Can you surprise the oracle?

Source Code changes the object of choice. You never press Cooperate or Defect. You submit a program, every opponent reads it, and the two programs choose simultaneously in a one-shot Prisoner’s Dilemma. This is Tennenholtz’s program equilibrium: mutual transparency creates cooperative equilibria that ordinary action-by-action reasoning cannot express.

CliqueBot cooperates only with a byte-identical copy. It is safe and cooperative with itself, but one harmless source-code change destroys cooperation. The simulator implements Oesterheld’s ε-grounded FairBot: with probability ε it grounds by cooperating; otherwise it runs its opponent against itself and copies the answer. Two such programs recurse until one grounds, then cooperation travels back up the call stack—even when their source text differs.

The league makes bounded computation part of the game. A small ε reduces accidental cooperation with an exploiter but produces deeper, more expensive recursion. This browser version follows the simulation-based approach developed further in Cooper, Oesterheld, and Conitzer’s AAAI 2025 work. Every match is independent and one-shot; this is not an iterated Prisoner’s Dilemma wearing new clothes.

Money Pump is an executable decision-theory problem. Two boxes cost 1 credit. Each holds either 0 or 3. Before you act, the oracle predicts buy F, buy D, or walk, then puts 3 in every box it predicted you would not buy. It commits to that entire state before showing the offer.

If the oracle’s prediction is right when you buy, you lose 1; if it is wrong, you gain 2. That makes 66⅔% its break-even accuracy. Walking guarantees zero. The puzzle is that the contents are already fixed, yet your choice is evidence about the forecast that fixed them. The mode adapts Oesterheld and Conitzer’s Adversarial Offer, designed to turn Newcomb-style disagreement into an exploitable market.

A behavioral study of Newcomb’s problem reported opposite majorities under closely related framings: 85% one-boxing in one experiment and 66% two-boxing in another. Money Pump removes the imaginary perfect predictor. You face the actual browser model, its actual accuracy, and an auditable record of what it forecast.

Rebellion turns “I’ll just do the opposite” into the challenge. Hard Mode supplies the prophecy you can see. A separate 21-model reaction ensemble predicts whether you will obey or defy it. That second forecast is hidden and scored.

Before each choice, the hidden forecast and its confidence are locked behind a SHA-256 commitment with a fresh 128-bit nonce. Forecasts open eight at a time; the revealed payloads are hashed again in your browser so the oracle cannot rewrite its answer after seeing your move.

Hard combines context-tree weighting, online logistic models, long-pattern matching, and specialists for rhythms, runs, balancing, and reactions to the oracle. A weighted-majority layer learns which predictors understand you, at three different time horizons.

Its cold-start knowledge comes from five label-symmetric population models covering 387,612 transitions in an open human-randomness study. It also learns from your pauses, but never your pointer position or the key you are about to press. The forecast is computed before your choice is compared or learned.

Classic is the original five-move lookup table. It remembers what followed each exact sequence and chooses the more common result.

Accuracy is scored after 20 warm-up predictions. A genuinely random source is impossible to predict better than 50% in the long run.

Transparent-agent background: robust cooperation via provability logic and simulation-based program equilibrium. Decision-theory background: functional decision theory and an experiment on ordinary people’s Newcomb choices. Prediction machinery: context-tree weighting, online expert weighting, and human sequence data. Inspired by the Aaronson Oracle. Everything runs in your browser; your play is not stored or sent anywhere.