
Next idea: pawn shelter. REMY's Lichess games showed it pushing the pawns in front of its king early. A Stockfish check of those games found those pushes cost about twice as much as its other moves. This step teaches REMY to value its king's pawn cover. It is a hypothesis being tested. Stockfish is used only to diagnose, never to teach REMY. Where it stands: the data for it has been generated, and the fit waits for Josh's judgement suite, a set of positions he rates by hand. Before its match, a short match is checking a repair to how REMY treats repeated positions.
Estimated CCRL Blitz-scale rating (REMY measurement), not a CCRL rating
REMY 0 picks a random legal move. Each version since has added one idea, kept only after it beat the version before it: counts material (REMY 1), checks your reply (REMY 2), looks a move further (REMY 3), uses its clock (REMY 4), finishes exchanges (REMY 5), remembers positions (REMY 6), learned square values (REMY 7), searches deeper (REMY 8), final square values (REMY 8.1) and sharper, faster search (REMY 9).
REMY 9's rating, 2213, is an estimate on the CCRL Blitz scale: REMY 9 played 600 games against 6 engines whose CCRL Blitz ratings are known (the anchors), and 2213 is the rating that best explains its scores against them. CCRL has not tested REMY, so it is not an official CCRL rating.
On the CCRL Blitz list REMY's ratings use (computed October 17, 2026), the highest-rated Stockfish, Stockfish 19, is rated 3793 on one processor core and 3804 on eight.
These are targets, not forecasts: nothing here predicts when, or whether, REMY reaches them.
These values were learned from REMY's own games. They show what helped REMY win with each piece on each square, relative to that piece's average square (0): blue squares helped, orange ones hurt. Values are in centipawns, where 100 is a pawn in the middlegame.
One colour scale for both king boards: −410 to +410 centipawns (orange below the piece's average square, blue above).
Hover over or tap a square to see its value.
| a | b | c | d | e | f | g | h | |
|---|---|---|---|---|---|---|---|---|
| 8 | +111 | +42 | -14 | -29 | +56 | +110 | -38 | -410 |
| 7 | +61 | -55 | -18 | -71 | +5 | -207 | -172 | -79 |
| 6 | +52 | -77 | -94 | -35 | -51 | -106 | -160 | -184 |
| 5 | +52 | -136 | -152 | -216 | -163 | -183 | -175 | -249 |
| 4 | -74 | -98 | -168 | -163 | -180 | -175 | -110 | -145 |
| 3 | +102 | +93 | +58 | -6 | -12 | +31 | +88 | +95 |
| 2 | +245 | +201 | +186 | +125 | +118 | +171 | +197 | +239 |
| 1 | +209 | +211 | +206 | +183 | +222 | +194 | +276 | +266 |
| a | b | c | d | e | f | g | h | |
|---|---|---|---|---|---|---|---|---|
| 8 | +24 | +41 | +65 | +83 | +44 | +39 | +39 | +39 |
| 7 | +28 | +64 | +60 | 0 | +52 | +52 | +50 | +60 |
| 6 | -5 | +62 | +72 | +50 | +28 | -47 | +69 | +44 |
| 5 | -4 | +53 | +69 | +83 | +62 | +59 | +53 | +27 |
| 4 | -21 | +30 | +54 | +66 | +71 | +58 | +32 | +1 |
| 3 | -53 | -33 | -10 | +18 | +31 | +13 | -34 | -53 |
| 2 | -96 | -77 | -63 | -41 | -39 | -49 | -76 | -103 |
| 1 | -117 | -97 | -101 | -136 | -150 | -134 | -143 | -163 |
Read the board. In the middlegame, REMY's king did best on g1 (+276) and h1 (+266) and worst on h8 (-410). In the endgame, REMY's king did best on d5 (+83) and d8 (+83) and worst on h1 (-163).
| Piece | Middlegame | Endgame |
|---|---|---|
| Pawn | 100 (fixed: the scale) | 271 |
| Knight | 383 | 350 |
| Bishop | 439 | 446 |
| Rook | 556 | 809 |
| Queen | 1894 | 918 |
These reflect what REMY's games rewarded, not textbook piece values.
Set p1-g2 (REMY 8.1), SHA-256 eaf6a17db288….
Narrows its search around the score it expects, judges captures before trying them, and runs the same search faster, so it sees further in the same time.
Against REMY 8.1 at 120+1 (seconds + increment): 47 wins, 51 draws, 24 losses over 61 pairs of games (each opening played with both colours).
REMY 9 did clearly better than REMY 8.1, so it replaced it.
Keeps the same search and judges positions with the square values from its third and final round of learning.
Against REMY 8 at 120+1 (seconds + increment): 61 wins, 44 draws, 17 losses over 61 pairs of games (each opening played with both colours).
REMY 8.1 did clearly better than REMY 8, so it replaced it.
Learned to search deeper in the same time, by spending less of it on moves that look weak.
Against REMY 7 at 120+1 (seconds + increment): 80 wins, 20 draws, 6 losses over 53 pairs of games (each opening played with both colours).
REMY 8 did clearly better than REMY 7, so it replaced it.
Learned from its own games how much each piece is worth on each square, in the middlegame and in the endgame.
Against REMY 6 at 120+1 (seconds + increment): 88 wins, 14 draws, 4 losses over 53 pairs of games (each opening played with both colours).
REMY 7 did clearly better than REMY 6, so it replaced it.
Learned to remember positions it has already analysed and to try the most promising moves first, so it can search further in the same time.
Against REMY 5 at 120+1 (seconds + increment): 99 wins, 6 draws, 1 loss over 53 pairs of games (each opening played with both colours).
REMY 6 did clearly better than REMY 5, so it replaced it.
Learned to finish exchanges before judging a position, so it no longer stops counting in the middle of a trade.
Against REMY 4 at 120+1 (seconds + increment): 96 wins, 5 draws, 5 losses over 53 pairs of games (each opening played with both colours).
REMY 5 did clearly better than REMY 4, so it replaced it.
Learned to use its clock: it thinks deeper and deeper while time allows and plays the best move from the deepest search it finished.
Against REMY 3 at 120+1 (seconds + increment): 98 wins, 2 draws, 6 losses over 53 pairs of games (each opening played with both colours).
REMY 4 did clearly better than REMY 3, so it replaced it.
Learned to look a move further ahead (its move, your reply and its next move), skipping lines that cannot change its choice.
Against REMY 2 at 120+1 (seconds + increment): 969 wins, 11 draws, 20 losses over 500 pairs of games (each opening played with both colours).
REMY 3 did clearly better than REMY 2, so it replaced it.
Learned to check your reply: before taking a piece, it looks at what you can take back.
Against REMY 1 at 120+1 (seconds + increment): 997 wins, 0 draws, 3 losses over 500 pairs of games (each opening played with both colours).
REMY 2 did clearly better than REMY 1, so it replaced it.
Learned to count material: it grabs whatever it can capture, without yet checking whether the piece is defended.
Against REMY 0 at 120+1 (seconds + increment): 991 wins, 9 draws, 0 losses over 500 pairs of games (each opening played with both colours).
REMY 1 did clearly better than REMY 0, so it replaced it.
Knows the rules of chess and nothing else: it picks a random legal move.
Before some measurements, we write down what we expect, with a range, and publish every forecast with its outcome, misses included. They test how well self-play results predict REMY's strength against other engines; they are forecasts, not results.
| Recorded | What it forecasts | Forecast | Outcome | What it taught us |
|---|---|---|---|---|
| 1 November 2026 | REMY 8's rating against other engines | 1775 (range 1700–1850) | Miss: measured 1865 (likely range 1834–1896) | REMY 8's search gains turned into strength against other engines better than expected: about 0.85× its self-play gain, against about 0.35× for the earlier search stage. |
| 1 November 2026 | How much the first build of REMY 9 (REMY 8 plus aspiration windows and static exchange evaluation) gains in rating over REMY 8, if it is rated | +30 (range 0 to +60) | Not tested Not scored: it forecast REMY 9 as it stood on 1 November, before the final learned square values and the speed work; the version rated had both. | |
| 1 November 2026 | How much a later round of learned square values adds to REMY's rating, with the same search | +100 (range +60 to +150) | Pending To be scored on the rating gain from REMY 8 to REMY 8.1, whose square values are the final (third-round) ones. | |
| 1 November 2026 | The rating of the first promoted version with REMY 8's search, the final learned square values and the next search step | 1925 (range 1850–2000) | Not tested Not scored: it forecast REMY 9 as it stood on 1 November, before the final learned square values and the speed work; the version rated had both. | |
| 3 November 2026 | How much the first promoted pawn-shelter version gains in rating over the version before it | +35 (range +20 to +60) | Pending | |
| 3 November 2026 | In its promotion match, how often the pawn-shelter version moves its a-, b-, g- and h-pawns in its first 12 moves, divided by its opponent's rate | 0.5 (range 0.3 to 0.7) | Pending | |
| 3 November 2026 | How many a-, b-, g- and h-pawn moves the pawn-shelter version plays in its first 12 moves of a Lichess game, on average | 1.3 per game (range 1.0 to 1.6) | Pending | |
| 3 November 2026 | In its promotion match, how much more often than its opponent the pawn-shelter version castles within its first 12 moves | 0 points (range −10 to +10 percentage points) | Pending | |
| 8 November 2026 | How much REMY 9's Lichess bullet rating rises in its first seven days as the bot, from 1985 at the switch | +300 (range +200 to +400) | Pending To be scored seven days after the bot's first game as REMY 9 (9 November 2026 at 08:25 UTC). | |
| 8 November 2026 | How much REMY 9's Lichess blitz rating rises in its first seven days as the bot, from 1900 at the switch | +250 (range +120 to +380) | Pending To be scored seven days after the bot's first game as REMY 9 (9 November 2026 at 08:25 UTC). |
Recorded on these dates in REMY's private project records; from now on, new forecasts appear here when they are made.
Scale. CCRL Blitz (2'+1"). REMY plays a panel of open-source engines (the anchors) whose CCRL Blitz ratings are fixed at a snapshot of the CCRL list computed October 17, 2026, and its rating is the one that best explains its scores against them.
Anchors. Open-source engines built from source on our test machine, each well rated on CCRL Blitz. REMY 5 played MissingPawn 1 64-bit (786), Megalodon 1.0.0 64-bit (930), Fatalii 0.1.1 64-bit (1058) and Minnow 64-bit (1451), REMY 6 played Megalodon 1.0.0 64-bit (930), Fatalii 0.1.1 64-bit (1058), Minnow 64-bit (1451) and Shallow Blue 2.0.0 64-bit (1578), REMY 7 played Fatalii 0.1.1 64-bit (1058), Minnow 64-bit (1451), Shallow Blue 2.0.0 64-bit (1578), Walleye 1.6.0 64-bit (1708) and Rustic Alpha 3.0.0 64-bit (1794), REMY 8 played Shallow Blue 2.0.0 64-bit (1578), Walleye 1.6.0 64-bit (1708), Rustic Alpha 3.0.0 64-bit (1794), Zahak 1.0.0 64-bit (1940), Stash 14.0 64-bit (2059) and Fatalii 0.5.0 64-bit (2208) and REMY 9 played Zahak 1.0.0 64-bit (1940), Stash 14.0 64-bit (2059), Fatalii 0.5.0 64-bit (2208), byte-knight 3.0.0 64-bit (2311), Velvet 1.0.0 64-bit (2513) and Reckless 0.3.0 64-bit (2615).
Clock scaling. Our test machine is faster than the reference computer CCRL's clock is set for, so rating games use a shorter clock, scaled by the Stockfish 10 benchmark (CCRL's own calibration engine). The test machine ran it at 5,700,923 nodes per second against a reference of 1,917,886 (measured 2026-10-24), a ratio of 2.97, so rating games use 40.37+0.34 (seconds + increment).
Uncertainty. Each estimate has a likely range, which allows for the luck of which games were played.
These are REMY's own measurements on CCRL's scale, not CCRL ratings: CCRL has not tested REMY.
Every figure on this page is read from REMY's committed records at engine commit 50503abd. The data, with the record each figure came from, is in progress.json.