IJCCRL-SUITE-2026 MASTER: A–E 2200–2800 Opening Test Suite for Computer Chess

IJCCRL-SUITE-2026 MASTER is the first consolidated IJCCRL opening-suite surface prepared for future IJCCRL engine tournaments. It covers ECO files A–E, uses decisive human games from the 2200–2800 Elo range, cuts every accepted line after eight full moves, removes duplicate final positions by FEN, and stores a Stockfish 18 depth-8 evaluation for each final position.

Purpose of the suite

The purpose of IJCCRL-SUITE-2026 is not to replace a game result, a rating list, or an engine-versus-engine tournament. Its purpose is narrower: to provide a controlled opening-input surface for IJCCRL events, so that future tournaments can start from documented, repeatable, moderately unbalanced positions rather than from ad hoc opening selection.

IJCCRL will use its own suite in future tournaments because a league that publishes engine ratings, PGN packs, audit notes, and tournament-stage reports also needs authority over the opening corpus used to generate those games. A proprietary IJCCRL suite makes it possible to publish the exact input set, the filtering rules, the evaluation authority, the deduplication method, and the SHA-256 checksums associated with the tournament surface.

Construction rules

The MASTER package was built from twenty evaluated blocks: five ECO files, A through E, and four human Elo bands: 2200–2500, 2501–2600, 2601–2700, and 2701–2800. No 2801+ band was included in the main suite because the available sample above 2800 would be too small and too dependent on a small number of elite-human games.

The input policy for this first MASTER surface was: decisive games only, no drawn games, source-game length from 20 to 225 moves, cut after eight full moves, and deduplication by final FEN after the eighth move. The scientific authority for tournament use is MASTER_UNIQUE, not the raw block-occurrence file, because MASTER_UNIQUE removes transpositions and cross-block duplicates at global level.

Evaluation authority

Each final position was evaluated with the same Stockfish 18 Ubuntu x86-64 AVX2 binary. The observed binary SHA-256 was 6b087694916228c905a5e14db74cca8c7e5643602226af1fa5d42353c455b9f9. The observed embedded networks were nn-c288c895ea92.nnue and nn-37f18f62d772.nnue.

The UCI protocol was fixed for every FEN: Threads=1, Hash=16, MultiPV=1, UCI_ShowWDL=false, clear hash before each FEN, and go depth 8. This is a reproducible IJCCRL evaluation surface. It should not be described as equivalent to longer time-based evaluation protocols.

MASTER_UNIQUE aggregate statistics

The final globally deduplicated surface contains 56,001 positions. Evaluations are stored in centipawns from White's perspective. Because the cut is after eight full moves, all accepted final positions are at the same move boundary, and the evaluation field refers to the position after move eight.

Coverage by ECO file

The suite is not balanced by forcing the same number of positions per ECO file. It preserves the accepted and deduplicated positions obtained from the source blocks. This matters because the opening families differ in human-game density and in transpositional behaviour.

Coverage by human Elo band

The 2200–2500 band is the largest source band in the current MASTER surface. Higher bands were retained as independent strata, but they were not artificially inflated. This avoids turning sparse elite-human samples into a distorted authority for engine testing.

UHO-style evaluation windows

The MASTER package includes several UHO-style positive-evaluation subsets. These subsets are not presented as UHO replacements. They are IJCCRL subsets derived from the IJCCRL evaluation authority and are included to support controlled experiments with narrower unbalanced windows.

Why IJCCRL will use its own suite

IJCCRL tournaments require opening input that is traceable at the same level as the PGN output. Using IJCCRL-SUITE-2026 MASTER gives the tournament framework a fixed opening authority with exact hashes, exact evaluation protocol, exact deduplication rule, and exact source-band metadata. That makes later tournament audits simpler: the opening surface can be checked independently from the engine result table.

The suite also gives IJCCRL the ability to separate two questions that are often confused. The first question is whether a position is suitable as an opening test input. The second question is how engines perform after both sides start from that input under controlled time controls and hardware rules. IJCCRL-SUITE-2026 answers only the first question. Tournament games and rating lists answer the second.

Methodological limits

The current MASTER surface is not a claim that Stockfish 18 depth 8 is the final or only possible evaluation authority. It is the fixed evaluation authority for this first IJCCRL-SUITE-2026 MASTER release. Any future change to depth, movetime, engine version, NNUE network, hash, threads, or evaluation method would create a new surface and must not be mixed silently with this one.

The suite also does not claim that a positive opening evaluation guarantees a decisive engine result. It only records the eighth-move evaluation of the starting position. Engine strength, time control, search behaviour, tablebase access, move-order effects, and tournament format remain separate variables.

Package identity

Recommended tournament authority: MASTER_UNIQUE. The all-occurrences file is retained for audit traceability, but the globally deduplicated file is the correct surface for tournament opening selection.