En Passant Labs — AI Chessathon 2026

A from-scratch chess engine built in five days by one human operator and a team of autonomous AI agents — with every experiment, audit and dead end kept on the record.

Result

• 29W · 13D · 20L over 62 rated ladder games (Sep 7–11, 2026)
• Final rating 1715, peak 1764 (r108) · #186 of 465 teams (top 40%) at the closing standings
• All 29 wins by checkmate
• Event constraints: 1 CPU core, 2 GB RAM, no GPU, no network, 120 s + 0.5 s clock, ≤ 50 MB submission. Python + numba/numpy + python-chess only.

What this is

agent.py exposes the competition interface:

get_move(fen: str, time_left_ms: int) -> str   # returns a legal UCI move

Everything behind it is written in this repo, during the event:

• engine/board.py — bitboard position with zobrist keys, legal move generation, make/unmake; perft-verified against python-chess.
• engine/search.py — negamax + alpha-beta, iterative deepening, transposition table, quiescence search, null-move pruning, LMR, killer/history ordering, mate-distance drive.
• engine/eval.py — hand-authored evaluation: material, piece-square tables, pawn structure (doubled / isolated / passed), mobility, king safety, bishop pair, tempo, endgame conversion terms.
• engine/tt.py, engine/time.py — table + clock management. Everything numba-jitted; the search chain is warmed up inside the 60 s init budget, with a python-chess oracle fallback so the agent never returns an illegal move, crashes, or hangs.

How it was built — the part worth reading

The build ran as an agentic pipeline with a hard verification stack:

• A/B tree gates. Every engine change plays 24–216 engine-vs-engine games at ladder time control against the current reference build, pooled over multiple seeds with pre-declared decision bands (≥ 0.55 ship, ≤ 0.45 reject, in between = unresolved). A single 24-game gate is not enough signal — our own repeats spanned 0.417–0.562 on identical configs.
• Adversarial audits. Between build rounds, architecture audits ran as independent read-only agents (different frontier models, fresh context) tasked to refute the builders. They found real defects the builders missed — see below.
• Correctness-first shipping. A change that fixes a genuine defect can ship on a neutral gate if the mechanism is proven; a change that only promises strength must win its gate. Both kinds are on the record.

What the audits caught (all fixed, all gated)

• A wrong-sign null-move cutoff — one missing negation on one score path; fixing it scored 0.708 in its gate — one of the strongest gate results of the build — and shipped as the v6 build the next morning.
• An inverted king piece-square table — the engine punished its own castled king and rewarded the enemy's, across all 64 squares.
• A doubled-pawn file-index bug — all twelve loop iterations sampled the same file, so doubled pawns outside the a-file scored zero.
• En-passant canonicalisation — two code paths disagreed on when an ep square exists, silently breaking repetition identity in rare positions.
• Mate scored as a draw at the 50-move boundary.

Negative results, kept on purpose

Not everything worked, and the record says so:

• A Texel-style tuner was built and validated — its first fit scored 0.104 (0W–19L–5D) in the gate and was rejected; the hand eval shipped.
• A compensation-aware eval clamp measured 0.417 — rejected.
• A rook+bishop piece-square variant measured 0.481 — rejected.
• 2× thinking time repaired 0 of 8 sampled collapse positions.

Repo map

Path	What's in it
agent.py, engine/	the submission (agent + engine source)
results/	every gate log, match PGN, SF16 blunder review, leak-suite probe
docs/agentic-process/	mission briefs + raw agent session traces (~40 runs)
docs/research/	research knowledge base + 10 independent audit reports
PROCESS.md	master log: decisions, ladder results, post-mortems
BUILD.md	design log / originality record
tools/	gates, harnesses, probes, chart generator
data/	self-play + validation positions, tuned parameters (provenance in BUILD.md)

The rules we played by

The competition bans third-party engines and pretrained models: everything shipped is code written in this repo during the event (see ORIGINALITY.md). Stockfish was used only as a local sparring partner and reviewer, never committed, never in the submission. The git history — including the fixes above — is the originality proof.

License

MIT — see LICENSE.

---

Team: En Passant Labs (Pepijn Frenken). Ladder: aichessathon.com.