Decision Spaces

Preface:

The future belongs to those who can articulate what they want with precision. Decision spaces are the language of that articulation.

As AI systems continue to become more capable, the bottleneck will shift from computation to preference elicitation. Through MetaPoll, a single person or group of any size may quickly and efficiently map a decision space that can be fed as a highly structured prompt with explicit trade offs encoded, providing execution logic that allows agents to act longer without supervision while producing better, more aligned results.

What is a Decision Space?

A decision space represents the set of choices available to a person or agent at any given moment.

Formally, in decision theory, a decision space $A$ is defined as the set of all possible decisions an agent could take. This mathematical view treats decision-making as selecting one element from the set $A$.

To derive $A$, we can express the relationship between potential and actual decision spaces as:

$$|P| - |J| - |U| + |J ∩ U| = |A| ~(Actual~Decision~Space)$$

To help conceptualize what's going on here, we can visualize the landscape as something like this:

Where:

• $P$ = All theoretically possible options (including absurd ones like "buy a yacht" when you have $10 in your bank account)

• $J$ = Infeasible options (blocked by physical laws, financial constraints, legal restrictions, etc.)

• $U$ = Unseen options (feasible choices the agent doesn't recognize or incorrectly believes are infeasible)

• $J ∩ U$ = Overlapping options (accounting for double counting of options that are both infeasible and unseen)

• $A$ = Actual decision space (what remains)

Demonstrated by $U$, our decision spaces are bounded not just by external rules or laws, but by our own perception.

Chess as an example

If someone is playing a game of chess, $P$ is the number of moves that could be made even if they're not legal moves, such as sliding the Knight diagonally like a Bishop, or moving the King like a Queen. Essentially, it's moving any piece to any square.

You could even argue that $P$ includes ridiculous moves like, rolling the king across the board to play bowling with the other pieces.

$J$ is then the number of moves that are illegal due to the rules of chess, $P-J$ leaves you with only legal moves left.

However, as most people who have played chess know, just because there is a good legal move to make, it doesn't guarantee that you will see it. $U$ is all the moves that are legal to make yet you don't see them or even mistakingly think that they were not legal moves.

The impact of unawareness $U$ on our daily lives cannot be overstated.

Many humans consistently make suboptimal decisions because they literally cannot see better options that exist. Our awareness acts as a filter - expanding it reduces $U$ and reveals a richer decision space $A$ that was always there.

While $U$ is our blindness to the true shape of our option space. Sometimes this blindness is simple ignorance. Yet other times: we may see an option and our internal models immediately distort it, marking viable options as infeasible.

What does this look like in practice? It's the person who has the skills, resources, and idea start a successful business, but tells themselves they can't do it, and believes the lie. People can shrink their own decision spaces and hold themselves back.

Essentially, $U$ is an awareness problem that includes both unseen options and misperceived ones. We consider both of these scenarios as "unseen" because misperception is the lack of seeing something for what it is. (In context of the State Change Loop, $U$ interacts with both the Sensing and Judgement nodes.)

$J ∩U$ is some housekeeping in the event options get double counted by belonging both to J and U. For example if there are 100 cups, 60 are lost, 20 are broken, 15 are both lost and broken, then there are not only 20 normal cups because the cup that is lost got double counted when it was also counted as broken.

The way to correct the double count is to add back the cups that are united in both sets.

$100 - 60 - 20 + (15) =35~cups$

Simplifying and summing up:

When you subtract all the infeasible moves and unseen moves ($P - J - U$ ), you are now left with only seen legal moves which we call $A$ which leaves you with the set of options you need to make a decision within. A decision space.

Why should you care about decision spaces?

The same AI systems that are turning every industry upside down are about to collide with decision-making, and they're missing the most critical data: what humans actually want.

People who aren't thinking about the impact rich decision space mapping will have in our future are missing something obvious coming.

Just as AI is trained on massive datasets, AI makes better choices when it considers rich preference data when making decisions for how to fulfil the user's request, just as a good, well structured prompt is the difference between a good or bad output. In other words: garbage in, garbage out.

When trying to deploy AI at scale for society at large, those that don't leverage decision space mapping will be stuck with 18th-century voting systems trying to coordinate 21st-century AI.

While everyone myopically obsesses over training data for language models, almost nobody is building the infrastructure to capture rich human preferences at scale.

The organizations, products, and communities that figure this out first will have an enormous upstream advantage in the AI-powered future.

Consider how badly binary choice compression loses information. When 100 people vote between two options, the result might be 51-49, but this ignores 99% of their internal preference data and tells us nothing about:

• What alternatives they would have preferred

• Which options of each dimension are valued or opposed

• A preference mapping of tradeoffs between competing dimensions

• How they would have voted if the options were different

Why haven't we mapped decision spaces before?

Our decision-making infrastructure has historically been constrained by practical limitations. When you need to coordinate millions of people without computers, you're forced to aggressively prune and compress complex preference landscapes into simple binary choices: (yes or no), (candidate A or B), (proposal pass or fail). This wasn't a design choice - it was a technological constraint. MetaPoll has changed the equation.

The transition from binary choices to rich decision spaces will be one of the most under-appreciated shifts in how humanity will coordinate on mass. When rich human preference data is connected with advanced AI systems it will unlock the next generation of collective intelligence.

The Decision Space Timeline

Next, let's take a quick bird's eye view of the story up to now.

Ancient Democracy (500 BCE): Athenian citizens dropped stones into urns—white for yes, black for no.

Condorcet's Paradox (1785): Marquis de Condorcet mathematically proved that collective preferences could be cyclic (A > B > C > A), revealing fundamental issues with pairwise comparison when dimensions interact.

Early Decision Theory (1940s-50s): Statisticians and economists like Abraham Wald framed decision-making as choosing from a set of actions to maximize expected utility. While they didn't use the term "decision space," their mathematical formulation inherently treated decisions as points in a set to be searched for optimal choices.

Bossert's Decision Space Framework (1998): Thomas Bossert coined the term "decision space" to describe the range of choices local officials have when responsibilities are decentralized. His framework quantified autonomy across different functions (finance, staffing, procurement), allowing researchers to link decision-making latitude to organizational outcomes. This marked the first practical, measurable application of the concept.

Broader Awareness (2000s): Decision space terminology gains braoder adoption, including fields such as engineering design, multi-criteria decision analysis (MCDA), neuroscience, and business strategy. Notably, game designers adopted decision space terminology to describe "all meaningful choices available to a player at a given time."

Data Driven Decisions (2010s): The dominance of big data led to "data-driven decision-making". Companies heavily invested in analytics, setting up dedicated data science teams to ensure decisions were backed by solid data. The term "Decision Intelligence (DI)" was coined in the 2010s, integrating data science, social science, and managerial science to design better decision processes.

LLMs and MetaPoll (2020s): New generative AI systems such as ChatGPT capable of processing massive text sets and thus preference data suddenly made high-dimensional preference learning feasible. MetaPoll launches and sets the stage for rich decision space mapping.

Autonomous Agents (2030s): Agentic AI both living as software systems and inhabiting physical robotic form will make decisions and take actions on our behalf for extended periods of time without supervison. An Agent's "source code" will be made up of a preference mapped decision spaces, which creates the organized and defined spec of a human's intentions and values allowing AI agents to decide and act without the human present while still maintaining alignment with the human's goals.

Examining Decision Spaces: From Tic-Tac-Toe to Global Coordination

To ground decision spaces in reality, we're going to go on an exploration of decision spaces ranging in size from tiny to gigantic.

We'll start with signal dimensional Tic-Tac-Toe, then explore chess, look at an AI agent, a human life, and then shift to organizations increasing in size until we reach global decision spaces.

The goal here is to illuminate how decision spaces work in the real world so they feel more concrete and real to you.

DISCLAIMER: While we've taken as much care to calculate as accurately as we can, we acknowledge that our calculations, approach, and assumptions may be incorrect. This math has not been formally proofed as we lack the resources to do so at the time of posting this information. However, we feel even our initial findings are worth showing. That being said, if you see any issues with the math and results posted, we would welcome speaking with you to correct them.

Why we are using Pairwise Comparisons

While we explore the difference decision spaces, we will be using Pairwise Comparison as the method for mapping the decision space. Pairwise would actually be a terrible way to do this, however, pairwise has been used since the dawn of history since the greek days, and even still today. It is our opinion that using Pairwise to map the preferences within each example decision space will illustrate how infeasible it would have been to map decision large spaces until better methods were invented (shown later in this article).

Mathematical Framework - defining the variables:

General Decision space metrics:

• P = All conceivable options

• J = Infeasible options

• U = Unrecognized options

• (J ∩ U) = Union of options with both J and U

• A = P - J - U + (J ∩ U) (actual decision space)

Structural metrics:

• M = Number of dimensions

• B = Average options per dimension

• W = Total rankable options

Pairwise Comparison Computational metrics: Please note that to account for multi-dimensional decision spaces, bundles comparing 2 way cross dimensional comparisons K needed to be added.

• α (Alpha) = 1.3

• ρ (Rho) = 9

• K = ρ × W^α

• Y = ( W + K )

• C = Y(Y-1)/2 (pairwise comparisons needed)

• V = C × decision_time [6 sec] (time it takes the voter to vote)

• T = C × comprehension_time [3 sec] (time it takes a viewer to understand the results)

Single-dimensional spaces: Tic-Tac-Toe

Here, the decision space is a one-dimensional choice: you choose which square in a 2 dimensional grid to place a mark. Starting with an empty board, player 1 faces nine options in their decision space to place their fungible X or O symbol: A1, A2, A3, B1, B2, B3, C1, C2, C3. Each move reduces the next player's decision space by exactly one option.

The complete decision space has:

• M = 1 dimension (square selection)

• B = 9 subdimensional options

• W = 9 total options (W = M × B)

• C = 36 pairwise comparisons needed to fully map preferences (Example: Prefer A2 or C1?)

• V ≈ 3.6 minutes

• T ≈ 1.8 minutes

This simplicity makes tic-tac-toe computationally tractable. But most real-world decisions aren't so clean.

Multi-dimensional space: Chess

Chess explodes into multiple dimensions. Unlike Tic-Tac-Toe where you decide to place one fungible mark across the board, with Chess each piece type represents a different dimension, with unique sets of legal moves for each different piece as sub-options.

For example: a queen on G2 can move to 21 squares on an open board: along ranks (A2-H2), files (G1-G8), and diagonals, a knight will have different set of options, the king another and so on. Compared to Tic-Tac-Toe , Chess is a much richer and more complex game with a larger decision space.

Early, Mid and Late game will have varying sizes of decision spaces, for example at the end game if you only have a king and the king has two squares to move, the decision space will be tiny. However for the majority of the game your decision space will be many times larger.

A normal mid-game chess move involves:

• M ≈ 16 dimensions (remaining pieces)

• B ≈ 5 average legal moves per piece

• W ≈ 80 total options needed to fully map preferences

• C = 3,809,544 comparisons

• V ≈ 6,349 hours (for each voter to express their preferences)

• T ≈ 3,174.6 hours (to comprehend the results)

Even at a relatively small scale such as Chess, we can see how computationally challenging decision space mapping is using traditional methods. It's a bit like trying to stream a 4k movie with a slow internet connection from the 90's. Netflix couldn't have delivered a video streaming platform when internet bandwidth was limited to 56k.

Consider that upgrading from a 56k modem to a 1 Gigabit fiber line is a 17,857× improvement (1,000,000,000 / 56,000 = 17,857) ) However, the efficiency gains from MetaPoll exceed that. MetaPoll could cut ≈6,000 hours of preference expression down to ≈5 minutes, a roughly 92,000× improvement. Things that were simply infeasible before, now allow whole new possibilities to open up.

AI Agent Decision Space

Now, let's move out of the space of simple games and into an AI agent's shoes. An AI agent's decision space represents the array of actions it can select when operating autonomously on behalf of a principal (an individual human, a team, an entire organization, or a higher level AI agent or system acting as principal).

Let's use an example of a general AI assistant managing life, its dimensions might include:

Social:

• Calendar management (importance vs. relationship maintenance vs. time availability...)

• Email prioritization (sender importance vs. urgency vs. project relevance vs. cognitive load...)

• Information sharing boundaries (privacy vs. transparency vs. social norms...)

• Communication style per recipient (formality vs. warmth vs. efficiency...)

Financial:

• Savings rate (10%, 20%, 30%...)

• Risk tolerance (low vs medium vs high...)

• Expected return ( 5%, 8%, 12%, 20%, 50%...)

• Insurance optimization (coverage levels vs. deductibles vs. premium costs vs. risk probability...)

• When to pay bills (optimize for interest vs. cash flow vs. credit score...)

• Investment rebalancing triggers (risk tolerance vs. tax implications vs. market timing...)

• Automated purchase authorization limits (necessity vs. budget vs. long-term goals...)

• Subscription management (value assessment vs. usage patterns vs. future needs...)

Health:

• Medication reminders (criticality vs. side effects vs. daily routine...)

• Doctor appointment scheduling (urgency vs. specialist availability vs. work conflicts...)

• Exercise prompting (fitness goals vs. energy levels vs. injury risk...)

• Meal planning (nutrition vs. preferences vs. budget vs. time constraints...)

• Sleep optimization (duration vs. schedule consistency vs. social obligations vs. work demands...)

• Symptom monitoring (when to alert vs. track vs. seek care vs. wait and see...)

• Supplement management (evidence quality vs. cost vs. interactions vs. perceived benefit...)

• Mental health support (intervention timing vs. privacy vs. professional help thresholds...)

Informational:

• Notification filtering (real-time alerts vs. batched summaries vs. do-not-disturb respect...)

• File management (chronological vs. topical vs. project-based vs. frequency of access...)

• Data retention (storage costs vs. future value vs. privacy vs. legal requirements...)

• Knowledge synthesis (when to summarize vs. preserve detail vs. create connections)

• Archive decisions (what to keep vs. delete vs. cold storage vs. immediate access)

Plus other dimensions like:

• Home management (temperature, lighting, security, maintenance, supplies...)

• Transportation (route preferences, timing, cost vs. speed trade-offs...)

• Shopping (brand preferences, quality thresholds, substitution rules...)

• and more...

The agent's "unseen" set U swells not from only ignorance, but from incomplete preference maps:

For example, what if an AI was inside a personal robot at their owner's home? The principal says to the AI "Please take care of my home while I'm away for the next 2 weeks."

Perhaps a week into the caretaking job, a wild fire approaches the home, the AI recognizes the danger to the house, but it's never been told what is a higher priority to save in the case of an emergency and communication with the principle is not available. Does the AI save your $2m dollar painting first, or your dog, or try to go fight the fire to save the house?

Without a preference map already detailed by the principal, the AI robot now needs to guess what you'd like it to do.

Estimates for a general personal AI agent:

• M ≈ 500 operational dimensions

• B ≈ 4 options per dimension

• W ≈ 2,000 total options

• C ≈ 15,846,877,279 comparisons

• V ≈ 3,015 years

• T ≈ 1,508 years

For organization-scale agents, multiply by factors of 5× to 10×. Without mapped preferences, agents traverse foggy terrains, where up front data gaps are simplicity shortcuts that lead to misalignment and suboptimal results.

MetaPoll provides detailed "prompts" in the form of MDCT preference trees, ensuring agents optimize for what we truly value, creating better more aligned results.

A Human Decision Space

Human decision spaces are far bigger and more complex than we realize. When was the last time you stopped to consider how many choices you make and remake in your life? If you're like most people, probably hardly ever or maybe never.

When we look at a human life, they begin as a baby with a minimal decision space (cry or don't cry, and even that is debatable), then the decision space expands dramatically through childhood (what games do I want to play?) and teenage years (who will choose to be my friends? how much will I study? will I rebel against my parents?).

By adulthood, the decision space becomes staggeringly complex:

Career dimensions: Job type, hours worked, income targets, location, industry, company size, remote vs. in-person...

Living dimensions: City, neighborhood, housing type, ownership vs. rental, roommates, temperature settings, furniture choices...

Health dimensions: Diet type, meal frequency, exercise routine, sleep schedule, medical checkups, supplements...

Financial dimensions: Savings rate, investment allocation, risk tolerance, retirement planning, insurance coverage...

Social dimensions: Friend groups, romantic relationships, family planning, community involvement...

Day to day interaction decisions: Do I lie to a coworker when they ask if their sweater looks good on them? My older kid wants me to come see his model airplane, but my younger kid wants to show me their new finger painting, what do I do? ...

How dimensions interact

Seeing all these lists of dimensions doesn't really give you a concrete idea of what it might look like in day to day living. Let's use a hilariously small example of deciding what temperature to heat your home at.

Imagine you come home from a day at work during the winter. The house is cold, you need to decide what temperature you want to set the thermostat at {68℉, 69℉, 70℉, 71℉, 72℉, 73℉, 74℉, 75℉}. You think you'd like it nice and toasty warm around 74℉, but then you remember gas prices have gone up 35% last month and you have some major car repairs next week that are going to stretch your budget to the brink.

Is your priority to get the car repairs done and stick to your budget (setting heat to 70℉) ? Or dip into your savings so you can be warm ( 74℉) while also getting the car repairs done? Or should you keep the temperature low ( 68℉) so you can save up more money?

Maybe you should get a second job so you don't need to make these choices, but do you see increasing financial income as a higher priority than maintaining your physical and mental health and having the time to spend with your family?

Maybe you could invest in more energy efficient insolation so your heating costs would go down, but can you save up enough money for the upfront costs without ruining your monthly budgets, and is that more important than going to visit your Dad (out of country) next year for his 80th birthday?

This is just a tiny example of how complex a human's decision spaces are. Dimensions interact with eachother in complex ways frequently.

Estimating decision space size:

• M ≈ 1,000 life dimensions

• B ≈ 5 options per dimension

• W ≈ 5,000 total decision points

• C ≈ 170,702,942,388 pairwise comparisons to fully map the decision space

• V ≈ 32,477 years

• T ≈ 16,239 years

The sheer scale explains why humans use heuristics, habits, and social copying rather than optimizing from first principles.

Realistically, a human is only making a tiny fraction of these decisions on a day to day basis. However, the takeaway is that it's a huge task for an AI to accurately understand and model all your life's preferences.

Organizational Decision Spaces - A Hospital

Organizations today are simply groups of many people, systems and technology working together to accomplish goals.

So what does this mean for the decision space? Amplify everything by 10x or more. To illustrate, consider a mid-sized hospital, where the decision space must balance patient care with administrative realities in real time. Imagine trying to coordinate hundreds of departments that all affect each other.

Adding to the challenge, hospitals operate under uncertainty, fluctuating patient volumes, and dynamic health threats that inflates $U$ through overlooked contingencies and $J$ via legal or ethical barriers. For example, what if a major earthquake occurs, sending a huge wave of injured patients. What about a mass shooting? A fast spreading pandemic? Hospitals need to be able to adapt to situations as they develop to accommodate many possible scenarios.

Key dimensions include:

• Medical: Treatment protocols, drug formularies, surgical techniques, diagnostic workflows, specialist rotations...

• Staffing: Recruitment standards, shift scheduling, skill compositions, professional development, performance evaluations...

• Infrastructure: Equipment acquisitions, maintenance routines, bed and surgery time assignments, laboratory expansions, emergency response times, digital systems...

• Financial: Reimbursement strategies, insurer contracts, charitable care guidelines, capital investments, medical malpractice capital retention pool policies...

• Regulatory: Audit procedures, data reporting, quality benchmarks, safety standards...

When the emergency room gets overwhelmed, it pulls nurses from other floors. When you cut the maintenance budget to afford new MRI machines, equipment starts breaking down more often, causing delays that frustrate doctors and endangers patients.

Estimates for such a hospital:

• M ≈ 10,000 operational dimensions

• B ≈ 10 options per dimension

• W ≈ 100,000 total options

• C ≈ 407 trillion comparisons

• V ≈ 77.5 million years

• T ≈ 38 million years

With so many decisions to make at the hospital scale, it's clear why organizations lean on hierarchies, specialized committees, and data dashboards, yet these often create data silos missing holistic views that could reveal hidden efficiencies or risks. Preference maps could help the hospital staff know what actions to prioritize when situations change quickly in unexpected ways. When that earthquake hits, should you abandon elective surgeries or delay cancer treatments? During chaos, knowing what to sacrifice is as important as knowing what to save.

As we scale upward to multi-national corps, governments, and global scale spaces, the exponential growth demands technologies capable of taming the chaos without losing the nuance.

Multi-national Company

Multinational corporations make a single hospital look simple, they operate across borders, amplifying complexity through many diverse product markets, regulatory environments, logistic networks, and cultural contexts.

Consider a simplified example of a tech giant like a hypothetical global software/hardware firm:

• Product dimensions: Software features, hardware integrations, release cycles, pricing models per region, customer satisfaction, competitor offerings, localization languages...

• Market dimensions: Entry strategies, competitive positioning, partnership alliances, advertising channels, influencer strategy, customer segmentation...

• Supply chain dimensions: Vendor selection, inventory management, logistics routes, tariff navigation, product insurance, disruption contingencies, sustainability sourcing...

• Human resources dimensions: Global recruitment, compensation structures, training programs, retention programs, performance evals, in person/remote work policies...

• Financial dimensions: Currency hedging, tax optimization, capital allocation, merger evaluations, dividend planning, stock buy backs, risk management...

• PR dimensions: Community management, public good will, consumer trust, crisis management responses, growth campaigns, spokesperson management...

• Legal and compliance dimensions: Intellectual property protection, data privacy regulations (e.g., GDPR vs. CCPA), antitrust monitoring, ethical AI guidelines...

• Innovation dimensions: R&D investments, patent filings, emerging tech adoption (AI, blockchain), internal skunkworks programs, collaboration ecosystems...

These dimensions interact internationally, a product feature might comply in one jurisdiction but violate another, or supply disruptions in one region cascade to altered financial forecasts elsewhere. Hundreds of thousands of employees across tens of thousands of teams, working all over the world create immense complexity and are so large that no one person can know and understand the entire company from end to end. Voting isn't just impractical here, it would take longer than the age of the universe.

Estimating for a multi national corporation:

• M ≈ 100,000 operational dimensions

• B ≈ 15 options per dimension

• W ≈ 1,500,000 total options

• C ≈ 464 quadrillion pairwise comparisons

• V ≈ 88 billion years

• T ≈ 44 billion years

This astronomical scale underscores why corporations rely on hierarchical delegation, AI-assisted analytics, and scenario planning—yet even these tools can struggle with the full multidimensional interplay, often leading to overlooked synergies or risks.

National Government

While the multi-national companies struggle with cross boarder complexity, national governments concentrate that complexity internally at even larger scales. By needing to orchestrate society-wide coordination, governments must balance immediate domestic needs with long-term strategic imperatives across vast populations and territories.

Picture a modern democracy like the United States, the decision space encompasses interlocking layers of policy, administration, and oversight, where choices in one domain ripple through others—often with electoral, economic, or geopolitical consequences.

Key dimensions include:

• Economic policy: Fiscal budgets, taxation structures, monetary controls, interest rates, trade agreements, subsidy allocations, securities laws, inflation targets...

• Foreign affairs: Diplomatic relations, alliance formations, sanction regimes, aid distribution, military deployments, trading partner policies, treaty negotiations...

• Domestic services: Healthcare systems, education curricula and policies, fire fighter services, , infrastructure projects (roads, communications, energy grids), environmental regulations...

• National defence: Intelligence operations, priority resource budgeting, R&D costs, training, reserve and drafting policies, logistical deployments, operation priority, rules of engagement...

• Domestic Justice and security: Law enforcement priorities, judicial reforms, cybersecurity measures, immigration policies, prison policies, prisoner rights, death penalty policy, counter-terrorism strategies...

• Social dimensions: Equality initiatives, cultural preservation, demographic planning (aging populations, migration), educational, regulations and subsidies, social security policies, parental rights, labor rights...

• Technological and innovation: R&D funding, digital governance (e-gov platforms), nano tech, military developments, AI ethics guidelines, space programs...

• Crisis management: Pandemic responses, natural disaster relief preparation and response, economic recovery plans, climate adaptation strategies...

These arenas are deeply interdependent—a tax policy might fund education but constrain foreign aid, while environmental rules could clash with industrial growth targets. Moreover, governments must navigate constitutional constraints, public opinion, and international law, adding meta-layers of feasibility checks. Considering all the ways things can drift out of alignment, an AI system could use all the help it can get by feeding it highly organized preference data.

Conservative estimates for a large nation-state:

• M ≈ 500,000 operational dimensions

• B ≈ 20 options per dimension

• W ≈ 10,000,000 total options

• C ≈ 64 quintillion pairwise comparisons

• V ≈ 12 trillion years of voting time per participant

• T ≈ 6 trillion years to interpret the results

The scope is immense. Even with bureaucratic hierarchies, legislative committees, and advisory bodies trying to handle the complexity, the vulnerabilities to gridlock, corruption, and contextual ignorance remain, where unseen interactions (high U) or infeasible paths (high J) lead to policy paralysis or unintended outcomes that are not inline with what the public wants, yet they must live with the consequences.

To think that a single binary vote for a single candidate comes close to representing a citizen's voice appears ridiculous when considering how large the decision space is. There's no way for a single "yes" or "no" answer to accurately represent that much data. It would be like a single black or white pixel representing a beautiful landscape photo. Not possible, no matter how great your imagination is.

Global Decision Space

At the top of today's human coordination lays a cross-national global scale decision space, where sovereign nations, international organizations, multinational entities, and our global commons (shared ocean and atmosphere) intersect to address planetary-scale challenges.

What does that look like today? Decisions must harmonize divergent national interests, while considering cultural paradigms, global logistics networks, and planetary systems (ocean currents, solar activity, magnetic field strength, global temperatures, plate activity, sea levels, natural disasters, etc) often under the shadow of existential military risks that live across borders.

A fraction of the dimensions span:

• Economic: International trade protocols, cross national resource pricing, global financial record and settlement system, international commodities markets, international financial aid programs, fintech regulations...

• Peace and security: Arms control treaties, cybersecurity norms, non-violent conflict resolution mechanisms, refugee protocols, warfare regulation, nuclear non-proliferation...

• Global health and biosecurity: Community health strengthening, pandemic preparedness, antibiotic resistance policies, mental health initiatives, food security frameworks...

• Technological frontiers: AI governance standards, space resource utilization, satellite and spacecraft technology, global internet sovereignty, biotechnology regulations, quantum computing collaborations...

• Environmental stewardship: Temperature policy, fishing and hunting policies, biodiversity policies, ocean regulations, fresh water management, forest management policies, air pollution regulations...

• Social and cultural: Human rights policies, education policies, gender and race policies, migration protocols, child and parental support policies, cultural heritage policies...

• Resource and infrastructure: Energy generation and infrastructure strategies, water rights regulations, global supply chain protocols, satellite spectrum allocation, disaster response coordination...

The world is incredibly complex, we still don't fully understand all the feedback loops that multiply across dimensions: a climate policy might alleviate health crises but exacerbate economic disparities, while technological advancements could either bridge or widen geopolitical divides.

Moreover, global spaces contend with sovereignty frictions, enforcement asymmetries, and the tragedy of the commons, inflating $J$ through diplomatic infeasibilities and $U$ via cultural or strategic blind spots. Every decision here can massively impact the second order decision spaces of millions of lives, companies, cities, nations, cultures.

Estimate is:

• M ≈ 2,000,000 operational dimensions

• B ≈ 25 options per dimension

• W ≈ 50,000,000 total options

• C ≈ 4 sextillion pairwise comparisons

• V ≈ 803 trillion years

• T ≈ 401 trillion years

The numbers here are just so mind bogglingly intense that they illuminate why global coordination often stagnates in procedural quagmires or defaults to lowest-common-denominator compromises. Planetary scale coordination will require us to unlock and use new technologies like rich preference mapping to distill coherence from chaos, enabling humanity and AI to navigate shared fates with unprecedented clarity and alignment. The TLDR? Our current decision infrastructure is woefully inadequate.

Why today's methods fail

Now that we've got a sense of what a decision space is, let's examine why we haven't traditionally mapped them.

Existing methods fail for essentially two primary reasons we'll discover.

1. Time cost: Current approaches scale catastrophically—requiring up to super-exponential time as complexity grows. For realistically sized decision spaces, the time needed to express preferences can exceed multiple human lifetimes. (Making it obvious why we couldn't do this before.)

2. Structural incompatibility: Existing methods force multi-dimensional decision spaces into one-dimensional rankings. Imagine trying to capture a mountain range on a single line—you lose elevation, valleys, alternative routes, and the relationships between peaks. Similarly, these methods compress rich preference landscapes into flat lists, creating massive blind spots where crucial trade-offs and interdependencies simply vanish from view. Lossy data capture ruins the context needed for collective decision-making in an increasingly complex world.

To further bring this to life, we'll use our medium large scale 100,000-option Hospital example and try calculate how long it would take to map the same decision space using each different approach.

• Pairwise comparison: The full 407 trillion comparisons would take a single voter ≈ 77 million years.

• Multiple choice: Even with 10 options per question, you'd need 10,000 polls. Dimensional tradeoffs are not accounted for. The combinatorial explosion makes comprehensive coverage impossible.

• Ranked choice and Quadratic Voting: Ranking 100,000 options directly would take each voter approximately 1,200 hours just to read through once, and still wouldn't capture how dimensions interacted.

• MCDA - (Multi Criteria Decision Analysis) would take [ ]

Our existing methods create crippling interpretation costs with up to super-exponential interest. It's infeasible for existing methods to map even moderately sized decision spaces. This means there are many gaps in the map that an AI agent would need to guess. (Lack of information for the prompt on what the agent is supposed to do)

What a new solution for the AI era looks like:

Preference Language: Ways to express complex, multi-dimensional trade-off preferences that capture the nuance of human values.

Aggregation Mechanism: Methods to combine billions of rich preferences into collective decisions without losing critical information.

Continuous Updates: Operating as a living system where preferences evolve continuously rather than through single point polling events, capturing shifts in collective sentiment as they emerge.

Verification System: Ensuring strict and robust anti-fraud and inclusion proofs that ensure the preferences being expressed are genuine and not manipulated.

Privacy Preservation: Allowing rich preference expression while protecting individual privacy.

Data Availability: Maintaining a dependable permanent, queryable record of preference data that supports both real-time access and historical analysis with support for advanced analytics.

The transition to rich decision spaces requires new infrastructure.

MetaPoll has:

• A high efficiency standard for preference expression that accounts for multi-dimensional choices

• An interface where people can express their values and explicitly encode trade offs

• Engineered an aggregation system that is privacy preserving and mathematically verifiable to be without fraud to enable global scale collective intelligence

• A continuously living polling architecture that shows evolution of preferences within a decision space

• Ensured these systems are simple, accessible and usable by everyone, not just the nerds

How does MetaPoll stack up?

MetaPoll: Mapping the hospital's entire 100,000 option decision space would require one MetaPoll that can be fully completed in around ≈ 162 hours (a worst case linear scaling scenario), not an easy task by any means. However, MetaPoll's true strength emerges when we account for its crowdsourced nature: individual voters can concentrate their efforts on the domains where they hold genuine expertise and strong views, perhaps those directly tied to their professional roles. Under this more realistic lens, the effective voting time per participant drops to about ≈ 10.3 hours, roughly equivalent to a day or two's work, sufficient to map out their personal slice of the multi-dimensional preference landscape. And this is just the starting point; subsequent updates often demand only minor adjustments, with future votes potentially wrapping up in under ≈ 30 minutes. In this way, the once-prohibitive costs of maintaining continuous consensus begin to feel not just achievable, but elegantly sustainable.

Mapping efficiency of MetaPoll

By using Multi-Dimensional Consensus Trees, combined with crowdsourced selective expression MetaPoll compresses an exponentially complex decision space into logarithmic to linear voter effort.

Voters rank options at each level of the tree, expressing granular preferences without examining every combination—they can simply skip areas where they lack strong opinions.

This creates a powerful network effect: each new participant brings their unique expertise and interests, gradually filling gaps in the preference map. As decision spaces grow larger, MetaPoll can leverage the power of crowdsourcing the right people to evaluate the right areas, ensuring comprehensive coverage without overwhelming any one individual more than they want.

Since voters only rank what matters to them, we expect real-world mapping time to scale sub-linearly, making the system viable even for large scale, complex decision spaces.

To sum up: crowdsourcing the preference expression, and spreading out the work load among many voters. When compared to basic Pairwise, MetaPoll may map a rich decision space (100,000 W) in ≈99.99999% less time, reducing voting time from ≈77 million years down to just ≈85 hours roughly a ≈7.9× billion improvement in coordination efficiency.

AI Alignment and preference data connection

As AI systems grow more powerful, the decision space problem becomes existential. We've built AI models trained on petabytes of data—every book, article, and conversation available on the internet. These systems can generate poetry, solve complex problems, and engage in nuanced reasoning.

Yet when we try to coordinate humanity or understand collective preferences, we're still using voting systems that compress thousand-dimensional human values into a handful of disconnected binary choices—whether planning city infrastructure, allocating resources, or making policy recommendations—it operates on this impoverished preference data. It's like training an LLM where every sentence has been reduced to "positive" or "negative"—the nuance that makes the system useful disappears.

Consider the absurdity: we train AI on datasets with trillions of tokens to capture the subtle patterns of human language, but then try to steer these same systems using polls that reduce complex societal questions to "yes/no" or at best, ranking five pre-selected options. It's like building a supercomputer and then controlling it with smoke signals.

When it comes to what we actually want (our group's prompt), AI systems get a few bits of information from traditional voting. Without a good prompt, AI gives cookie-cutter responses that tend towards being more deterministic, like how when asked for a random number between 1 - 25, every AI model picked 17.

The parallel to prompting is instructive: just as "write something about economics" produces mediocre output while "analyze the game-theoretic implications of quadratic funding mechanisms for public goods" yields more sophisticated analysis.

Feeding AI systems impoverished preference data (AKA: Garbage in Garbage out) produces crude approximations compared to what detailed preference expression can achieve.

MetaPoll's Multi-Dimensional Consensus Trees (MDCT) fundamentally change this equation. Instead of lossy compression, they capture rich preference information across hundreds of dimensions—essentially creating an incredibly detailed prompt for AI systems, complete with explicit priority ordering across all the trade-offs humans actually care about.

How can we hope to solve the AI alignment problem if we're not even starting with a clear picture of what the goal is in the first place?

The future of AI alignment may depend on our ability to articulate preferences at the resolution reality demands. Rich decision space preference expression provides the language for that articulation.

The Agency Problem

What if we could get rid of polling completely and just get AI to watch us, learn our behavior and then predict what we want? Sounds really nice on the surface until you think a little deeper about what kind of world that would create.

Consider how recommendation systems work: they observe your behavior—what you click, how long you watch, what you skip—and build a model of your preferences. But this creates a fundamental problem: the system can only learn from the choices you make within the options it presents.

A room with no doors

If the algorithm decides you're a "comedy person", it will never discover that you might enjoy documentaries too. The feedback loop becomes self-reinforcing: limited options lead to limited data, which leads to even more limited options.

Without a way to actively express preferences beyond the algorithm's narrow view, people become trapped in increasingly constrained bubbles.

Now imagine this problem magnified. As AI systems become more autonomous, making more decisions on our behalf, the gap between what we actually want and what the system thinks we want becomes highly problematic.

This isn't hypothetical. Recommendation systems already create filter bubbles and hyper-polarized echo chambers. The difference is that current platforms still maintain some user agency through search and explicit choice.

What happens when AI agents are making thousands of micro-decisions for us every day, each based on it's guesses rather than our choices?

The key insight is that behavioral inference creates a narrowing spiral:

Past behavior → AI prediction → Limited options → Constrained behavior → Reinforced predictions

Breaking this spiral requires explicit preference expression—the ability to say not just what we've done, but what we want to do, what we value, and which outcomes we prefer over others. Without this, we lose the capacity to explore, grow, and change.

Living our own lives with the assistance of AI

MetaPoll shifts the dynamic. Yes, AI systems can continue working for us, but by providing rich, high-dimensional preference maps about what we want, we ensure we remain seated in the driver's seat of our own lives.

Imagine two different AI assistants.

The first one silently observes your daily routine—noting that you always order the same coffee, take the same route to work, read similar articles. Over time, it begins making choices for you based on these patterns, gradually narrowing your world to match what it thinks you want.

The second assistant has access to your explicitly stated preferences: yes, you enjoy your usual coffee, but you also value trying new experiences 20% of the time. You want efficiency on weekdays but spontaneity on weekends. You're interested in gradually expanding your technical knowledge, even if your current reading habits don't reflect that yet.

The difference is profound.

• The first system locks you into an ever-tightening spiral of your past choices.

• The second helps you become who you want to be.

Two Paths Forward

Decision spaces are the architecture determining everything from individual life satisfaction to civilizational trajectory. We've ignored them not because they're unimportant, but because they were computationally intractable. That's no longer true.

The transition to rich decision spaces isn't just an optimization - it's a necessity for maintaining human agency in an AI-mediated world. As AI systems become more capable, the gap between what they could optimize for (if they knew our true preferences) and what they actually optimize for (based on crude signals) becomes an increasingly dangerous chasm that swallows human agency.

A future where AI systems predict what we want based on what we've done and keep us locked in echo chambers would be a nightmare.

Instead we need humans working with AI systems that help us achieve what we actually value, based on preferences we can actually express.

Once a human or group's preferences are mapped through applications like MetaPoll, that map becomes a resource for automation. AI agents can access the detailed preference landscape and know exactly what trade-offs to make to execute the person or community's will. Without such mapping, AI must guess—and those guesses compound into a future where humans have no agency in their own lives.

We need to adopt and master this infrastructure now, before AI systems become so powerful that their optimization based on impoverished preference data becomes impossible to escape.

We are at a fork in the road: either we express our preferences within a rich decision space, or we descend into in a bottomless prison of no true choices—AI systems guessing what we collectively want while we uselessly shout "yes" or "no" into the void.

The difference between these two futures is the difference between a world where humans maintain agency and one where we become passengers in our own lives.

Let's choose our future wisely.