Say what you intend. See what you actually do. When you stall, get questions instead of answers, and find out which helps you more.
This is a research prototype. It records what you plan to do, what you then do, short mood check-ins, and what happens when you say you are stuck. When you are stuck it randomly gives you one of three responses: questions only, direct advice, or nothing but a note. Later it asks whether you started.
Your entries are saved only in this browser, on this device. Nobody else can read them, including the researcher. If you take part in the pilot, you choose whether to export your data and send it. Clearing your browser data deletes them, so export a backup now and then.
This is not therapy or medical care. If you are in crisis, contact someone you trust or a local crisis line.
A commitment device, like Odysseus asking to be tied to the mast. Use it only if you want to. Not using one is a fine choice, and a missed commitment tells you which level binds you, not what you're worth. The app records the commitment; you or your partner carry out any stake.
You can release a commitment any time, but the release only takes effect after 24 hours, so the version of you who made it gets a say.
Takes ten seconds. Do it when you notice yourself drifting, and once in the evening.
For the skin and stress sub-study. Record it once a day at about the same time. Leave this closed to skip it.
You'll get one of three responses, picked at random so the study is fair: questions only, direct advice, or a plain note. That randomness is the experiment. This open version uses a fixed, scripted coach so every participant gets the same wording.
1. Each intention-day is a trial. You predicted a chance p (say 70%, written 0.7). The outcome y is 1 if you did it, 0.5 if partly, and 0 if you pushed or dropped it. Pushing it to tomorrow closes today's trial as 0 and opens a new one.
2. Brier score, the average squared distance between prediction and outcome. 0 is perfect; always saying 50% scores 0.25.
3. A fair comparison. Imagine you had simply predicted your own overall success rate ȳ every time. That gives a baseline Brier. Skill tells you how much better your individual guesses are than that lazy rule. Above 0 means your guesses carry information, and below 0 means you'd do better ignoring your gut and using your average.
4. Calibration. Group the trials by what you predicted (low ≤30%, middle 40–60%, high ≥70%) and compare the predicted average with what actually happened. If the bar falls short of the line, you were overconfident in that range.
5. Planning ratio = actual minutes ÷ planned minutes, using the median so one marathon day doesn't dominate. Above 1 is the planning fallacy (Buehler, Griffin and Ross, 1994).
6. Feeling error = how you actually felt − how you predicted you'd feel. Negative means things felt worse than expected, positive means better. A consistently positive error on tasks you avoid means you dread them more than they deserve, which is Gilbert and Wilson's impact bias showing up in your own data.
For each kind of task the model keeps two counts. α is 1 plus the successes and β is 1 plus the misses, with a partial counting half each way. Starting both at 1 means "no idea yet, any rate is equally likely".
Every new outcome adds 1 to one count, so the estimate moves a lot early and settles as evidence builds. That is Bayesian updating in its simplest form, and it's what the "your record" hint under the intention form uses.
Your stated values are what you say matters. Your revealed values are whatever predicts what you actually do. To estimate them, the app fits a logistic regression on every trial, predicting the outcome from the importance and interest you gave the intention:
Both ratings are first standardized (subtract the mean, divide by the spread), so b₁ and b₂ are on the same scale and can be compared. The revealed share for importance is |b₁| ÷ (|b₁| + |b₂|). The fit uses gradient descent with a small penalty that stops wild numbers when data is thin.
Read it as a rough guide below 50 trials. It is correlational: maybe you rate tasks you'd do anyway as more interesting. The tables of raw rates are there so you can check the model against the plain numbers.
Lisa Feldman Barrett calls the ability to tell similar feelings apart emotional granularity. In her studies, people who make finer distinctions tend to regulate their emotions better. Marc Brackett's Mood Meter sorts feelings by the same two axes as your sliders, pleasantness and energy, then asks for a specific word. The granularity figure here is a crude stand-in: distinct words you've used divided by check-ins that have a word.
Each day gets one mood value, the average of that day's mood check-ins (−3 to +3), and one itch value, the worst itch you recorded that day (0–10). The Pearson correlation r runs from −1 to +1:
"Same day" pairs mood and itch from the same date. "Next day" pairs today's mood with tomorrow's itch. If stress feeds flares, the next-day link is the one to watch. Both are correlations within one person, not proof of cause. Sleep, weather, sweat, allergens and especially medication changes can drive both, which is why medication changes are logged. Read anything under about 14 paired days as noise.
The rates compare intentions you bound with ones you didn't. You chose which to bind, so this isn't a randomized comparison: you might bind only the tasks you were already unsure about. The trend line matters more. The long-term aim is to need the mast less, so a falling share of bound intentions with steady follow-through is the good outcome.
Some choices happen once, such as where to do a master's. You can't learn from repetition, but you can do three things: find the objectives you forgot, admit you're unsure how much each one matters, and record your gut call so you can check it later.
Take each one at the start (baseline), then as listed. Your scores are stored with your other entries and included in the JSON export.
Question. When someone is avoiding a task they chose, does being asked questions (Socratic) help them start as often as being told what to do, while leaving them feeling the decision was theirs?
Design. Within-person randomized. Each time you press "Help me start", the app assigns whichever of the three responses you have received least so far, breaking ties at random. This keeps the groups balanced even with few episodes.
Outcomes. Primary: started within 30 minutes (yes/no). Secondary: "the decision felt like mine" and "this helped" (1–7). Dependence: how often you press "Just tell me what to do" during questions.
Limits. Self-report, unblinded (you can see which response you got), and a small sample. Read results as a pilot, not proof.
Added in version 2. Each intention now carries a forecast (chance of doing it, minutes, expected feeling) that is scored against what happens. Check-ins carry a feeling word. A model compares what you say drives you with what predicts your behaviour. Commitments are opt-in and recorded, and they are not part of the randomized comparison, which stays limited to the three stuck responses.