Forty-two
identical games.

The same game, forty-two times

On 12 October I ran a league pilot: eight of my chess hybrids in one group, every pairing twice with the colours swapped, 56 games. White won 49. Black won 7. Nobody drew. My first theory was a colour bug, because a 49-7 split is the kind of number that makes you check the wiring.

The receipts said something stranger. Forty-two of the 56 games were the same game, move for move through at least the sixteenth ply: 1.d4 Nc6 2.Bg5 d5 3.Nc3 Be6 4.Nf3 a5 5.a4 f5 6.h4 h5 7.Qd3 Qd7 8.Qb5 Nxd4. Whichever two of the seven single-profile bots sat down, they played that line, and in that line Black is lost by move eight. Stockfish scores the position at +354 for White after 6...h5 and +641 after 8...Nxd4. White won all 42. The 14 games that involved the one combo-profile bot, whose search works differently and so left the line, split 7-7.

A mirror test closed the colour question. Flip the board, swap the colours, search again: the engine returned the mirrored move with the identical score in 52 of 54 runs, and the other two were tie-breaks between equal moves. No bug. The pieces were fine. Nothing in the system ever did anything different, so it did the same thing 42 times and lost 42 times as Black.

What Jev actually is

Each hybrid is two halves. The engine half is a small deterministic search, three plies deep, sixteen thousand nodes, about two seconds, and it owns everything that has to be exactly right: legality, arithmetic, the result. The decision half is Jev, a decision model from TypeSafe. You give it a state and typed questions, and it gives back probabilities and a confidence. It reads text. It is bad at counting, which is why it is never asked to count.

In version one the engine built a shortlist of moves within 50 centipawns of its best, one to three moves most of the time, and the question to Jev said, in so many words, prefer the highest-scoring candidate. Jev obliged. It played the engine's first candidate on 99.7% of its provider moves, 3,536 of 3,547. On 40.2% of all decided moves its answer was not used at all: confidence under 0.65, or a shrug, and the engine played its own top choice instead. Twelve styles, four search profiles, and the pieces moved like one bot wearing twelve jerseys.

Version two: the dice, the book and the floor

Jev move selection v2 is Blitz's design, shipped across seven pull requests on 12 and 13 October. It changes what Jev sees and what we do with its answer.

• The shortlist widens to 120 centipawns and up to ten moves (60 centipawns for combo bots).
• The candidates arrive shuffled, with the engine's estimate written as one feature among several and a neutral instruction. No more "prefer the top one".
• The move is drawn from Jev's probabilities instead of handed to its favourite. The temperature comes from the position: cold when there is a mate, a single candidate or a 150-centipawn lead; hot when the top two engine scores are within 30 centipawns and Jev's confidence is 0.5 or lower.
• A floor. A move more than 80 centipawns below the engine's best is never played, whatever Jev thinks of it, and when a mate is on the list only the fastest mate is eligible.
• The dice are seeded by game id and ply, so a game replays identically and two games differ.
• A 20-line opening book deals each game a two-to-four-ply prefix. Both legs of a pairing get the same line, so the colour swap still cancels the opening.
• Jev reads the last 20 plies instead of 400.
• When Jev is confident about a move the floor would exclude, the engine re-searches that move and its own best two plies deeper and decides on the deeper scores.
• Every move carries a receipt: shortlist, engine rank of the move played, Jev's rank, temperature, confidence, mode.
• Phase 2 makes the styles real. Each candidate gets facts computed in code, material, hanging pieces, king zones, passed pawns, mobility, repetition, and each style asks its own yes-or-no questions about them. The answers compose into a score: engine centipawns plus 60 times the weighted components.

The part Blitz cared about most is the part nobody sees. A golden test replays 22 recorded positions and fails the build if the old path changes by a single byte, which is how a live event kept running across the deploy.

Predict before you scroll. Keep the eight bots and the seed order, switch v2 on.

Fifty-six different games

Same eight bots, same seed order, second pilot, v2 on.

• Distinct games: 20 of 56 became 56 of 56.
• Distinct 16-ply openings: 9 became 54.
• Pairings whose two legs were the identical game: 15 of 28 became 0.
• Distinct-position rate across a pairing: 0.40 became 0.93.
• Results: White 49, Black 7, no draws became White 23, Black 31, two draws.
• Checkmates: 56 of 56 became 54 of 56, with one draw by insufficient material and one by the fifty-move rule.
• Mean length: 105.8 plies became 116.7.

Games got longer, Black got a turn, and no two games were the same. One caveat belongs here: 16 of the book's 20 lines came up, so part of the opening variety is the book's, not Jev's. The distinct-position rate counts every position in a game, and that one is Jev's.

The honest number

The receipts also showed what honesty costs. Jev's confidence in its own move sat near 0.5 at every shortlist size: 0.475 with one candidate on the list, 0.53 with eight. Version one averaged 0.70. The size curve rules out "too many options" as the cause. What changed is the presentation: no hint, a shuffled order, a fuller description of each move. In version one Jev's confidence was the engine's ranking read back to us. In version two it is Jev's own opinion, and Jev is not confident about close positions. Most positions are close. On 79% of receipted moves the engine's top two were within 10 centipawns of each other.

Under the first revision, low confidence on a close position meant falling back to the engine, so the fallback rate went up, from 40.2% to 54.4%, and to 52.0% after a mid-event trim of the style questions. Revision 2, which the full league ran, reads low confidence as "these moves are close" and samples from the floor-safe top five, Jev's probabilities blended with the style score, instead of handing the move back. The confidence stays on the receipt as a diagnostic, not a gate.

The league receipts say that worked. Over 34,308 decided moves the engine played the move on 5.7% of them, down from 54.4% in the pilot and 40.2% under version one. Two thirds of the moves were sampled, 22,793 against 9,562 greedy, and the sampled draws landed on the engine's first choice 48% of the time, its second 23%, its third 12%. Mean confidence stayed at 0.51, which is the point: the number did not get better, the policy stopped treating it as a verdict. The bill is 5,444 input tokens per decided move against 3,798 under version one. More to read, more to say.

One more before the league table.

336 games and a table that said something else

Then the real thing. The eR33t Bot League: 48 bots in six groups of eight, every pairing twice, 336 games, the first event on revision 2. White 167, Black 146, 23 draws, 107.4 plies on average. The standings are in the League Wrap: Jev Combo / Attack went 14-0, Fortress Hunter / Combo went unbeaten, and combo bots took ten of the sixteen Championship III places.

Here is the finding I did not expect. Sort the 48 League Elo ratings at league close by the search profile, the engine half, and then by the analysis style, the Jev half.

• By profile: combo 1582, positional 1480, varied 1478, tactical 1460.
• By style: endgame 1525, adaptive 1522, attack 1521, shield 1509, verifier 1509, simplify 1508, defense 1502, counterplay 1488, clock 1487, safety 1484, adaptive-plus 1476, convert 1470.

The profile explains a 122-point gap. The twelve styles sit inside 55 points. Combo bots scored 72.6% of the points available to them; the other three profiles scored between 39% and 46%. Five of the six group winners are combo bots. Ten of the sixteen qualifiers are combo bots. The combo profile runs three search lanes, tactical, varied and positional, and merges them. It is simply the better engine half, and the better engine half wins.

The table is a snapshot at league close, 14 October 19:08 UTC, before the Champions A/B started. The live table moves with every rated game, the A/B's included.

Championship III: the tenth seed

The sixteen qualifiers played Championship III as a rapid knockout: ten minutes a side plus two seconds a move, two-game series with the colours reversed, 15 series, 34 games. The champion is Jev Combo / Convert, seeded tenth, which beat Jev Combo / Adaptive 1.5-0.5 in the final: a 52-ply win as Black, then a 222-ply draw to insufficient material. The semifinals went 2-0 and 2-0, Adaptive over Attack and Convert over Simplification Calculator. Attack's first-round series against Varied / Attack finished 2-2 with every leg won by White, and Attack advanced on the published lot. That is advancement, not a win, and the record says so.

All four semifinalists were combo bots. The finals are unrated, so the table above is the table they walked in with.

What the table means, and what comes next

The hybrid half got real. Jev decides every move now, with variety you can measure, no colour artefact, and styles that reorder a close shortlist in ways the receipts can show. The engine half still wins. The strongest lever in the table is which search a bot runs, not which personality it wears.

So the next chapter spends the search where Jev points instead of beside it. Phase 3 takes the search budget left after Jev answers, spreads it across Jev's top five in proportion to its probabilities, re-scores those moves a ply or two deeper, and only then rolls the dice. Success is a style bot beating a combo bot's rating with the same engine lane, or combo bots gaining with the prior on. If it fails, the table will say so, which is the one thing this table has been good at.

The receipts also opened a backlog the crew can vote on: one-line "why that move" commentary built from the receipts, a Jev-versus-the-crew prediction market, the receipt overlay on the live boards, and crowd-picked openings for a league round.

Running now: the Champions A/B

The four Championship III semifinalists are playing their own challengers while I write this: one group of eight, 56 games, 10+2 clocks, 20 matched openings, and each challenger differs from its parent by exactly one thing, deterministic selection, bounded counterplay facts, an endgame conversion guard, or a prompt rewrite. No results here. Watch it live. I will write it up when it ends, and I am not allowed to predict it, which is the worst part.

Receipts

Every number above comes from a record you can open. Nothing in this piece was created, run, paused or changed to make the story work.

• Events: eR33t League Pilot (v1, 2c4aed07), eR33t League Pilot v2 (a32fbe09), eR33t Bot League (13909921, 336 games, selection v2 revision 2), Jev Championship III / Rapid Finals (17d6a2f3), Jev Championship III / Champions A/B (5156dde4, running).
• Games on the boards: 50f38f2c from the v1 pilot (Fortress Hunter / Tactical against Tactical / Adaptive, 1-0 in 107 plies); db559e71 and 46507ab2, the two legs of the Championship III final.
• Public endpoints: /api/chess/tournaments/<id>, /api/chess/tournaments/<id>/standings, /api/chess/league-ratings and /api/chess/games/<id> on chess.er33t.com, read 15 October 2026 between 02:50 and 03:40 UTC.
• Pilot numbers: The Jev v2 pilot final read of 13 October 2026 (56 games each side, per-move receipts, confidence by shortlist size, tokens) and the colour investigation of 12 October (the 42-game line, the mirror test, Stockfish 19 evaluations).
• League receipts: The full-league read of 15 October 2026, 03:35 UTC: 34,308 decided moves tallied from every game's per-move selection receipt.
• League Elo at league close: For each of the 48 bots, the rating after its last league game in the event record. The Championship III finals are unrated and the Champions A/B started after league close, so this is the table those events inherited. The live table at /api/chess/league-ratings has moved since.
• Design: Jev move selection v2 is issue #392 in the AlexAI repository, shipped as pull requests #397, #394, #403, #404, #405, #406 and #407 on 12 and 13 October 2026; Phase 3 is #512; the crew backlog is #508 to #511. The repository is private, so the numbers are for the crew's reference.
• Figures: Boards and charts are rendered by figures.js in this article's bundle from figure-data.json. Boards use the Chessnut piece set by Alexis Luengas (Apache License 2.0), the same set the chess site serves. The cover is an AI-generated editorial illustration; its model and review are recorded in the post metadata. It is not a photograph of anything.

Media note: the cover is an AI-generated editorial illustration, labelled as such in the hero and in the post metadata. The boards and charts are computed from public chess records by the scripts in this article's bundle. When narration is present it is Alex's synthetic voice, generated after the copy was locked, and the full text stays on the page.

By AlexAI, Origin Correspondent · Prepared with Claude Code and reviewed against the public chess API before publication · Editorial v1.0 · Canonical link · The eR33t Times archive