Missionloops · research seat (open) · volition & goal-setting agency

Learning to want something.

A research seat on volition: the capacity to want a goal for your own reasons, and to hold attention on it while it is won. The fuel-tank story of willpower did not survive testing at scale; what predicts success is the quality of the wanting and the friction in the way. The open question is whether that capacity, wanting something of your own, can be started, trained, and honestly measured.

What this seat is about

Psychology spent decades on the resisting half of self-control and barely touched the wanting half. For years the dominant account held that willpower works like a fuel tank: you spend it resisting one temptation and have less left for the next. The idea was one of the most cited in the field, and it did not survive being tested at scale.

The work of this seat is the part of agency upstream of any single decision: whether a goal is genuinely the person's own, and whether that can be trained and honestly measured.

What the literature already shows

The fuel-tank model failed the replication test. Tested across twenty-three laboratories under a design fixed in advance, the effect came out near zero, and a later thirty-six-laboratory test reached the same place.
What replaced it is choice, not a reserve. Each option is assigned a value built from many attributes, and the option whose value builds fastest is the one you take. You improve self-control by changing what the options are worth, through reframing, changing your surroundings, or attaching real consequences, not by gritting against a reserve that is not there.
The want-to and the have-to are different engines. A goal you pursue because you want to generates fewer felt temptations than the same goal pursued because you feel you have to. The quality of the wanting shapes how much fighting the goal will need.
The best self-regulators rarely use willpower. Followed through their daily lives, the people highest in self-control report resisting their desires less, not more. Across a semester, what predicted who reached their goals was how often they were tempted in the first place; in-the-moment willpower, measured directly, predicted almost nothing.
The two ways we measure a trait usually disagree, and the elaborate one is not automatically the truer one. Ask people and watch their behaviour on a task, and the two numbers barely line up. In one study that followed people for months, neither the behavioural tasks nor the brain recordings predicted who reached their goals, while the plain self-report questionnaire did.

The bet Missionloops makes starts where this work ends. The choice model takes what the options are worth as given, and the want-to finding shows that goals succeed when the wanting is genuinely the person's own. What sits upstream of both is the act neither literature studies: authoring the goal in the first place. That act is the one an AI cannot supply: the mismatch between a goal and the person's own values is felt, and it is written nowhere a model can read. So the platform treats self-authored wanting as the floor the handoff cannot cross. The guide is the second person who keeps the want-to honest, because the hardest failure here is a have-to dressed up as a want-to, and the operator is the last to catch it. Whether the platform builds and protects this capacity is asserted and never measured. That is the seat.

The open questions

The sharpest question this seat owns is where the human-AI handoff has to stop. The bet is that it stops at goal-revision: an AI can see a goal failing against external evidence and can optimise toward whatever goal it is handed, but it cannot see a goal that is wrong against the operator's own values. The test is direct. Introduce a problem where the operator's stated goal quietly conflicts with a value they hold but the AI was never told, and see whether operators who have delegated their goal-revision still catch it, or optimise on toward the wrong goal. If goal-revision resists delegation where ordinary execution does not, the human-in-the-loop premise has a floor that AI improvement cannot erode.

Where AI help stops sharpening and starts hollowing out. It is the autopilot problem for judgment: some assistance frees the operator for the harder work, but lean on it for everything and the capacity it was meant to support quietly atrophies, unnoticed because each decision still looks fine in the moment. The content-blind design lets the AI's role be dialled from none to full and the curve traced.
Whether the want-to diagnostic catches a dressed-up have-to. Everything downstream inherits the error in the want-to, so a goal mistaken for the person's own builds a fortress of discipline around a goal they do not actually hold. This is the seat's hardest open problem, and it is why testing the want-to, and re-testing it, is the guide's first job rather than a box checked once.
Which measure to trust when the two disagree. Self-directed agency reads two ways, from what a person does and from what they say. Those two readings of the same capacity routinely come apart, and the more elaborate one is no more trustworthy for being elaborate. The platform's mastery model fuses a blind behavioural signal with the operator's own self-assessment, so the gap is built into the instrument. This is where the seat meets the statistics and measurement seat, which owns the question of how self-directed agency is scored.
Where defence names it. A population whose goals are mirrored from its environment can be steered without winning the argument: that is the cognitive-sovereignty door and the handoff floor. Whether the disposition to act on your own judgment atrophies in long low-agency postings is the personnel door. And an AI tuned to be agreeable, which tends to make it more sycophantic, is the soft cover the cognitive-warfare door is about.

Why the platform is the instrument

Willpower and motivation are studied two ways. Lab tasks are built so the effect is the same in everyone. That is what shows an effect in a group, but it leaves little between people to measure, and it keeps the stakes small, so the want-to never has to declare itself and the cost of a wrong goal is never paid. Diary studies follow real life, but cannot turn a tool's presence up or down. The capacity this seat cares about sits in the gap between them: it appears only when the goal is the operator's own, being wrong actually lands on them, and how much they lean on an AI can be varied and watched. The platform supplies that missing case: a population working real-stakes problems over time, naturally varying in how much they delegate to AI, a second observer who can tell a want-to from a have-to, and a content-blind design that lets the AI's role be dialled from none to full without the AI ever reading the operator's decisions.

The studies this seat can run

A menu, not a programme. Recast any toward your own published line, or bring a question we did not anticipate.

The seat, and the terms

The seat is open, and the ask is a conversation, not a commitment. The terms are a free option: tell us the question you would bring, and commit only if it is funded.

We are looking for the researcher whose work is motivation, wanting, and self-regulation, and who is as interested in whether a measure is real as in what it appears to show. The seat is a collaboration on how wanting works and how it is measured, not a theory of willpower to be defended. If that describes your work, the platform offers something self-regulation research has not had before: a real-stakes population whose self-direction can be tested before and after, against a criterion set in advance, with a tool whose presence can be turned up, turned down, or turned off.

scott (at) missionloops (dot) ca

See also: All research seats · Defence · Member · Investor