From Evidence to Action: The Social Science Behind Spillway
Spillway is not being designed as a conventional recommender system with an inbox attached.
Its architecture starts from a different question:
What can the system justifiably infer about this person and this situation, and what follows for action right now?
That question brings social-science ideas directly into software design: latent variables, measurement error, differential accuracy, selection effects, affordances, bounded rationality, and decision theory.
The epistemic layers themselves—source artifact, observation, evidence, knowledge, decision—are defined in From Data to Knowledge. The named failure modes that arise when those distinctions are collapsed are collected in Computational Fallacies. This page focuses on what social and statistical science add once the system begins reasoning about people and action.
Categories and priorities are latent constructs
Many things Spillway cares about are not directly observable.
A category such as “Students” or “Research” is inferred from indicators such as sender identity, wording, recurring relationships, links, prior filing decisions, or model interpretation. Importance is even more relational: important to whom, for what purpose, at what time, and relative to what else?
This is a measurement problem. The software observes indicators and tries to infer constructs the user actually cares about.
That means a convenient signal should not silently become the construct itself. A quick reply can be evidence of importance, ease, habit, urgency, or simply a good opportunity to act. A long delay can reflect low value, high difficulty, interruption, lack of prerequisites, or waiting for a better device.
Differential accuracy matters more than one headline number
A classifier can be excellent for one kind of message and weak for another while maintaining an impressive average score.
The useful quantity is therefore closer to:
P(correct | user, category, evidence, context)
than one global accuracy statistic.
This has operational consequences. A cheap deterministic cue may be sufficient for one class of messages while another deserves more inference or explicit review. The same principle applies to model-generated estimates of urgency, task duration, or importance: performance can differ by construct and context.
Evidence is conditional and often dependent
Several cues that agree may still share the same cause. An AI model may rely on the same words and links already counted by deterministic features. A learned sender association may itself have been produced from earlier messages containing those cues.
The system should therefore care about marginal informativeness—what a new observation adds beyond what is already known—rather than simply counting votes.
Similarly, uncertainty should survive downstream use. A weak inference should not become certain merely because another subsystem stored it as a Boolean.
Those are examples of a broader principle: the conditions under which a measurement was produced matter to how much it should influence a decision.
Behavior is produced by motives and opportunities together
Observed behavior is not a direct readout of preference.
Suppose someone repeatedly postpones an important manuscript while quickly clearing small administrative requests. The behavior could reflect preference, but it could also reflect initiation cost, lack of uninterrupted time, cognitive bandwidth, anxiety, task difficulty, or the tools available in the moment.
A useful conceptual reminder is:
action propensity ∝ motivation × opportunity / effective difficulty
This is not a literal universal equation. Its purpose is to stop the system from collapsing several latent causes into one score.
The relevant variables can include importance, urgency, difficulty, setup cost, available time, current bandwidth, prerequisites, deferrability, and whether the current environment makes action practical.
Device affordances are opportunity structures
A phone, iPad, and Mac are not merely differently sized displays. They create different opportunity structures.
A phone may make awareness, quick reply, capture, lookup, and simple triage easy while making long composition, multi-window comparison, attachment-heavy work, or tasks requiring several applications costly. An iPad can change that balance. A Mac can make substantial work much easier while being unavailable during travel, transit, meetings, or short gaps.
Those affordances change both what users can do and what they may be trying to accomplish.
On a phone, success may mean noticing that something important exists, deciding whether it deserves interruption, handling a tiny action, retrieving a link, or handing work off to a better context. On a Mac, the goal may be composing, comparing, editing, synthesizing, or completing substantial work.
That is why this sequence can be a success:
see on iPhone → defer → open on Mac → act
The phone created useful awareness and allocated work to a better opportunity. Treating “no phone reply” as failed engagement would confuse device constraint with low value.
Revealed behavior is selected by context
Social science routinely distinguishes what people want from what their circumstances allow them to do.
Device-specific behavior has the same identification problem. If the device determines which actions are feasible, then the observed sample of actions is selected by opportunity as well as preference.
This is especially important for learning systems. Repeated mobile non-action should not automatically become evidence that the underlying message is unimportant. Repeated pass-through on both phone and desktop is more informative. A phone exposure followed quickly by substantive desktop action can instead be strong evidence that the message mattered but was better suited to another context.
The system should therefore learn from trajectories, not isolated event counts.
Opportunity is dynamic
Even “this is a computer task” is not timeless.
If the user has not touched a Mac for several days, waiting for a theoretically better device may stop being useful. A task normally deferred to desktop may become worth attempting on an iPad or phone. Conversely, when a Mac is clearly available, deferring cumbersome phone work can be efficient rather than avoidant.
Device suitability should therefore be contextual and probabilistic, not a permanent binary label.
What matters is different from what should be done now
Spillway should preserve a distinction between underlying value and current action opportunity.
A two-hour task can remain critically important when the user has ten minutes. A five-minute task can still be the best use of those ten minutes. Repeatedly avoiding a hard task does not make the task less important.
This suggests keeping several dimensions distinct rather than collapsing them prematurely into one “priority” number: importance, urgency, action value, effort, initiation cost, deferrability, device suitability, awareness value, and uncertainty.
Prediction, preference inference, and recommendation are different problems
A system can become very good at predicting what a person will do next without learning what they believe they ought to do.
Spillway therefore aims to distinguish:
- Behavioral prediction: What is the user likely to do?
- Latent-value inference: What does the evidence suggest matters, motivates, constrains, or frustrates the user?
- Recommended allocation: Given values, commitments, constraints, uncertainty, device, and context, what deserves attention now?
Collapsing those questions is one route from helpful personalization to engagement optimization.
Sparse human statements can identify what behavior cannot
Behavior is often ambiguous. A brief explicit signal can be highly informative.
“Critically important, but I keep avoiding this” separates importance from difficulty or initiation friction in a way that repeated deferral alone cannot. “Not on mobile” distinguishes device suitability from overall value. “Done” distinguishes completed attention from mere lack of further interaction.
The value of these signals comes from identifying latent causes, not from turning the interface into a survey instrument.
Value of information connects inference to resource use
Additional computation has a cost in time, memory, energy, and sometimes money.
So the relevant question is not simply whether another model pass could produce more text. It is whether reducing an uncertainty is likely to change a consequential decision.
A rough conceptual rule is:
value of another pass
≈ expected information gain
× probability the new information changes a decision
× value of improving that decision
Knowing that a receipt is a receipt may not justify another expensive pass. Uncertainty about whether a message contains a hard deadline tomorrow might.
The broader design implication
Spillway is trying to estimate the value of allocating scarce attention and action opportunities across contexts without mistaking observable behavior for the user's underlying goals.
That requires preserving distinctions among motive, opportunity, difficulty, urgency, uncertainty, and device context—and being willing to say when the evidence cannot identify which explanation is correct.
Computational Fallacies · From Data to Knowledge · References and Further Reading · The Ideas Behind Spillway
Documentation provenance: Iterative human–AI construction. See Documentation Provenance.