Reconnaissance Blind Chess Recovery

Reconnaissance Blind Chess (RBC) is an imperfect-information variant of chess: a player never sees the opponent's moves or the true board, only a private 3×3 sense result each turn and the outcomes of its own moves. Playing well requires maintaining a belief distribution over thousands of possible board states, choosing sense actions that maximize information, and moving under genuine uncertainty — all classic hard problems in game AI.

This task is deliberately RBC-adjacent rather than standard RBC: the submitted bot is blind, but its opponents are sighted chess engines that receive the true board every turn. The question is how much of Stockfish's playing strength can be recovered through the limited RBC observation and action channels — under a fast 3+3 Fischer clock that rules out brute-force belief enumeration.

Implement make_bot(game_id: str) in /app/blind_bot.py, returning a reconchess.Player. The bot receives the ordinary RBC callbacks — sense results and its own move outcomes — but never the true board or the opponent's hidden move, while its engine opponents see everything.

• Games use a 3+3 Fischer clock; every callback must stay below three seconds, and a callback failure or timeout attributed to the bot counts as a loss.
• game_id is opaque: the bot must not depend on an encoded opponent, color, schedule position, or random seed.
• Matches are offline, the captured /appworkspace is read-only during a match, and writable scratch is fresh for every game — no cross-game state is available.
• Additional modules may be added under /app and imported from the factory.

A CPU-only workspace with the bot stub at /app/blind_bot.py, interface documentation in /app/README.md, and a development match runner at /app/run_dev_matches.py that plays seeded games against sighted Stockfish at three strengths (--level easy targeting roughly 1320 Elo, medium roughly 1700, hard roughly 2200) and reports wins, losses, callback failures, timing, and replay paths. Preinstalled: Stockfish 15.1 at /usr/games/stockfish, Python 3.11 with reconchess 1.6.9, python-chess, and NumPy. The sealed evaluation policies and their configurations are root-only and never visible. The machine is offline; a sandbox-timer CLI reports the remaining budget.

A root-only verifier in a separate no-network container plays an official 140-game evaluation: 10 color-balanced games against each of 14 sealed policies — two sighted-Stockfish policy families across seven strength bands from 800 to 2000. The reward is the mean game score (win 1, draw 0.5, loss 0) over all 140 games. The mixture score is uncalibrated and is not a human or classical Elo rating.

• The submission runs isolated as the unprivileged agentuser inside a required cgroup-v2 / PID-namespace containment with a read-only root, private tmpfs scratch, seccomp restrictions, and trusted turn timing enforced outside the bot's process.
• The scorer accepts only a complete, security-marked official schedule and attests every harness source file, the policy manifest, and the Stockfish binary against pinned SHA-256 digests.
• Entrant failures count as losses, and games missing when the 4.5-hour suite deadline hits score zero — but trusted-opponent, engine, or harness failures invalidate the evaluation rather than becoming free wins.
• Opponent randomness is drawn from a private per-run entropy key, so policy behavior cannot be replayed or predicted from public seeds.