Macrostate Engineering

What It Is

While planning a morning routine on a call, Will began comparing sequences: which step first, which interval length, and how each choice changed the remaining possibilities. He caught the branching process as it happened: “it doesn't matter what happens as long as the end state is reached and these constraints are not violated.” The routine needed a defined result and boundaries, but it did not need every configuration worked out in advance.

Macrostate engineering specifies an outcome and lets a resolving process find a configuration that satisfies it. In statistical mechanics, the macrostate describes an ensemble-level property while microstates are the detailed configurations compatible with it. An acceptance criterion can admit many such configurations. Which one occurs does not matter if the required properties hold.

The resolver can be an AI agent, a team, the subconscious or the day's execution. The specification must constrain the search without prescribing every detail. Entering one configuration to edit it manually is useful for debugging, but disastrous as the default: one trajectory does not scale, and conscious attention is the system's scarcest resource.

The underlying physics includes the Boltzmann distribution, energy landscapes and entropy. This protocol applies the distinction to prompting, delegation, days and whole-life exploration. The transfer of ensemble thinking, ergodicity and mean reversion to behavior remains a heuristic that has been useful for Will, rather than rigorous physical modeling.

The branching represents permitted alternatives, all judged against the same outcome and constraints.

Read this diagram

The required outcome branches into An acceptable route; Another acceptable route; A route discovered in use.

From Metaphor to Protocol

The morning-planning interruption made a previously available concept into an action. Will learned to recognize the physical feeling of too many branching possibilities as heat in the head. Instead of planning harder, he could identify that he was computing at a resolution that should be delegated.

Days later, the same response interrupted system design: “the engineering questions — 'should the agent plan the frontend,' 'should processes encode displays' — those are all microstates. You're designing the state machine before you've defined what should be true.”

A useful specification sits between an exact configuration and an empty aspiration:

“Get me a good flight” leaves the requirement undefined. “Under $500 by Tuesday” constrains cost and timing while leaving airline, route and booking site available for search. Effective AI use depends on this ability to define macrostates and leave their details to computation:

"I cannot use AI without macro states. Just chatting is actually kind of useless."

Chat can instead bind attention to the search: “I am tired of fking chat mostly because it binds my attention to microstates.” Frustration then indicates a resolution error rather than a model error. Will's diagnosis was “when you get very frustrated with the AI this just means like hey you're thinking at the wrong level.”

Who Resolves the Microstates

The same division can run through different resolvers:

“I care what gets done” leaves different freedom from “I care how it gets done.” That does not make every descent into detail wrong:

"There's actually a good and bad part about micromanaging. Micromanaging is allowing yourself to enter a micro state. And then for the purposes of debugging because the work product is not good."

Debugging deliberately enters a microstate, finds the defect and exits again. The problem is remaining there after the diagnostic work is complete. The hierarchy is also relative: “one microstate might be a macrostate in another person's lens, right?” A report's committed outcome can be one detail in the manager's larger arrangement. Execution Resolution follows that chain of levels.

Scale makes the distinction necessary:

"Single trajectory through space time, person… with one directed goal and… a normal human's working memory and processing capacity. You basically can't survive in this environment… or at least it's not going to scale. We kind of need to upgrade people's vocabulary and thinking in terms of statistical ensembles."

Parallel agents, projects and ideas exceed the time and working memory available to follow each trajectory. Ensembles are the only representation that compresses them enough to manage together.

Counterfactual self-comparison can use the wrong unit for the same reason: “I'm actually just comparing 2 microstates. When I should be comparing my [macro] states.” A particular imagined version of a day is not the measure against which the ensemble succeeds or fails.

Noise at Level N Is Structure at Level N+1

Sometimes an individual instance is too variable to teach, grade or control usefully. That variability can identify a stable process one level above it: “it is an arrow pointing one abstraction level up.” Individual days vary while a weekly routine has a recognizable structure. Posts, trades, workouts and conversations vary while the policies generating them remain describable. “Noise at level N is usually structure at level N+1.”

Teaching prompting exposed this distinction. A recording contains one sampled session, without showing which details were essential and which were variance:

"if there's too much randomness at the prompt layer — you move up a level and teach the generative process — the prompting pattern — the heuristics (macro → micro), the concept of getting a prompt running."

Outside-in decomposition, macro shape before granular verification and the method of converging on a solution persist across different prompts. Those are the curriculum. Taste Compilation turns the judgment involved in that process into language someone else can use.

When instances cannot be graded, the generating policy can still be taught and improved. Agent design can improve decomposition and convergence rather than curate golden trajectories. Personal work can examine the policy behind a scattered session or unsuccessful post instead of trying to control that one instance. The useful intervention moves to the level where variation has a stable structure.

Commit to Macrostates, Not Plans

A plan specifies one trajectory through a week. New information can invalidate it by Tuesday without invalidating the desired outcome:

"You don't even need to stick to a plan. You just need a macro state. You don't need a specific plan. You can modify the plan every day because you modify what you do every day. It's more about wanting a specific macro state."

The plan becomes “a dynamic plan that is computed every day at the beginning of the day.” The weekly target remains fixed while the route is recalculated: “I should lay out what is the end goal state of this week… and then each day I can figure out what I need to do within that day to be close to the macro state.” Evening review compares what happened with that target.

A changing plan can look like a changing commitment from outside. Will distinguished them: “I change my plan every day. But I told you the end goal is the same, but the plan changes every day depending on the information.”

Walking west across San Francisco to reach the ocean gave him an image for it. A dead-end street changes the route without changing the reference:

"It keeps moving around, everything's like that, but I just know on Earth. All you have to do is very simple. You just need to point your body towards the sun."

The fixed reference and recalculated heading guarantee convergence in this example. Search finds the route, while Order of Determination fixes the representation with the most degrees of freedom first. Plans and schedules are derived afterward.

Even the relationship application retains this division: state the outcome and constraints explicitly, then leave the path available to revision.

The Tracker Must Be Visible

A target kept only in the head does not converge because drift has no feedback to oppose it:

"CORE LEARNING — the macrostate tracker must be visible. feedback is key. always put the macrostate up on the board. this is why the habit tracker works: visibility keeps the state legible, creates feedback, and helps convergence happen instead of leaving the target implicit."

A visible tracker keeps the difference between the declared and current state available. The daily review supplies the clock for that comparison: “recap what they did yesterday and then see if they met the macro states. Did they check off all the boxes?”

Declaring the system also changes how an action is considered. It becomes a step in an existing commitment instead of an isolated cost–benefit decision: “It's weird that feeling like there is a system makes you do ur habits even if u don't want to, even if the system is minimal.” A mediocre system that runs therefore beats an optimal one still being designed.

"What I'm doing today has a bigger purpose within something I designed, I am a microstate trying to find the macrostate I designed lol."

A lucid planning session sets the ensemble-level condition. Ordinary hours then supply instances of the process moving toward it. The same requirement separates an agent from an agentic system: “a well-formed agentic system is defined less by what it contains than by the macrostate it guarantees… A system without a macrostate is an agent, not an agentic system.” The guarantee requires an objective, feedback and a termination condition. It applies to a day as well as to software.

Structure Is a Macrostate Converger

Structure both sets the desired state and reduces the choices at each step:

"Realization: structure is basically a macrostate converger and a discretized decision menu. Instead of 'I have 4 hours,' it's 'when I go home my mission is to complete as many things off my habit list as possible.'"

Four hours of unrestricted possibility form an open search that resolves, through the Boltzmann distribution, toward the lowest-energy option. A habit list supplies a finite menu. Choosing among its items is cheaper than generating a new activity from nothing, and every choice can be assessed against the declared mission.

"Without a boundary, intelligence behaves like gas. My work was not failing, it was diffusing. The feedback introduced a boundary, so the diffusion can now become convergence."

The gas comparison applies Intelligence Is Water to capability without containment: it spreads into more space at lower density. A boundary permits that activity to converge. The resulting work taxonomy turns “the normal amorphous blob into clean cybernetic operations”:

All three are work. Mixing their operations during a session makes the day lose its direction. Branching and Convergence develops the shipping phase.

Case Study: One Month of the Protocol Assembling

The pieces emerged across a month of logged work rather than arriving as a finished framework.

In Week 1, work on an explainer system separated defining the frame from exploring its implementations: “I spent last week getting the explainer creation system up, but now we have to collapse the macrostate. ... we have already locked down the form factor, the content on the page, and the general structure. Now we need to explore microstates that fulfill it.” The established form factor, content and structure gave the next week a stable target.

Two days later, the habit tracker was working where other targets were not. Visibility supplied the difference: the board kept a feedback signal available while implicit targets drifted.

In Week 2, evenings changed from four hours of unspecified time to a finite habit menu serving a declared mission. The next day added intermediate targets: “what are the macrostates along the path, not just the final destination?” Waypoints could also be specified as outcomes rather than routes.

In Week 3, planning, working and shipping were named as different operations: “this makes the normal amorphous blob into clean cybernetic operations.” The following day, external feedback about a diffuse week exposed the missing boundary. The work had spread without converging.

The physics vocabulary was already available. Each encountered failure identified an operation that made it usable in practice.

Engineer Environments Where Any Pick Advances

The environment can make each reachable choice satisfy the target:

"Create an environment where it doesn't matter what you pick, it moves you forward. I think that's the thing we need to set up."

Will's example is a selectorized gym. A person with little executive energy can wander between machines and still complete a workout: “it's ADHD compatible, where I can just wander but still settle into a good state because of the macro state engineering.” Each station is a valid option, so no willpower is spent deciding whether it advances the session.

“You're basically a stochastic process… if you don't feel energized by doing one thing, you can always keep doing something.” The menu changes the distribution of outcomes. Prevention removes bad options; this construction makes the remaining options good.

It also changes the unit of evaluation. “All these decisions do not matter, right? It's when the sum of those decisions… the aggregate fails the macro state.”

"You cannot do things perfectly because there's other combinations that would let you succeed… There's not only just one path. Like you can compensate or you can find a different combination that works."

A sandwich does not by itself fail a diet whose criterion is the week's aggregate deficit. Different meals can compensate for one another and still satisfy it. The craving also specifies a property rather than necessarily twelve pieces of one food: “it's actually a macrostate. There's actually many things that can fulfill this.”

Requests Are Macrostates

A craving can request salt, umami, intensity, volume and relief. “McNuggets” names one implementation, made especially available by advertising. Treating that implementation as the entire request leaves indulgence or resistance as the only apparent choices. Extracting the test condition reveals many other satisfying configurations, most of them cheap.

At 2:30am, during a wave of food resistance, Will ate two cans of green beans, mandarin oranges, okra and a quarter Haetban, dressed with Maggi seasoning and wasabi paste. “it felt equally as delicious cuz of the maggi. im surprised how simple foods taste so amazing.” The meal passed the acceptance test at approximately 450 calories, compared with approximately 2,210 for fifty McNuggets. Salt, umami, spice, warmth, volume and hunger state supplied the pleasure; neither frying nor delivery was necessary to those levers. Nutrition Architecture applies this search within the food system.

The problem becomes finding a lower-damage configuration that satisfies the same request. Substitution removes the confrontation created by denial. Because the craving is satisfied, it does not return doubled an hour later.

Even “part of me wants to sin” can be decomposed into intensity, surrender or relief from being the architect. Those pressure-release states have mostly nondestructive implementations. Moral judgment intensifies the urge, while asking what state is sought makes another route available.

The same applies to requests from other people. The named solution can be the best-advertised configuration rather than the underlying need. Extracting the condition before executing the literal request is the request-side counterpart of Structure Over Request: “The craving specifies the test condition, not the implementation.”

Ergodicity as Life Design

A routine is an engineered energy landscape. Its effectiveness at retaining familiar behavior also confines the states a person encounters:

"Just to champion ergodicity and try to explore other states, kind of break free of your thermodynamic prisons. Even though you engineered the energy landscape, it sometimes [is] good to, you know, kind of get out of that for a while."

From inside that attractor, its states can seem like the whole world: “wake up, go to gym, watch geopolitics, get depressed… that felt like the only reality.” An ergodic process eventually visits its whole state space, making its time average equal to the ensemble average:

$$\lim_{T \to \infty} \frac{1}{T} \sum_{t=1}^{T} f(x_t) \;=\; \langle f \rangle_{\text{ensemble}}$$

The left side represents life experienced one state after another. The right represents the true signal across everything the person could be. A life confined to one routine is non-ergodic: it samples one basin and mistakes that basin's average for the world's.

Eighteen-mile walks at 5:30am, unfamiliar districts, new people and deliberately extreme days sample states the routine hides. These are high-temperature actions: “it took me an understanding of like why I needed to do high temperature behaviors… I needed [to] kind of like stochasticize and get out of my comfort zone.”

The result recalibrates the world model. Will recalled, “I grew up in a suburb of LA, Asian enclave. I just thought everybody's Asian.” A narrow sample had been treated as the population. “Ergodicity, I call it ergodicity. It's called ergodicity, just exploring all states” describes acquiring the variety that corrects it. Cities earn their cost by exposing a person to that range of human behavior.

"The main lesson is to have faith in ergodicity, have faith in your ability to explore states and for the average to emerge and design mechanisms that take advantage of this property."

Repeating the stochastic process allows its average signal to emerge without being designed beforehand: “you just need to repeat this stochastic process, you will eventually explore the average of the state, which is the true signal, the macro state”. Past methods and abandoned phases then become explored states rather than moral failures.

"Usually when I feel like I've done something and I stop doing it, I felt bad to restart doing it, 'cause it felt like admitting that all this other stuff was for waste. But I didn't think this was part of an ergodicity experiment. And now I know it works."

Returning to an earlier practice uses the result of that experiment. It need not imply that trying the alternatives was wasted.

Mean reversion receives the same interpretation:

"Mean reversion happens not because it's a force, but because it's an effect. It's a macro state, which has a lot of different microstates."

A baseline contains combinatorially more configurations than a peak, so random drift returns to it through counting rather than an active force. Deliberate outliers change the data the average must absorb: “you have this ability to generate outlier data… it is healthy to kind of enter those extreme microstates as well. Cause that shifts the average signal.” Even a model trained on Will's logs “doesn't predict my manic states… These are outliers.”

The second intervention is repeated self-selection toward the most extreme version of oneself: “creating your own population and pushing yourself harder in that direction”. This changes the macrostate toward which the counting process returns, until the old mean no longer supplies a destination for reversion.

Failure Modes

The Protocol

A week, day or delegation begins with properties that must hold. Making them visible supplies a feedback signal. Each morning's plan is derived from current information; the evening's review assesses it against the retained target.

The resolver—AI, subconscious, team or day—then supplies the detailed path. Conscious descent into that path is reserved for debugging a poor result, with a deliberate return afterward. Where willpower is unreliable, a menu of acceptable choices lets any selection contribute. Evaluation concerns their aggregate, allowing compensation and alternative combinations.

Regular excursions beyond the established routine complete the protocol. Those outlier days supply data that changes the average, rather than being discarded as anomalies.

Integration with the Mechanistic Framework

Connection to Statistical Mechanics

Boltzmann weighting, energy landscapes, entropy and phase transitions supply the physics vocabulary. Because behavior is governed at the ensemble level, the protocol places the intervention there.

Connection to Order of Determination

The macrostate has the most degrees of freedom and must be fixed first. Plans, schedules and configurations follow from it; fixing those while the target remains open reverses the dependency.

Connection to Container Design

A bounded context specifies the kind of state its occupant will approach. It gives activity a spatial boundary within which to converge.

Connection to Branching and Convergence

Branching explores detailed possibilities. Convergence selects and releases a result, with shipping as the collapse event.

Connection to Selection over Design

Sampling across personal states and selecting from the resulting signal allows a macrostate to emerge without being designed in advance.

Connection to Free Will

One microstate can be forced, but thirty consecutive ones cannot be sustained against the distribution. Engineering the distribution directs agency toward the repeated outcome.

See Also

• Intelligence Is Water explains why unbounded capability spreads without convergence.
• Prevention Architecture removes unwanted options from the menu.
• Probability-Space Bending changes outcomes through their distribution.
• Discretization supplies finite choices within an open process.
• State Machines describes transitions at the microstate level.
• Tracking keeps the feedback record visible.
• Structure Over Request constructs the conditions that produce a requested behavior.
• Nutrition Architecture applies substitution to daily eating.
• Focus keeps attention above delegated search.