Rewind the Robot Train it, save every version, and roll back when the data goes bad.

1. Quick start — read this, start playing

   • Your Version Log starts with v0, the base model: every skill 1, total 4. Lay 3 Dataset cards face up and shuffle the 6 Event cards face down.
   • On your turn, take one face-up Dataset. Pick any version on your log, copy its datasets into the next row, add the new one, and write which version you built from. That's a commit.
   • Add up each skill: 1 + every dataset's number. Skills stay between 0 and 5. Refill the face-up row.
   • When everyone has played, flip one Event. If it changes a dataset, fix every version that has it. After 6 rounds, circle the version you release. Highest total wins.
   • That's the whole game. Everything below is for when a question comes up.

2. Setup

   • Each player takes a Version Log and a pencil. Row v0 is already filled in: no datasets, every skill 1.
   • Shuffle the 12 Dataset cards into a face-down pile and turn 3 face up. Shuffle the 6 Event cards face down in a second pile.
   • Pick who goes first. Play passes to the left.

3. Your turn

   • COMMIT: take one face-up Dataset card. Choose any version on your log as the parent: the newest one, or an older one if you'd rather branch.
   • Write the next version number, its parent, and every dataset in it: the parent's datasets plus the new card's code.
   • Work out each skill: 1 + the numbers from every dataset in that version. A skill can't go below 0 or above 5. Write the total.
   • OR MERGE instead of taking a card: pick two versions. The new version gets every dataset from both, counted once each. Work out its skills the same way.
   • Turn a new card face up so there are 3 again. Keep the Dataset card in front of you so everyone can see what you trained on.

4. End of each round

   • When everyone has had a turn, flip the top Event card and read it out. It's true for every player.
   • If it changes what a dataset does, write the new numbers on that card. Then fix every version on your log that has that dataset, old ones too. Cross out the old skills and write the new ones.
   • If it says a dataset can't be released, put an X next to every version that has it.
   • After 6 rounds (6 Events), the game ends.

5. How to score

   • Test Day: circle ONE version on your log to release. It can be any version without an X, even v0.
   • Your score is that version's total: Spot, Chat, Walk, Safe added up (20 is perfect). Highest score wins. Ties share the win.
   • Solo: score 15 or more, then try to beat your best with a new shuffle.

6. Tips

   • Release your best version, not your newest. In 500 practice games, picking the best version on the log beat shipping the newest one 303 times and never did worse.
   • Roll back. When an Event hurts your newest version, build on an older one instead. Doing that beat always building on the newest 336 times and lost 5.
   • A dataset that adds a lot to one skill can quietly take from another. Try it on a branch and keep the old version.
   • Every Event card is public. Count which datasets are still under suspicion before you lean on them.

7. What the game teaches

   • A model is its data. The same recipe with different datasets gives a different robot, so teams write down exactly what went into every version.
   • Problems in training data often show up later: mislabeled photos, rude chats, data nobody agreed to share. One bad dataset changes every version built with it.
   • Teams keep old versions so they can roll back. Newer isn't always better.
   • Branching lets you try a risky idea without losing a version that works. Merging combines two lines of work.
   • Reproducibility: if you can't say what's in a model, you can't fix it, and you can't prove it's safe to release.

8. Variants

   • Open book: shuffle the Event deck and lay all six face up before you start. Can you plan around every problem?
   • Short log: you can only keep 4 versions. To save a 5th, erase one. (Real storage isn't free either.)
   • Teacher mode: build one class log on the board. Teams vote on which dataset to commit and which version to build on.

Downloads

• Instruction sheet

  Plain paper · 1 page · 8.5 × 11

  One 8.5 × 11 page for the table — quick start, setup, turns, and scoring.

  Open Download

• Full rules

  Plain paper · duplex · 2 pages

  Same game, longer type — plus what the cards teach and the variants.

  Open Download

• Version Log

  Plain paper · 1 page · landscape

  Seven version rows: parent, datasets, four skills and a total. One per player.

  Open Download

• Dataset + Event cards (Avery 5388)

  Avery 5388 · 6 sheets

  12 Dataset cards to train on and 6 Event cards that change them later.

  Open Download