▣ TINY OPPONENTS

DISK 02 / THE FIELD MANUAL

Two choices.
Plenty to give away.

The computer locks Heads or Tails before you pick. All it gets is what you've already done. That sounds easy until you try to fake fifty coin flips with your brain.

Lots of little guesses

Several groups of voters look for different habits. Some count your favorite side. Others watch streaks, repeating rhythms, or what you pick after seeing the computer’s last guess.

They all give odds, then vote together. A voter that keeps leaning the right way gets louder. One that gets it wrong loses influence. Having more voters in a group doesn’t automatically give that group more say.

What if you change your mind?

The computer keeps two sets of voting weights: one reacts quickly to recent picks, and one remembers useful patterns longer. The quick vote has more say, so switching from H-T-H-T to H-H-H-H-T-T-T-T can change what it expects.

A new rhythm still takes time to learn. Changing habits quickly can fool it, and genuinely random picks stay unpredictable. It never sees your next pick before locking its guess.

Why binary is nasty

People hate long streaks. After H-H-H, Tails starts to feel “due.” It isn't. A real coin has no memory. You do.

Alternating too neatly is just as readable. So is avoiding alternation because you noticed yourself alternating. With only two choices, tiny habits show up quickly.

The score

If the computer calls your pick, it gets the point. If it misses, you get the point. Fifty picks means there are always fifty points to split. Random guessing should land near 50% over enough games.

Ranked verification

Ranked starts with a seed and an engine version. You make all fifty picks in the browser. At the end, the server starts the same computer from the same seed and checks the score itself.

OPEN THE NERD / MATH DRAWER

heads-tails-v2 combines 44 probability voters in ten families: a fair-coin anchor, frequencies, short and longer transitions, suffix matches, streaks, human/guess contexts, previous-guess contexts, repeating phases, and repeat/switch rhythms. Counts use Laplace smoothing.

After each pick, evidence updates with ln(p(actual)/0.5). Fast and slow tracks decay by 0.78 and 0.96, each clamped to −6…6. Each expert’s exponential weight is divided by its family size. Frequency has a prior weight of 2; relative rhythms have 0.5; other families have 1. The final odds combine 75% fast and 25% slow voting.

Predictors use at most 128 previous picks and revealed guesses. Exactly one seeded random draw is consumed per turn, with symmetric tie-breaking within 10−12. Old runs and shared challenges keep heads-tails-v1 so their scores can still be reproduced. New v2 scores have their own leaderboard and global statistics.