🧬 biology

You Are in Control of Your State: Why Human Outcomes Are Controllable Through Causal State Intervention

This paper argues that human behavioral variability stems from dynamic, latent internal states that can be causally intervened upon to precisely control outcomes, a framework supported by six strands of scientific evidence and a large-scale observational study of over 200,000 users.

Original authors: Suraj Biswas, Saurav Gupta, Pritam Mukherjee

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suraj Biswas, Saurav Gupta, Pritam Mukherjee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Puzzle: Why Do We Act Differently?

Imagine you are a student. You get the exact same comment on two different essays from your teacher.

Monday: You read it, feel motivated, rewrite the essay for two hours, and submit a better draft.
Tuesday: You read the exact same words, feel annoyed, close the laptop, and don't touch it for three days.

The input (the comment) was the same. You (the person) were the same. But the result was totally different.

Most scientists and AI systems try to solve this by looking for more data points: "Was she tired? Was it raining? Is she an introvert?" But even with all that data, they can't perfectly predict why you acted one way one day and another way the next.

This paper argues that the missing piece isn't a new data point; it's your internal "State."

The Core Idea: You Are a Dynamic Weighting System

The authors say that at any given moment, your brain isn't just a static computer. It's more like a mixing board with hundreds of sliders (biological, emotional, physical, mental).

The State: This is the current position of all those sliders.
The Weighting: Depending on where the sliders are, the same input gets "weighted" differently.
- Example: If your "Sleep" slider is low and your "Stress" slider is high, a critical comment from a boss gets weighted as a "personal attack."
- Example: If your "Energy" slider is high and your "Confidence" slider is high, that same comment gets weighted as "useful feedback."

The paper claims that outcomes are controlled by this internal state, not just by the outside world.

The "Bottleneck" Problem: Why Your Brain Lies to You

The paper introduces a fascinating concept called the Attentional Bottleneck.

Imagine your brain is a massive stadium receiving billions of signals every second (heartbeats, background noise, memories, smells). But your conscious mind (what you can talk about) is a tiny, narrow tunnel that only lets a few signals through at a time.

The Filter: Your current "State" acts as the bouncer at the tunnel. It decides what gets through to your conscious mind.
The Trap: When you ask yourself, "Why did I get so angry at that comment?" you are only seeing what made it through the tunnel. You might say, "Because the comment was rude."
The Reality: The real reason might have been that your blood sugar was low, or you had a bad night's sleep, which changed the "bouncer" to let the anger through.

The Paper's Claim: Trying to change your behavior by just "thinking harder" or "reasoning it out" is like trying to fix a traffic jam by shouting at the cars already stuck in the tunnel. It's inefficient. You need to fix the traffic before it hits the tunnel (by fixing your sleep, energy, or stress levels).

How to Actually Take Control

The paper argues that you are in control, but not in the way you think. You can't just "will" a specific outcome (like "I will be calm"). Instead, you control the State and the Timing.

Think of your day like a river:

Sometimes the river is calm and wide (you are rested, focused).
Sometimes it's a raging, narrow torrent (you are stressed, tired).
The Intervention: If you try to steer a boat (make a decision) in the raging torrent, you will crash. But if you wait for the calm patch of water (a "state-transition window"), a tiny nudge can steer the boat perfectly.

The Takeaway: To control your outcomes, don't just focus on the decision. Focus on:

Upstream Channels: Fix the biology (sleep, food) and physiology (breathing, movement) that feed into your brain.
Timing: Recognize when you are in a "bad state" and wait for a better moment to make big decisions or accept feedback.

What This Means for Technology (AI)

The authors look at how current apps and AI work. They say most AI is like a correlation machine: "You clicked this button yesterday, so we will show you this ad today." It looks at your past habits (features) but ignores your current internal state.

The paper argues for a State-Aware System. This system would need to:

Estimate your current "mixing board" settings in real-time.
Know when to intervene (timing is more important than the content of the message).
Understand that the same advice given to a tired person vs. a rested person will have totally different results.

Summary of Claims (What the Paper Actually Says)

Variability is Real: The fact that you act differently on different days isn't "noise" or error; it's a feature of how humans work.
State is Causal: Your internal state causes the outcome, it doesn't just correlate with it.
Control is Conditional: You can control your life, but only by managing your internal state and choosing the right moment to act.
The "Verbal Narrative" is Flawed: What you say about why you did something is often a compressed, state-dependent story, not the full truth.
Current AI is Limited: Most personalization systems fail because they ignore the "state" and only look at static user profiles.

What the Paper Does NOT Claim:

It does not claim to have a specific "magic pill" or clinical cure.
It does not claim to have proven these things with final, clinical medical data (it uses observational data from a digital platform).
It does not say you can control everything at all times; it says control is "bounded" and depends on your current state.

In short: You are not a robot with a fixed program. You are a dynamic system. To get the result you want, you have to tune the system (your state) and wait for the right moment to act.

Technical Summary: You Are in Control of Your State

Problem Statement

The central puzzle addressed by this paper is the persistence of within-person variability in human behavior. Empirical observations across psychology, behavioral economics, and computational psychiatry demonstrate that the same individual, presented with identical observable inputs, produces divergent outcomes on different occasions. Conversely, different individuals under nominally similar conditions produce non-uniform outcomes that cannot be fully predicted by observable covariates (e.g., mood, sleep, time of day) or interaction terms.

Current correlational personalization systems fail to account for this residual variability, often treating it as noise. The authors argue that this variability is not random error but the signature of a real, characterizable feature of human cognition: the dynamic latent state of the person at the moment an input is processed. Existing models that treat the person as a static feature vector are insufficient because they ignore the time-indexed weighting of biological, physiological, and neuropsychological dimensions that governs decision-making.

Methodology and Framework

This paper is a position contribution rather than a pre-registered clinical or empirical study. It develops a theoretical framework grounded in six established strands of literature: causal inference, predictive processing, allostasis, attentional bandwidth, chronobiology, and computational psychiatry.

The core methodology involves:

Defining State as a Weighting Vector: The authors define the state ( $S_t$ ) not as a static feature vector, but as a time-indexed weighting vector over dimensions (biological, physiological, neuropsychological) that determines how an event ( $X_t$ ) is evaluated to produce a decision ( $A_t$ ) and outcome ( $Y_t$ ).
Causal Architecture: The relationship between state, decision, and outcome is posited as causal, not correlational. The framework utilizes Pearl's causal hierarchy to argue that observational distributions underdetermine interventional outcomes.
The Bandwidth Argument: The paper leverages the "attentional bottleneck" (the gap between ~10¹¹ neural events and ~40–60 bits of conscious report) to argue that conscious verbal narratives are state-dependent compressions of reality, making them poor targets for intervention.
Observational Context: The framework is motivated by a 24-month observational window (2023–2026) from a deployed behavioral platform involving over 200,000 consented users across four occupational personas (students, developers, founders, researchers). The authors report aggregate qualitative patterns from this deployment to ground the framework but explicitly state that quantitative effect sizes and clinical measurements are not reported in this paper.

Key Contributions

1. Theoretical Formulation

State as a Dynamic Weighting Variable: The paper formalizes the state as a latent vector that evolves through slow (circadian), medium (social/work), and fast (immediate event) update operators. It posits that the same observable features can yield different outcomes because the weighting of those features changes with the state.
Causal Path Identification: The authors map the causal path from state to outcome through three intermediate channels:
- Biology: HPA-axis dynamics, endocrine signaling.
- Physiology: Heart-rate variability, sleep architecture, energy availability.
- Neuropsychology: Attention allocation, appraisal, working memory.
- Crucially, the decision itself is identified as the least efficient intervention point because it is downstream of the state-determined filtering process.

2. Operational Requirements for State-Aware Systems

The paper derives six operational requirements that any system aiming to support state-conditional intervention must satisfy, contrasting them with current correlational baselines:

C1. Real-time state estimation: Must operate at a cadence matching state dynamics (e.g., hourly, not daily).
C2. Causal-path attribution: Must attribute outcome divergence to specific channels (bio/physio/neuro) rather than just associations.
C3. Intervention timing: Must target state-transition windows rather than engagement triggers.
C4. Per-individual outcome forecasting: Must produce probabilistic forecasts at the individual level, not cohort averages.
C5. Counterfactual replay: Must be able to reconstruct what would have happened under alternative interventions for past episodes.
C6. Welfare separation: Must structurally separate proximal engagement signals from distal welfare signals.

3. Testable Predictions

The framework commits to seven falsifiable predictions to distinguish state-aware models from correlational ones:

P1: Within-person variance in outcomes (for fixed events) exceeds between-person variance.
P2: State-aware models can recover the specific causal channel of divergence; correlational models cannot.
P3: Intervention timing (proximity to state-transition) dominates intervention content in effect size.
P4: Individual-level forecast calibration improves with observation depth for state-aware systems but plateaus/degrades for correlational baselines.
P5: State-aware systems exhibit counterfactual coherence in paired-episode predictions; correlational baselines fail symmetry tests.
P6: State-aware systems can optimize for distal welfare at a controlled cost to engagement; correlational systems face a steep trade-off frontier.
P7: Structural state dynamics generalize across personas, whereas correlational models trained on one persona transfer poorly to others.

Results and Observational Findings

As a position paper, the authors do not report statistical results or quantitative findings. Instead, they report qualitative patterns observed in the 24-month deployment that are consistent with the framework:

Within-week regularities: Day-of-week effects (e.g., Mondays vs. Fridays) create a stable "envelope" for state trajectories that overrides individual event features.
Bimodal intervention outcomes: When interventions are timed relative to state-transition windows, outcomes cluster bimodally (high-effect vs. low/null-effect). Engagement-triggered systems that average across timing lose this signal.
Variance hierarchy: Within-persona variance is consistently larger than between-persona variance, suggesting that state is a more critical variable than occupational persona for predicting outcomes.
Timing sensitivity: The deployment suggests that a "modest intervention delivered at a state-transition window" has an outsized effect compared to a "useful intervention delivered at an unfavorable moment."

Significance and Claims

The paper claims to provide a prescriptive and operational account of human outcomes that moves beyond correlation. Its significance lies in:

Reframing Control: It argues that human outcomes are controllable, but only through interventions on the state and the timing of intervention, not by willing the outcome directly or adding effort to the decision layer. Control is conditional on the state-trajectory.
Critique of Current AI: It posits that the default architecture of AI personalization (user-features × context $\to$ engagement) is structurally misaligned with the problem of human welfare because it operates on stable features rather than time-indexed states and conflates engagement with welfare.
Design Implications: It offers a blueprint for digital health, education, and AI systems, arguing that "timing is not a small refinement on top of content optimization; it is the larger lever."
Ethical Stance: It asserts that state-aware systems are more powerful instruments than correlational ones, necessitating strict "welfare separation" to prevent the extraction of engagement at the cost of user well-being.

The authors explicitly limit their claims: they do not assert that specific interventions work, nor do they claim their platform is a clinical instrument. The empirical observations serve as motivating context, with quantitative validation deferred to a future companion paper. The paper serves as a position contribution to commit the field to a research program focused on causal state representation and individual-level controllability.