A game for the first year of a PhD. You make the thirteen decisions a study is made of; each one names the pitfall it risks, what it costs, and the fix. Then you judge other people's plans and see whether your confidence matches your accuracy: that is taste, and it is trainable.
The pitfalls are well documented, but which option is “sound” and which plan is “stronger” is partly one researcher's judgment. Where the game disagrees with you, the reveal gives its reasons; you are allowed to disagree back.
The pitfalls are universal; the wording isn't. Machine learning is the main focus; the other domains reword the same traps and add a few of their own.
Runs in your browser; nothing is sent anywhere. Progress stays on this device. Dotted terms have a one-line explanation on hover; every card also lists its terms underneath, tap-friendly.
The outcome is what you're judged on; the process is your biggest lever.
The verdict is the game's reading of your choices, not a grade. Some “defensible but costly” calls are the right call in a real project; the reveal says why the game scored them as it did.
Rigor after each decision. Color = the call you made. Hover a dot for the card.
Open one to see what it is, how to notice it next time, and what to read.
Two different skills. Most students struggle with the first.
The answers are one researcher's judgment, and taste is partly subjective. Each reveal names the feature and the principle behind its pick so you can argue with it; the preference items have no right answer at all.
Taste is the judging half of a generative adversarial network (GAN), the discriminator: it gets sharper by judging many examples and being told the answer. The answers it is told here are one researcher's judgment; the calibration numbers measure agreement with that judgment, not truth.
Each dot: how often you were right when you said that confidence. On the diagonal = calibrated. Below = overconfident.
Descriptive, not graded. Two sources feed it: the preference items, where both sides are sound and your pick is the only signal, and the graded items, where the game's answer sits at the right-hand pole, so a miss records the instinct that pulled you left. Hollow mark: this round. Filled mark: all rounds on this device. Each pole has a virtue and a vice; neither is the goal.
The thing that separated strong from weak. Open one for the argument.
Five questions a supervisor, a reviewer and a funder all ask, with anchored scales. Score a sample idea first and compare with the student who proposed it and two reviewers, or score your own and get a one-paragraph scope statement.
Sample ideas for your domain first, then the others.
Every pitfall in the game, by phase. Open a tile for what it is, what it smells like, and the fix. Dots show what your runs on this device hit (red) or walked past (green).
Every term the game uses, one line each, in nine groups. Open a group, or type to search across all of them.
Everything linked in the game, in seven groups by what it helps with. P2P is The Path to PhD, the newsletter this game grew out of.