Reality Gets a Vote

Reality Gets a Vote

Reality gets a vote. It does not get a dictatorship.

That sentence connects several worlds I had kept in different rooms: scientific method, machine learning, large language models, quantitative research and capital allocation.

They share one quiet operating system:

Prior → evidence → updated belief → action → more evidence → better belief.

In RobinOS language: form a view, observe, adjust, move again.

The loop feels simple. Human beings spend remarkable energy escaping it. We prefer certainty when evidence is thin and surrender when one observation hurts. We marry hypotheses, defend sunk costs and promote recent pain into a universal law.

Bayesian thinking offers a gentler discipline. I have a working belief. Reality is allowed to argue. The size of the update should match the quality of the evidence.

A Scientist Dates the Hypothesis

A scientific theory begins as a prior: an organized belief about how the world works. An experiment creates evidence. Replication changes confidence. A good scientist dates the hypothesis, tests it, and ends the relationship when reality brings enough receipts.

The process contains conviction and humility at the same time.

Conviction is necessary because experiments cost time and attention. Humility is necessary because a beautiful explanation has no voting rights beyond its evidence.

One failed experiment rarely erases an entire theory. It may reveal a flawed instrument, a weak assumption, an uncontrolled variable or a genuine contradiction. The first job is classification.

This is where Bayesian thinking becomes practical. Ask:

• How strong was the prior?
• How reliable is the new observation?
• How likely would this evidence be if the belief were correct?
• Which part of the belief did the evidence actually test?
• What observation would create a larger update?

The questions slow down the emotional leap from surprise to verdict.

Gradient Descent With Feelings

Machine learning performs a version of this loop without drama. A model starts with parameters. Data exposes error. Training nudges the parameters. Repeated updates shape an internal map.

Gradient descent says: I was wrong by this amount; move in that direction.

Large language models make the process visible at another scale. Each token changes the probability landscape for the next token. A sentence emerges through continuous reweighting. It is a tiny probability dragon breathing one word at a time.

Humans add identity. We say, “I was wrong,” and hear, “I am foolish.” The model only updates.

That is one reason Bayesian language can feel joyful. It separates the person from the probability. A belief can move without the self collapsing.

I can hold a strong view and remain available to evidence. I can change my mind without rewriting the past as stupidity. The earlier belief may have been reasonable under the earlier information set.

The Alligator Learns to Distinguish Pain

The Flash Crash Lab forced this lesson into the physical world.

The prior was that a specific pattern of fragility, breakdown and volume shock could identify rare panic dislocations. Historical research supported the idea. Then production produced ugly observations: reporting defects, execution blocks, state mismatch and market paths that hurt.

The emotional update was enormous. The evidence update should have been narrower.

An execution rejection says something about the operating system. A losing counterfactual says something about the setup. A NaN reporting defect says something about evidence quality. A cleanly executed losing trade says something more directly about the strategy.

When every observation receives the same label, the posterior becomes a mood.

The lab improved when each event carried separate states. Signal quality, execution state, evidence status and realized outcome could update different parts of the system.

The strategy no longer needed to survive every bad event. It needed to survive the evidence relevant to its claim.

Capital Allocation Is Belief With Size Attached

Every position is a belief with size attached.

A small position can express weak confidence, high uncertainty or limited risk capacity. A large position expresses stronger conviction and greater responsibility. Cash is a form of humility with optionality.

This framing changes the meaning of action. Buying does not mean certainty. Selling does not mean the original thesis was foolish. Position size can move as evidence, price and opportunity cost change.

Bayesian capital allocation asks two questions together:

1. What do I believe now?
2. How much should that belief be allowed to matter?

Price adds a third. A company can improve while the investment becomes less attractive because the market has already paid for the improvement. A company can disappoint while the expected return improves because expectations moved faster than fundamentals.

Belief, evidence and price must update together.

The discipline protects curiosity. I can study an extraordinary company without needing to own it today. I can own a company while acknowledging the probability of being wrong. I can preserve a research queue instead of forcing every interesting fact into a trade.

One Observation Is a Vote

Imagine a strategy with a long evidence base and one disappointing forward result.

The observation deserves a vote. Its weight depends on data quality, sample independence, execution fidelity and whether the environment matched the strategy’s intended regime. A clean contradiction carries more weight than an ambiguous operational event.

Now imagine thirty clean contradictions across the intended environment. The posterior should move sharply.

Bayesian thinking is not a permission slip to protect a favorite idea forever. It is a rule for proportional surrender.

The same rule applies to people and projects. One awkward meeting is evidence. One failed launch is evidence. One brilliant week is evidence. None deserves to become a total biography without a larger record.

Reality gets votes. The counting method matters.

The Joy of Being Correctable

I used to hear uncertainty as weakness. I now hear it as room.

Room for the next experiment.

Room for a smaller position.

Room for a surprising colleague to be right.

Room for a machine to reveal which part of my intuition survives contact with data.

Correctability keeps intelligence alive. A system that cannot update becomes a monument to its first draft. A person who cannot update becomes a prisoner of an earlier self.

The joy lies in remaining curious after evidence arrives.

You do not need to give reality a dictatorship. Give it a vote, record the vote honestly, and return for the next one.

A small field guide for updating without panic

The first step is to write the prior before the new evidence arrives. A prior does not need a precise decimal probability. It can be a bounded statement: high confidence, working belief, open question, weak hypothesis. The act of writing prevents the later mind from pretending it always knew the result.

The second step is to name what the new observation can test. A customer cancellation may test willingness to pay. It may say little about product utility if procurement changed. A losing trade may test one setup in one regime. It says little about a different regime. An employee departure may test team health, compensation, personal circumstance or all three. Evidence needs a jurisdiction.

The third step is to grade evidence quality. Direct measurements deserve more weight than recollection. Repeated independent observations deserve more weight than one correlated cluster. Point-in-time records deserve more weight than a narrative written after the outcome. A source with incentives can still be useful; its incentives belong in the weight.

The fourth step is to state the update in words. “My confidence moved from high to moderate because the expected operating proof did not appear.” “The product thesis remains intact; the distribution assumption weakened.” “The strategy evidence is inconclusive; the execution defect is confirmed.” Clear language prevents a numerical costume from hiding a vague thought.

The fifth step is to resize action. A belief update that leaves behavior unchanged may be intellectually decorative. Change the experiment, position, deadline, budget, control or monitoring trigger in proportion to the evidence. The action can be small. A small update often deserves a small move.

The sixth step is to define the next evidence before waiting for it. What would raise confidence? What would lower it? When will the observation arrive? Who owns collection? This converts uncertainty from fog into a research queue.

The seventh step is to preserve the old view. Keep the prior, evidence and update together. A history of changing beliefs becomes an asset. It reveals whether you react too strongly to recent events, protect favored ideas, ignore base rates or demand impossible certainty before acting.

I also use three emotional checks.

Am I trying to make the pain stop? Pain can create an update larger than the evidence.

Am I trying to protect an identity? “I am a good investor” and “this investment is good” are separate propositions.

Am I still curious about the next observation? Curiosity is often the first casualty of shame and certainty.

The best posterior is rarely the cleverest sentence. It is the belief that leaves you correctly sized, still observant and ready to change again.

A boardroom version

The same method works when the belief belongs to a company rather than one person.

Imagine a product team that expects a new feature to improve retention. The launch produces strong usage and flat renewal. The team can defend the thesis, declare defeat or separate the observations. Usage supports product curiosity. Renewal disconfirms the immediate commercial bridge. The next experiment should test whether the feature attracts the wrong cohort, arrives too late in the journey or solves a problem customers enjoy without paying to solve.

An investment committee can do the same. A company delivers revenue growth and weaker cash conversion. The growth thesis receives support. The financing or margin thesis receives pressure. Position size can change before the entire company judgment becomes binary.

A research lab may see a strategy behave well before fees and poorly after fees. Signal quality and business quality then carry different posteriors. The correct action may be to preserve the signal research while blocking capital until execution economics improve.

This decomposition keeps reality constructive. Evidence can close one path and open a better one. A failed bridge tells the team where value stopped moving.

The boardroom tool needs a compact record:

1. prior belief and confidence range;
2. dated observation with provenance;
3. which causal link the observation touches;
4. revised range rather than one theatrical probability;
5. proportional action;
6. next evidence and owner.

The record makes updating legible across time. It also prevents seniority from becoming a substitute for likelihood. The loudest voice can propose a prior. Reality retains the deciding vote.

The language of an honest update

I prefer updates that describe direction and consequence before pretending to offer precision. “Confidence moved from strong to mixed; reduce the experiment and test retention by cohort” is often more useful than a probability with no calibration history.

The sentence should name the observation, the affected causal link and the action. It should also preserve what remained stable. One disappointing metric may reduce confidence in conversion while leaving customer love or technical feasibility intact.

This language keeps teams from treating every update as a public confession. Learning becomes normal operating behavior. People can change their view early, while the cost of change is still small.

The investment transfer

An investment journal can implement the method with four dated lines: prior, new fact, changed implication and position response. The entry stays short enough to write before emotion edits the memory. A later review can compare the actual update with what the investor claimed would change the view.

Over time, this reveals personal calibration. Some people update too quickly on price and too slowly on business evidence. Others protect a thesis until the exit becomes expensive. Reality keeps voting; the journal shows whether the portfolio counted the ballots.

The discipline is gentle, cumulative and difficult to fake after the fact.

Source note

This essay draws from Robin and Teddy’s conversation on Bayesian thinking and from sanitized Flash Crash Lab learning. It states a decision philosophy rather than a trading recommendation or claim of current performance.

#BayesianThinking #DecisionScience #CapitalAllocation #RobinOS #Joy