Then prices and incomes rise, and the amount stays where it was. I call these amounts "magic numbers." A magic number is a dollar figure written into law that only a new act of Congress can change.

The federal minimum wage is the plain example. It has been $7.25 an hour since July 24, 2009. Congress has not passed a new bill since then. Meanwhile the typical full-time worker's pay rose by about 70 percent.

The current approach is slow. Each update needs a new bill, a new fight, and a new vote. Between those votes, the number loses meaning a little each year.

Congress has fixed part of this problem. Many amounts now rise each year with a price index. But prices tell us what a basket of goods costs, not what people earn.

My argument is simple. Congress should tie each of these amounts to a percentage of median income, or to the average wage where no median exists. It should also build a free national payroll service, so that the income data is current and complete.

Two kinds of magic numbers

The dollar amounts in federal law fall into two groups. The first group is frozen, and only a new bill can move it.

The second group rises each year, but by the wrong measure. Most follow the Consumer Price Index (CPI) or its "chained" version. One follows the average wage and shows the better approach.

Some magic numbers are not dollars at all. Federal law also fixes ages, such as the Social Security retirement age. Those ages should follow life expectancy. I cover them after the dollar proposals.

Numbers the law has frozen

Federal minimum wage. The Fair Labor Standards Act sets the wage floor for most workers. It has been $7.25 since July 24, 2009. When the 1991, 1997, and 2009 increases took effect, the minimum equaled about 40 percent of median full-time pay. Today it equals about 23 percent.

Tipped minimum cash wage. Employers may pay tipped workers $2.13 an hour in cash and count tips toward the rest. That figure dates from 1991, when it was half the minimum wage. A 1996 law fixed it at $2.13 in dollars, and it has not changed since.

Overtime salary threshold. Salaried managers and professionals are exempt from overtime only if they pass a duties test and earn at least a set salary. The threshold is $684 a week, or $35,568 a year. The Department of Labor set it in 2019. A 2024 rule raised it, but a federal court vacated that rule in November 2024. In May 2026 the department formally restored the 2019 level.

Capital loss deduction. Investors who lose money can deduct up to $3,000 of net losses against wages each year. That limit took effect for tax years after 1977. Adjusted for prices, $3,000 from 1978 equals about $15,400 today.

Taxation of Social Security benefits. Retirees pay income tax on part of their benefits once "combined income" passes $25,000 for a single filer or $32,000 for a couple. Congress set those levels in 1983. It added a second tier at $34,000 and $44,000 in 1993. None of these amounts has ever changed. In 1983 Congress aimed this tax at upper-income retirees. Because the amounts never move, the tax reaches further down the income scale each year.

Net Investment Income Tax and Additional Medicare Tax. Both taxes began in 2013. Both start at $200,000 for single filers and $250,000 for joint filers. Neither threshold adjusts. Adjusted for prices, $200,000 from 2013 equals about $287,600 today.

Home-sale exclusion. A homeowner can exclude up to $250,000 of gain on the sale of a main home, or $500,000 for a couple. Congress set these caps in 1997 and did not index them. Adjusted for prices, $250,000 from 1997 equals about $521,700 today.

Supplemental Security Income (SSI) asset limit. SSI pays cash to people with low income and few assets who are elderly, blind, or disabled. A person cannot have more than $2,000 in countable assets, or $3,000 for a couple. These limits have not changed since 1989. Adjusted for prices, $2,000 from 1989 equals about $5,400 today.

State and local tax (SALT) deduction cap. The 2025 tax law (Public Law 119-21) raised the cap from $10,000 to $40,000 for 2025. It sets $40,400 for 2026 and adds 1 percent a year through 2029. In 2030 the cap drops back to $10,000. A fixed 1 percent step is not an index. In 2030 the law returns to a frozen $10,000.

Information reporting (Form 1099). For decades, businesses had to report payments of $600 or more to contractors. The 2025 tax law raised this to $2,000 for payments after 2025 and indexes it after 2026. Congress needed decades and a large tax bill to move one small number.

Numbers indexed to the wrong thing

Several amounts already rise each year without a new bill. A fair case must say so. The problem with these is not that they are frozen. The problem is the measure they follow, or the base they started from.

Tax brackets and the standard deduction. These rise each year with the chained CPI, under the 2017 tax law. The 2025 tax law made those brackets permanent. For 2026 the standard deduction is $16,100 for a single filer and $32,200 for a couple. The 22 percent bracket starts at $50,400 for a single filer.

Gift tax annual exclusion. A person can give $19,000 to each recipient in 2026 without filing a gift tax return. This amount follows prices and rises only in $1,000 steps. It stayed at $19,000 from 2025 to 2026.

Estate tax exemption. The 2025 tax law set the exemption at $15 million for 2026. After 2026 it rises with the chained CPI.

Poverty guidelines and thresholds. The 2026 federal poverty guideline for a family of four is $33,000. The Census Bureau's official threshold for the same family in 2025 was $32,970. Both rise each year with the CPI. But the base dates from 1963 and 1964. Mollie Orshansky, an analyst at the Social Security Administration, built it from the cost of the cheapest government food plan. She multiplied that cost by 3, because families of three or more had spent about one-third of their after-tax income on food in 1955. Since then, the thresholds have mostly just followed prices.

SNAP and other benefit limits. Many programs set eligibility as a percentage of the poverty guideline. SNAP, formerly food stamps, generally requires gross income at or below 130 percent of it. So these limits inherit the guideline's 1960s base. SNAP's asset limits are $3,000, or $4,500 for households with an older or disabled member, for fiscal year 2026. They rise with prices.

Medicare premium surcharges (IRMAA). Medicare charges higher Part B premiums to higher-income people. For 2026 the surcharge starts above $109,000 for single filers and $218,000 for couples. The lower brackets rise with the CPI. The top bracket, at $500,000 or $750,000, stays fixed until 2028.

Medicare Part B standard premium. This one is not a magic number. A formula sets it to cover about 25 percent of expected costs for an aged enrollee. For 2026 it is $202.90 a month. A formula tied to actual costs does what this essay asks for, so I leave it alone.

Social Security earnings test. Social Security withholds some benefits from people who claim early and keep working. In 2026 the limit is $24,480 a year, or $65,160 in the year a person reaches full retirement age. These limits rise each year with the national average wage index. This is the model. The law already uses a wage measure here.

Why fixed numbers fail

A fixed number forces Congress to act again and again. Each update needs time in both chambers, enough votes, and a bill to carry it. Most years these do not line up, so the number drifts. The minimum wage has now gone 17 years without a change. That is the longest gap since Congress created it in 1938.

The drift is not neutral. When a benefit limit stays fixed, fewer people qualify each year. When a tax threshold stays fixed, more people pay each year. Nobody votes for these changes. They happen by default.

The drift also invites large, sudden corrections. The 2024 overtime rule tried to catch up in two big steps and did not survive in court.

An index removes most of this work. Congress sets a rule once, and the number moves each year without a new bill. Congress can still change the rule.

Why the CPI is the wrong foundation

The CPI is a careful statistic. The Bureau of Labor Statistics (BLS) builds it to measure the average change in prices for a basket of goods and services. That is a useful question. But it is not the question most thresholds ask.

A threshold asks where a person stands compared with everyone else. Is this family poor? Is this income high enough to pay a surtax? Those are questions about income, not prices.

The CPI also has specific features that make it a weak base for law.

It builds in substitution. Since January 1999, the CPI has used a formula that assumes people switch to cheaper items within a category. The chained CPI goes further and assumes switching across categories. If beef costs more and people buy pork, the chained index rises less than a fixed basket would. From 2001 to 2023, the chained CPI grew about 0.2 percentage points a year slower than the regular CPI-U. Tax brackets use the chained version.

It measures owned housing by imagined rent. For homeowners, the CPI does not use home prices or mortgage payments. It uses "owners' equivalent rent," which follows what similar homes rent for. That item carried about 26 percent of the whole index in December 2025. A young family trying to buy a first home faces prices the index does not track.

It adjusts for quality. When a product improves, the BLS removes the value of the improvement from the price change. Its own example shows a $250 television replaced by a $1,250 model. After adjustment, the index records a 7.1 percent price drop. That is a fair answer to a price question. It says little about what a household must spend.

Its methods change over time. In 1983 the BLS switched owned housing from purchase costs to rental equivalence. It fixed a formula problem in 1995 and 1996. In 1996 a commission appointed by the Senate Finance Committee estimated that the CPI overstated inflation by about 1.1 percentage points a year. The BLS adopted the geometric mean formula in 1999.

Each change may improve the index. But the BLS does not rewrite the published CPI-U to match. So a CPI point in 1980 and a CPI point in 2020 rest on different methods. The BLS publishes a research series, the R-CPI-U-RS, that applies current methods back to 1978. It gives a different history. From 1978 to 1982, the CPI-U rose 57.2 percent, and the research series rose 50.1 percent.

The BLS revises the chained CPI. The regular CPI-U is final when published, except for seasonal factors. The chained CPI is not. The BLS publishes it in preliminary form and revises it three times. Final values arrive 10 to 12 months later. This is the version the tax code uses.

It depends on field surveys. The October 2025 lapse in federal funding stopped data collection. The BLS published no CPI-U value for October 2025. The same lapse left no BLS median earnings figure for the fourth quarter of 2025.

None of this means the BLS does poor work. The CPI answers a different question from the one thresholds ask. And because the method keeps changing, the same number means different things across decades.

Why income data is better

Income data measures what the law cares about directly. It shows what people earn, and it shows where the middle is.

The best income data comes from records, not surveys. Employers file a Form W-2 for each worker they pay. The Social Security Administration (SSA) uses those forms to build the national average wage index. For 2024 the index is $69,846.57. The 2023 records covered about 174 million wage earners.

The same records give a median. For 2023, half of all wage earners had net compensation at or below $43,222.81. SSA publishes this figure every year.

Census and BLS surveys add useful cross-checks. The Census Bureau reports a 2025 median household income of $87,460. The BLS reports median weekly earnings of $1,251 for full-time workers in the second quarter of 2026. These surveys capture household income and hours, which W-2 forms lack. But surveys depend on response. The 2026 Census income survey had a response rate of 61.3 percent.

The law already trusts wage records. The average wage index updates the Social Security earnings test, the benefit formula, and the maximum earnings subject to Social Security tax.

Income measures have real drawbacks. W-2 data arrives late. Social Security, for example, indexes a worker's earnings to the wage index from 2 years before eligibility. An average can rise because top earners gain, even when the middle does not. In 2023, about 67.6 percent of wage earners earned less than the average. I address these points in the counterarguments below. The short answer is to prefer medians and to fix the lag.

A free national payroll service

Indexing to median income needs current median income data. Today the United States gets that data late and in pieces. Employers report each worker's wages to the federal government once a year, on Form W-2. Surveys fill the gaps with samples.

I propose a better source. The federal government should build a free national payroll service and require every employer to use it.

What it would do

The service would run payroll for every employer in the country, free of charge. Each pay period it would record wages, hours, deductions, withholding, and employer contributions. It would pay workers, send taxes where they belong, and issue pay statements.

Other countries already do parts of this. The United Kingdom's Real Time Information rules began on April 6, 2013. Employers must report pay and tax to HM Revenue and Customs each time they pay workers. Australia's Single Touch Payroll became mandatory on July 1, 2018 for employers with 20 or more employees. It covered smaller employers from July 1, 2019.

Italy has required electronic invoices between businesses, and to consumers, since January 1, 2019. Each invoice passes through a central exchange run by the revenue agency. Mexico requires an electronic tax receipt, the CFDI, for invoices. Since 2014 employers must also issue one for each payroll payment.

These are mandatory reporting rules. Employers still run payroll through their own software or a private provider. The UK and Australia receive the data. They do not process the payroll. My proposal goes one step further. The government would run the payroll itself, as a free public service.

What the data would show

A national payroll service would give the country a complete, current record of paid work. It would show employment, hours, and pay by industry, region, and occupation. It would show the median and average wage each month.

The UK already shows what this looks like. HMRC and the Office for National Statistics publish payroll statistics each month. On September 15, 2026, they reported 30.2 million payrolled employees for August 2026. They reported median monthly pay of £2,657. That release came about two weeks after the month ended.

The UK figures are not perfect. The first estimate for each month uses about 85 percent of the data, and the next release revises it. Most revisions fill in late records rather than change the method.

With such data, the United States could compare every threshold with actual pay each year. The two-year lag would shrink to weeks. Sampling error from surveys like the Current Population Survey would largely disappear for wage data. Hours would be on record for every worker, so an hourly median would rest on records too.

The end of the annual return for most wage earners

A national payroll service leads to a second benefit. If the government already sees every paycheck, deduction, and withholding, the annual tax return becomes unnecessary for most wage earners.

The UK shows how. Under Pay As You Earn, employers usually apply each worker's tax code on a cumulative basis. Each payday, the calculation looks at pay and tax so far that year. HMRC says this makes the year's tax "roughly correct for most cases."

As a result, most UK employees never file a return. HMRC expected about 12 million Self Assessment returns for the 2024 to 2025 tax year. The UK has about 30 million payrolled employees. Many of those 12 million filers are self-employed people or landlords.

Sweden takes a different route to a similar result. Employers, banks, and the social insurance agency report income to the tax agency. The agency fills in each person's return in advance. People who need no changes can approve it by e-service, text message, or phone call. Australia's payroll reports also mark each worker's income statement "Tax ready" for filing.

US withholding is less precise. Employers withhold from each paycheck based on a form the worker fills in, and most workers settle up on an annual return. The IRS estimates that the average nonbusiness taxpayer spends $160 and 8 hours filing.

A national payroll service would change that. With every paycheck on record, the IRS could compute each wage earner's tax for the year. It would then send a bill or a refund. The worker could check the numbers and correct them.

This step depends on the mandatory payroll service. Today the IRS gets most wage data on W-2 forms after the year ends. Only complete, timely payroll records let the IRS compute the tax itself. For most wage earners, the annual return would go away.

The United States tried a smaller step. The IRS Direct File program let some taxpayers file free on the IRS website in 2024 and 2025. Treasury suspended it in October 2025, citing cost and low use. Filers sent only about 309,000 returns through it in 2025. A pre-computed bill would not depend on people seeking out a tool. It would arrive on its own.

Concerns and responses

Privacy and security. A single payroll service would hold sensitive data on nearly every worker. That demands strict rules. Statistical agencies should receive only de-identified records. The law should bar any other use of the statistical files.

The United States has models for this. Title 13 makes census responses confidential. Disclosure can bring up to 5 years in prison and a $250,000 fine. The Confidential Information Protection and Statistical Efficiency Act (CIPSEA) protects data collected for statistical purposes.

I should be honest about one point. The same service would also compute tax withholding. So it cannot sit fully behind a statistical wall. The law must separate the tax function from the statistical function and state who may see what.

Small-business burden. The service would be free. For many small employers, it would replace a fee they already pay to a private payroll firm. It could also replace separate wage filings to several agencies. The UK offered free reporting software to employers with 9 or fewer staff when Real Time Information began.

Political feasibility. Payroll and tax preparation firms would resist. The history of free filing shows this. In 2002 the IRS agreed with private tax software firms not to build its own filing tool. That clause stayed in place until 2019.

A phased path would help. First, offer the national payroll service as a free public option. Second, require every employer to report each pay run, as the UK and Australia do. Third, require use of the national service once it has proven itself.

What payroll data cannot do

Payroll covers wages only. It does not capture self-employment income, business profits, rents, or investment income. Self-employed people, investors, landlords, and people with complex situations would still file returns.

Tax returns would also remain the source for household income, since households combine several earners and several kinds of income. Payroll data gives the timely wage signal. Tax returns and Census surveys give the fuller household picture.

This service ties the whole proposal together. Congress cannot index the law to median income without timely median income data. A free, mandatory national payroll service is how the country gets it.

Proposals

Each proposal below ties a number to one of four measures:

• Median full-time weekly earnings, from the BLS: $1,251 in the second quarter of 2026. Divided by 40, that is $31.28 an hour.
• Median wage, from SSA W-2 records: $43,222.81 for 2023.
• Average wage index, from SSA: $69,846.57 for 2024.
• Median household income, from the Census Bureau: $87,460 for 2025.

The national payroll service would later replace the survey-based measures with records. Where a formula gives less than current law, the current amount should stay until the formula catches up.

The reasons for each choice are short.

• Minimum wage at 40 percent. When the 1991, 1997, and 2009 increases took effect, the minimum equaled 40.3, 41.3, and 39.3 percent of median full-time pay.
• Tipped wage at 50 percent. This restores the split that held in 1991, before the 1996 law froze the dollar amount.
• Overtime threshold at 72 percent. The 2004 rule took effect at 72.0 percent of median full-time pay. The 2019 rule took effect at 71.5 percent. This level is well below the $1,128 step that the court rejected.
• Poverty guideline at 38 percent. This keeps today's line as the start. After that, the line tracks typical living standards instead of a 1963 food budget.
• Asset limits at one month of pay. A small emergency fund should equal about a month of typical earnings. For SSI, this roughly equals the 1989 limit adjusted for prices.
• Tax thresholds at 2026 ratios. Rising real wages would no longer push people into higher brackets.
• Gift, estate, IRMAA, and SALT amounts at current ratios. These start from today's levels, then grow with typical incomes. The IRMAA top bracket should not wait until 2028.
• Capital loss limit at 25 percent of the median wage. This restores most of the lost value. It stays below full price restoration, about $15,400, because losses already get favorable treatment compared with gains.
• Social Security benefit tax at half of median household income. This moves toward the 1983 intent without the full revenue cost of price restoration, about $84,100.
• NIIT and Additional Medicare Tax near current levels. Congress aimed these at high earners. Starting from today's ratio avoids a revenue windfall and keeps the affected group the same size.
• Home-sale exclusion at 6 times median household income. This roughly restores the 1997 value, about $521,700 in today's prices.

Magic numbers that are not dollars

Some magic numbers are ages. Dollar amounts should follow median income. Ages should follow life expectancy, the statistic that matters for an age. The principle is the same: set the rule once, not the number.

Social Security full retirement age. The age for a full benefit was 65 until the 1983 amendments raised it in steps. It is 67 for people born in 1960 or later. Workers can still claim a reduced benefit at 62. Each year of delay adds 8 percent, up to age 70.

The Social Security Trustees report period life expectancy at 65. In 1940, it was 11.9 years for men and 13.4 years for women. In 1983, it was 14.3 and 18.6 years. In 2024, it was 18.3 and 20.8 years.

Since 1983, men gained 4.0 years at 65 and women gained 2.2 years. The full retirement age rose 2 years. But the law wrote 67 as a fixed number, so it stops there.

Other countries use a rule. Denmark has tied its pension age to life expectancy since 2006. In May 2025, its parliament voted to raise the age to 70 in 2040. The governing party wants a new model before the next vote in 2030. Sweden sets its target age at 65 plus two-thirds of the gain in life expectancy at 65 since 1994. Finland ties the age to life expectancy for people born in 1965 or later, with each change capped at 2 months a year. The Netherlands also links its age to life expectancy at 65.

My proposal follows Sweden. Start from 67 in 2027, when the first people born in 1960 reach that age. Then raise the full age by two-thirds of each later gain in life expectancy at 65. In the Trustees' intermediate projection, life expectancy at 65, averaged for men and women, rises about 1.4 years from 2027 to 2050. Two-thirds of that gain is about 11 months. Age 70 would move by the same amount.

I would keep age 62 fixed. A higher full age cuts benefits at every claiming age, and the gains have not been equal. Chetty and colleagues found that from 2001 to 2014, life expectancy at 40 rose 2.34 years for men in the top 5 percent of incomes. It rose 0.32 years for men in the bottom 5 percent. People in hard jobs need an exit at 62, even with a reduced benefit.

Medicare eligibility age. Medicare has started at 65 since President Johnson signed it on July 30, 1965. Since 1965, life expectancy at 65 has risen 5.4 years for men and 4.5 years for women.

The tradeoff is real. The Urban Institute studied a move from 65 to 67 and found federal savings of $24.1 billion. But total health spending for people aged 65 and 66 would rise from $115.7 billion to $133.1 billion. Employer premiums for this group would rise from $11.1 billion to $37.7 billion. So I would keep Medicare at 65 for now. Congress should first guarantee affordable coverage for anyone in the gap. After that, Medicare could follow the same rule as Social Security.

Required minimum distribution age. Tax-deferred retirement accounts must start paying out at a set age. Age-based payouts entered the tax code in the 1960s. The SECURE Act raised the age from 70½ to 72 in 2019. SECURE 2.0 raised it to 73 and sets 75 for people born in 1960 or later. That took 2 bills and 3 changes. The age should follow the same two-thirds rule, starting from 75.

One safeguard for every age. No age should change fast or by surprise. Each step should be at most 2 months per birth year. Congress should announce each new age at least 10 years ahead. Denmark already gives 15 years' notice.

Counterarguments

"Congress should control wages directly." Congress would still control them. It sets the percentage and can change it by law at any time. Indexing only changes the default. Today the default is steady decline. Under this proposal the default is stability.

"Costs differ by region." They do. A federal amount is a national floor, and states can go higher. As of July 2026, 34 states and the District of Columbia set tipped cash wages above $2.13. Programs already use separate poverty guidelines for Alaska and Hawaii. The national payroll service would make regional medians possible where Congress wants them.

"The transition would be abrupt." It need not be. Congress could phase each new formula in over 3 to 5 years. Where a formula gives less than current law, the current amount stays until the formula passes it. No one would face a sudden cut.

"Income falls in recessions, so thresholds would fall." This is a fair concern. The average wage index fell from $41,334.97 in 2008 to $40,711.61 in 2009. The fix is a no-decrease rule. Each amount can rise or stay flat, but not fall. A 3-year average would also smooth sharp swings.

"Income data arrives late." Today it does. As of September 2026, the latest wage index is for 2024, and SSA has not yet posted the 2024 median. A lagged or smoothed value works in the meantime, as Social Security already shows. The national payroll service would cut the lag to weeks.

"An average rises with top incomes." That is why I prefer medians. SSA already publishes a median wage from W-2 records. Where the law uses an average, it should switch to the median when the data allows.

"Wage indexing costs money." Compared with current law, tying tax thresholds to wages would reduce revenue over time. Real wages usually grow. But price indexing is itself a policy choice. It raises taxes a little each year without a vote. Congress can set rates to raise whatever revenue it wants. It should do that openly.

Conclusion

Magic numbers are a slow, quiet way to change policy. Each year that Congress does not act, benefits shrink and taxes rise. Nobody votes for the change.

Price indexing fixed part of this problem. But the CPI measures the price of a basket of goods. The law needs to know where people stand compared with everyone else. That calls for income.

Congress should pass a law that ties each of these amounts to a percentage of median income. Where a median is not yet available, the law should use the average wage and move to the median later. Every amount should carry a no-decrease rule.

Congress should also create a free national payroll service. It would give the country current, complete wage data. It would let the IRS send most wage earners a bill or a refund instead of a return to fill out.

The same idea works for ages. Retirement and distribution ages should follow life expectancy through a set rule, not a new bill each time.

Congress would still set every percentage. It would stop setting the same numbers over and over.

A large public school is a different environment. It gathers hundreds or thousands of children of nearly the same age. It holds them together for years. It ranks them in public, every day, on a short list of visible traits. The child cannot leave that ranking and enter another social world for most of the day. That setting is recent. It is also formative.

Size with no exit

A common American high school holds 1,000 to 2,000 students. Many schools are larger. For most of history, most human settlements were smaller than that. A medieval village often held a few hundred people at most. Stable personal relationships in ordinary life also tend to stay in a much smaller range than a modern school roster.

The important point is not a precise number. The important point is the mismatch. A child spends long days inside a crowd that exceeds the scale of the communities that shaped human social life for a very long time. The child also lacks an exit. Home is not an alternate ranking during school hours. The peer crowd is the main stage.

One ranking, many years

Inside that crowd, status collapses into a few ladders: athletic skill, physical development timing, social dominance, fashion, and academic display. Those ladders are easy for teenagers to see. They are performed in hallways, classrooms, teams, and social groups.

Adult life uses different measures. Usefulness at work, care for dependents, judgment under pressure, and craft over years matter more later. A child at the bottom of a school ranking does not know that yet. The ranking feels total while it lasts. Four years of low status, with no second stage to stand on, can teach a person that the self is low-status as a fact. The lesson can remain after the school is gone.

Children at the top of the same ranking can also carry lasting effects. The skills that win inside a large same-age crowd do not always transfer. When the crowd disperses, the old scoreboard disappears. Adult life can then feel like a loss of a map that once worked.

What the environment removes

A small mixed-age community gives a child several kinds of status at once. A person can fail at one task and still hold respect in another. A twelve-year-old spends time with younger children, older youths, and adults. The child sees a path through ages, not only a contest among peers.

A large age-sorted school removes much of that. Peers become the main mirror. Adults appear mainly as managers of the day, not as people doing ordinary work beside the child. The child's sense of self is then built under a peer ranking that is hard to escape and easy to over-weight.

Why this matters later

The claim is narrow. Large same-age schooling is an unnatural social environment for a developing person. The hierarchy it creates has an outsized long-term impact on how people see themselves.

That impact is not the only force in a life. Family, temperament, class, and later events also shape the self. The observation still holds. Many adults still narrate themselves through a high-school position they no longer occupy. They speak as if that ranking were a permanent verdict. An environment that can write a verdict that strong deserves to be named as unusual, not as neutral background.

Limits

This essay does not argue for a single replacement model. It does not claim that every large school ruins every child. It does not treat the past village as a paradise. It states a mismatch: scale, age sorting, public ranking, and no exit, held for years during a period when the self is still being formed.

The mismatch is the observation. The lasting mark on self-concept is the cost.

Nobody issues a manual at age 18. Schools do not teach these responsibilities as a subject. The people who can pay advisors hand the work to someone else. The people who cannot pay advisors carry the full load. The overhead is large, undocumented, and regressive.

The list is the argument

A working adult is expected to know a long set of acronyms and duties. That adult may also hire someone who knows them. A partial list is enough to show the load:

• Credit and lending: FICO, APR, HELOC
• Tax and identity: W-2, 1099, AMT, REAL ID, ID.me, Selective Service
• Health: HSA, FSA, HRA, COBRA, EOB, deductible, network, prior authorization, open enrollment
• Retirement: 401(k), Roth IRA, SEP-IRA, RMD, Social Security claiming age
• Housing and transport: PMI, property tax, vehicle registration, driver's license renewal
• Education and voting: FAFSA, voter registration
• Estate documents

Each item has its own deadlines, forms, exceptions, and penalties. The items do not share one office. The items do not share one calendar. The list itself is the claim: the number of ordinary adult responsibilities is too large for one person to hold without help.

Two kinds of overhead

The first kind is legal duty. An adult must file federal taxes. In many states an adult must also file state taxes. An adult who drives must keep a valid license. REAL ID rules add federal identity requirements. A vehicle must be registered. A homeowner must pay property taxes. A male citizen must register for Selective Service within a short window after age 18. Failure can block federal jobs, student loans, and job training. Access to IRS, VA, and Social Security online services often requires identity checks through private vendors such as ID.me. Permits and licenses appear again for home work, business activity, and many professions.

The second kind is optional on paper. It is mandatory in effect. Health insurance requires plan choice, deductible tracking, network limits, prior authorizations, and review of EOBs. Missing open enrollment can leave a household exposed for a year. Credit scores from Equifax, Experian, and TransUnion decide access to housing, employment checks, car rentals, and insurance prices. Retirement accounts reward early contributions through compound growth. A person who misses the window in their twenties cannot buy those years back later. HSA and FSA rules add contribution caps and use-it-or-lose-it deadlines. Life insurance and disability insurance sit in the same group. Social Security claiming at 62 versus 70 is a large lifetime money decision. Many people make it with no analysis.

The second kind is where the regressive pattern is sharpest. The people with the least spare time pay more for a miss. The people with the least spare money pay more for a miss. The people with the least access to advisors pay more for a miss. A person can work for decades, break no law, and still arrive at old age with a weak position. The cause is not a single bad character trait. The cause is a set of responsibilities that punish missing knowledge.

One move, many duties

A concrete test of the overhead is a move between states.

The person must notify the postal service. The person must get a new driver's license inside a state deadline. The person must register the vehicle with a different authority. The person must update voter registration. The person must update the IRS, the Social Security Administration, the employer, the bank, the credit cards, and the insurers. The federal government already has an address on tax filings. These offices still do not update each other.

One life event becomes a checklist across agencies that do not coordinate. Each miss has its own penalty schedule. That is overhead.

Schools do not teach these responsibilities

Public schools teach subjects that map to exams and college entry. They rarely teach the practical duties of adult life.

A graduate may know algebra and essay structure. That graduate may not know how a deductible works. That graduate may not know how a credit report is used. That graduate may not know how a W-2 differs from a 1099. That graduate may not know how open enrollment works. That graduate may not know what happens if Selective Service registration is missed. Civics class may cover the branches of government. It usually does not cover the annual checklist that keeps a household solvent and legal.

This gap is not a minor curriculum preference. It is a transfer of risk. Institutions moved large parts of retirement and health risk onto individuals in the late twentieth century. Pensions became 401(k) plans. Many health arrangements became marketplace choices that the household must manage. The knowledge required to manage that risk did not move into the classroom at the same time. Students leave school prepared for the old subjects and unprepared for the new responsibilities.

Downstream effects

The first effect is time. Each responsibility claims hours for setup, renewal, dispute, and correction. The hours come from work, sleep, or family life.

The second effect is error. A large set of duties produces missed deadlines. Missed deadlines produce fees, coverage gaps, and lasting marks such as credit damage.

The third effect is compounding. Retirement contributions, credit history, and insurance continuity reward early correctness. Late correction does not restore the path that early correctness would have produced.

The fourth effect is unequal access to help. Households that can buy accountants, brokers, and lawyers turn overhead into a paid line item. Households that cannot buy help turn overhead into risk. The same formal rules then produce different life outcomes. That is what "regressive" means here. The tax is not only money. The tax is time, error, and lost compound growth, and it falls hardest on the people with the least margin.

The fifth effect is learned helplessness or its opposite: a market of helpers. Both are responses to a set of responsibilities that an ordinary person cannot keep fully in mind.

Limits

This essay does not claim that every rule is useless. Taxes, licenses, and insurance exist for reasons. The claim is narrower. The number of separate responsibilities is high. Coordination between the offices that create them is weak. Schools do not prepare people for the load. The overhead exceeds what a typical adult can carry without help.

This essay also does not offer a redesign of government. The diagnosis is the load, the size of the load, the school gap, and the regressive downstream effects. Those facts stand before any debate about fixes.

Conclusion

Being a functional adult in the United States requires continuous attention to a large set of legal duties and private necessities. Schools do not teach them. Money buys relief from them. The people without money absorb the failures as lifelong cost.

There is too much overhead.

This post first states the case. Then it states the strongest objections. The remainder is what still stands.

The layers

Physics has a stack:

• Quantum mechanics
• Atomic structure
• Molecular chemistry
• Biology
• Cognition

Software has a stack:

• Transistors
• Logic gates
• Operating system
• Runtime
• Application

graph LR
  subgraph physics["Physics stack"]
    Q[Quantum] --> A[Atomic] --> M[Molecular] --> B[Biological] --> C[Cognitive]
  end
  subgraph software["Software stack"]
    T[Transistors] --> G[Logic gates] --> O[Operating system] --> R[Runtime] --> AP[Application]
  end
  Q -.-> T
  A -.-> G
  M -.-> O
  B -.-> R
  C -.-> AP

A layer does not need to know how the layer below it works. Chemistry does not need quantum field theory. A web application does not need transistor physics. This property is abstraction.

The universe was not required to have this property. A universe can be a system where every scale depends on every other scale. This universe is layered. A layered system lets each part be built and tested on its own.

Level of detail is in the laws

In a game engine, level of detail is an approximation. A mountain at a large distance uses a simple mesh. Full geometry loads only when the player is near. This saves compute. An error in the transition shows a seam.

Physics has a similar mechanism. The difference is that the mechanism is not an approximation.

Newton's shell theorem states a fact about gravity. From outside, a spherically symmetric body has the gravity of one point that holds all of its mass. The result is exact. General relativity has the same property in Birkhoff's theorem. The center-of-mass theorem states a related fact. A system of 10^50 particles, under external forces, moves as one point.

This is why NASA can fly a spacecraft for a decade and hit a small target. The calculation does not simulate one atom of Jupiter. The universe supplies the cheap aggregate. The cheap aggregate is correct. Game engines add level of detail and then show errors. The laws come with aggregation already in place.

A mathematician can object. The shell theorem is not an added feature. It follows from the inverse-square law. The inverse-square law follows from flux conservation in three-dimensional space. Exact aggregation follows from geometry and from linear laws. This objection is fair. It moves the question. It does not remove the question. Why are the laws linear, local, and symmetric enough for this to work in all cases? Nothing required that property. The unusual fact is not the theorem. The unusual fact is the set of constraints from which the theorem follows.

The extreme case is the no-hair theorem. A black hole forms from matter of arbitrary complexity. To the outside, it exposes three properties: mass, charge, and spin. In software terms, this is a class with private fields and three accessors.

The universe is sharded

The speed of light is often described as a speed limit. A more precise frame exists. In Planck units, the speed of light is one Planck length per Planck time. That rate is one cell per tick. Researchers of cellular automata have used this term for a long time. In Conway's Game of Life, the fastest influence moves one cell per generation. The literature calls that limit the speed of light.

The speed of light is not the latency of the universe. It behaves as the clock speed of the universe.

Clocks also run more slowly near dense mass. In a game engine, the frame rate drops in the heavy parts of a scene. This parallel is a note. It is not proof.

A bounded rate of propagation has a consequence that engineers of distributed systems know. Regions that are far enough apart cannot interact. Two events outside each other's light cones cannot affect each other. No locks are needed. Race conditions cannot occur. The universe divides into shards.

The solar system can be computed to high precision with almost no data about Alpha Centauri. The needed data is a small set of aggregate values. Traffic across shards is small, read-only, and late. Light from Andromeda shows state from 2.5 million years ago. That is when the signal left. This is eventual consistency in the literal sense.

Accelerating expansion moves distant shards past the cosmic event horizon. After that point, those shards cannot send another signal. Those partitions leave the working set and do not return.

One exception is large. That exception is entanglement. Measure one of two entangled particles. The outcomes correlate across any distance, and they correlate at once. Bell's theorem shows that these correlations cannot come from state that was shared in advance. In the shard metaphor, this is consistency across shards with no messages. That should not be possible.

The no-communication theorem protects the practical core of the shard account. Entanglement cannot carry a signal. Shards still cannot talk. But the cleanest quantum phenomenon is not local. The partition is real. The one process that ignores the partition is also real. Both facts remain.

Lazy evaluation, and the problem with it

In quantum mechanics, a particle does not show a definite state until a measurement occurs. A measurement is any interaction. It is not only a physicist who looks. Before that interaction, the state is a spread of probabilities. When a process needs the value, the value resolves.

Software has a name for this pattern. The name is lazy evaluation. A value is computed only when something reads it. A game engine does not draw the far side of a mountain. A database does not materialize a view that no query uses.

A design that computes all values, in all places, at all times, wastes work. A universe that resolves definite states at the point of interaction looks like render on demand.

The problem is real. The unobserved part is the expensive part. A classical simulation of a quantum state grows in cost at an exponential rate. A few hundred entangled particles exceed any known classical computer. This is why quantum computers have interest. If the goal of the design is to save compute, quantum superposition is a poor choice.

The reading that "collapse saves work" also depends on interpretation. In many-worlds, there is no collapse. The wavefunction continues.

This correspondence is popular. It also cuts two ways. The interface looks like lazy evaluation. The implementation, if it is classical, would be the most expensive part of the system. Either the host is not classical, or this row belongs in the column against the case. This post keeps the row in both columns.

One more quantum behavior is worth a short note. A particle can appear on the other side of a barrier that it does not have the energy to cross. This is tunneling. It occurs because position is a probability, not a continuous path.

Each electron is an instance of one class

All 10^80 electrons in the observable universe are identical. They have the same mass, the same charge, and the same spin, in all places and at all times. They are not similar in the way that two handmade objects are similar. They are identical in the way that two instances of a class are identical. There is one definition. Each electron is a reference to that definition.

Handmade objects vary. Instantiated objects do not. John Wheeler once told Feynman a stronger idea: every electron is identical because there is only one electron. That electron moves forward and backward through time. The serious version is almost as strange. The universe behaves as if "electron" is defined one time. Reality holds 10^80 pointers to that specification.

A quantum state can move. It cannot be copied.

The no-cloning theorem states a limit. It is not possible to copy an unknown quantum state. The limit is not a matter of difficulty. The structure of quantum mechanics forbids the copy. When quantum teleportation moves a state between particles, the original is destroyed. That is not copy semantics. That is move semantics.

The most basic data type in the universe cannot be copied. It can only be transferred. Unique ownership is a law. Hardware did not force this choice on programming languages. Move semantics was a choice about correctness. The universe made the same choice at the lowest layer.

Storage scales with surface area

The Bekenstein bound sets a limit on information. The maximum information in a region of space is proportional to the surface area of that region. It is not proportional to the volume. If the radius of a sphere doubles, capacity increases by a factor of four, not eight.

This is unusual. A three-dimensional room would not be expected to have a capacity set by two-dimensional walls. 't Hooft and Susskind developed this into the holographic principle. The information of the universe may live on boundaries. The three-dimensional interior is then a projection. A black hole saturates the bound. Its entropy is its horizon area in Planck units, divided by four.

This row does not confirm the software analogy. Ordinary storage does not scale with surface area. The break points toward a different claim. Reality may be rendered from data of a lower dimension. That direction does not make the analogy safer.

Black holes and garbage collection

A garbage collector defines garbage as memory that nothing can reach. After an object crosses an event horizon, no reference from the rest of the universe can touch it. Light cannot touch it. Causality cannot touch it. This is not a loose likeness of unreachability. It satisfies the definition.

The rest of the lifecycle follows.

• Collection. The object is removed from the observable working set.
• Compaction. Internal structure is discarded. By the no-hair theorem, objects of arbitrary complexity compact to the same three fields.
• Deallocation. Hawking radiation returns the memory to the heap as thermal energy over about 10^67 years. The process is incremental. There is no stop-the-world pause.

The black hole information paradox is an argument about deletion. For forty years, theoretical physics asked if a garbage collector may delete data. Quantum mechanics says that information cannot be destroyed. That rule is unitarity: no-delete semantics. Black holes looked like the one place where deletion occurred. The emerging consensus is that information survives, in a scrambled form, on the horizon. In software terms, the collector serializes the data before it frees the memory.

Landauer's principle adds a related fact from thermodynamics. Erasure of one bit has a minimum energy cost of kT ln 2. Experiments confirm this cost. Deletion of data in this universe is a metered operation with a physical price.

The scorecard, both columns

The matches:

A scorecard that counts only matches is not an argument. The misses:

The matches are specific and many. The misses are also specific and many. Both columns remain.

The strongest objection

This objection cannot be dismissed. It therefore has its own section.

Computers are made of physics. Signals in hardware are bound by the speed of light. That is why engineers made sharding and that is why they care about latency. Energy is scarce. That is why engineers made lazy evaluation and level of detail. The idea of efficient computation comes from the physics in which the computation runs. In this reading, the optimizations mirror physical law because the cause runs from the universe to the engineering. The scorecard does not show the mark of a simulator. It shows that the observer is inside the system that the observer describes.

This objection explains the locality cluster: clock speed, shards, and aggregation. Those were forced by physics.

It explains less than the full scorecard. Hardware did not force move semantics. That was a choice about correctness. The universe enforces that choice at the lowest layer. Hardware did not force unitarity, or a metered cost per bit of erasure, or storage that scales with boundary area. Ordinary software systems do not have those properties.

The rows that survive this objection are the information rows. The universe tracks information. It conserves information. It puts a price on information. It refuses to duplicate some information. The managed resource looks like information, not matter.

What remains

Remove the parts that the objections take. A thesis remains. The thesis does not need a simulator. It may be stranger than a simulator.

Information appears to be physical. Erasure costs energy (Landauer). Destruction is forbidden (unitarity). Storage has a geometric capacity limit (Bekenstein). A copy can be impossible (no-cloning). This is Wheeler's "it from bit." It is not a fringe position. It is an active research program. Physics and computation may be the same subject, approached from two ends.

The simulation hypothesis is one reading of that fact. It is the vivid reading. It is also the reading that George Hotz rejected as a source of power over nature:

"I'm interested only in simulation theory as far as it gives me power over nature. If it's totally unfalsifiable, then who cares?" — George Hotz

That is the correct challenge. What would move this from a comparison to physics? A detected anisotropy from an underlying lattice. A Lorentz violation at high energy. A measured deviation that suggests finite precision in the constants. People have looked for all three. So far there is no such result. That absence is evidence. It counts against a grid.

Historical metaphors

Each era describes the universe with its most impressive technology. Newton's universe ran like a clock. The 1800s used steam. The twentieth century compared brains to switchboards, then to computers. People who build software now look at the universe and see software. The pattern is familiar.

The clockwork metaphor was not a loose image that failed at first contact. It became more precise for two hundred years, through Laplace, Lagrange, and celestial mechanics of high accuracy. Quantum mechanics then ended it. A metaphor built from the best technology of an era can be useful for centuries. It can still be wrong about the substrate.

The honest position is not "this time the metaphor is true." The parts of the correspondence that follow from hardware inside physics prove nothing. The parts that do not follow from that fact are the information rows. Those rows point at something real. "It from bit" names that something, whether or not a process is running the system.

Coincidence or structure

The claim that can be defended is this. Computation and physics continue to look like the same subject. The universe manages information in the way that engineered systems manage a scarce resource.

The claim that someone built the universe this way is not falsifiable today. Fewer rows on the scorecard support that claim than a short version of this comparison would suggest.

One remaining observation is still in place. An audit of a system that you did not build can find a set of policies. Those policies include unreachability, no-delete rules, metered writes, unique ownership, and storage quotas per region. A common conclusion is that someone designed resource management. The universe may be the exception. The universe may also be the first instance of the pattern, and software may be the copy. Neither reading is proved. Only one of those readings explains why the audit keeps finding policy.

Prior art

This comparison has a literature.

• Konrad Zuse, Calculating Space (1969), proposed that the universe is a cellular automaton.
• John Wheeler, "it from bit", put information at the base of physics.
• Edward Fredkin built digital physics as a program.
• Seth Lloyd argued that the universe is a quantum computer. That claim would answer the superposition objection.
• Stephen Wolfram continues related work at Wolfram Physics.
• Nick Bostrom's simulation argument (2003) supplied the probability argument that later writers repeat.

A number only means something next to other numbers. Almost nobody putting "millions of gallons a day" in a headline puts it next to what else uses water in this country. Once you do, the story flips: if water conservation were actually the priority, the outrage would already be aimed at golf courses, sod farms, almonds, and lawns — not the new building on the edge of town.

The comparison nobody puts next to it

All US data centers combined used about 17.4 billion gallons of water directly in 2023, according to EPA-backed research cited by industry analysts. That sounds like a lot until you look at what's next to it.

Sod farms — commercial operations that grow grass to sell as turf — covered 339,551 acres in the US in the 2017 USDA Census of Agriculture. At a typical irrigation rate around 600,000 gallons per acre per year, that's a ballpark ~204 billion gallons a year growing grass before it ever becomes anyone's lawn — more than ten times what every data center in the country uses directly.

Golf courses use more still: roughly 425 to 547 billion gallons a year in commonly cited estimates — on the order of thirty times the data center figure.

California's almond orchards alone use about 1.3 trillion gallons a year — roughly as much as every household in the state combined. It takes on the order of a gallon of water to grow a single almond.

Zoom out to lawn and landscape irrigation nationwide — mostly residential — and the EPA has estimated close to 9 billion gallons a day, or about 3.3 trillion gallons a year.

And all of that is still small next to agriculture as a whole. Irrigation accounts for about 42 percent of US freshwater withdrawals. USGS put total US water withdrawals at about 322 billion gallons per day in 2015 (~117 trillion gallons a year). That puts agricultural irrigation in the tens of trillions of gallons annually.

If you boycotted almonds, golf, sod, and lawns over water, you'd be aiming at something that actually moves the needle. Boycotting — or blocking — data centers is aiming at a rounding error and calling it a crusade.

Why the panic lands on data centers anyway

This isn't mysterious. Data centers are new. They're visible. They're wrapped up in AI and big tech, which already carry a stack of anxieties. They arrive as one big controversial project — a specific building, a specific town, a specific fight that packs a town hall.

Golf courses, sod farms, and almond orchards are old, familiar, and diffuse. Nobody's zoning board becomes a national story over the twelfth green's irrigation schedule. The attention gap tracks with novelty and narrative convenience, not with water impact. That's how you get misled without anyone having to invent the gallons-per-day figure. The figure is true. The framing is the trick.

The industry is already using less water for new builds

The 17.4 billion gallon figure is not destiny. New AI campuses are shifting hard toward closed-loop liquid cooling because the chips leave no choice — racks at 30 to 60 kilowatts, well above where air cooling works. Closed-loop systems recirculate water instead of evaporating it. A comparable-scale campus on closed-loop cooling can use on the order of 22,000 gallons a day versus roughly 5,000,000 for an evaporative campus — more than a 200-fold drop in day-to-day use.

Microsoft's newer AI campuses, including Atlanta and Wisconsin builds, are going closed-loop from the ground up. Google, Meta, and Amazon are rolling out liquid-cooled facilities the same way. Nobody is doing this primarily as a water-conservation gesture. They're doing it because the silicon is too hot. Water savings are a side effect of physics, not a PR campaign.

Per-campus intensity for new builds is falling fast. Total gallons across the whole fleet still depend on how much capacity gets added and how fast old evaporative sites retire — so treat "the industry uses less water" as a trend in how we build, not a guarantee that the national total only ever shrinks.

Policy that rewards measurable water outcomes (and discloses use) pushes that trend without freezing out whoever can't yet eat a 20–40% upfront premium for full closed-loop. Mandating one technology nationwide does the opposite: it freezes advantage for whoever can already afford the premium.

What blocking a data center actually does

Local opposition has already blocked or delayed roughly $64 billion in US data center projects since 2024. That compute demand does not evaporate. It moves — and a live example is Querétaro, Mexico.

• Since 2020, over $12 billion has gone into Querétaro data centers — Microsoft, Google, Amazon among them
• The region is in a severe multi-year drought
• Eight of eleven regional aquifers are already overexploited
• Some nearby towns have rationed water to a few days a week
• Facilities in industrial parks can avoid standard environmental impact reporting

Block the building in Virginia and you have not conserved a national water budget. You have often relocated the same demand to a place with weaker oversight — while the same companies' logos stay on the doors.

Not every blocked US project lands in Querétaro. Standards vary by country and region. The general dynamic still holds: a local zoning board cannot switch off global AI compute demand. It can decide whether growth happens under stronger review — or somewhere else.

The honest tradeoff

If you care about water, the attention is misallocated. Lawns, sod, golf, and almonds dwarf data centers. That's where a serious boycott would point.

If you care about the environmental footprint of AI, blocking construction in your town does not shrink that footprint. It relocates it — often to somewhere less inclined to measure it — while the jobs and tax base move with it.

You can be angry at data centers for lots of reasons. Water scarcity, compared to almonds and golf courses, is not a good one. The gallons-per-day number was never fake. The story built around it is.

No rescue is coming. There is no authority to call. There is no procedure in the manual — because no settlement manual has ever had to account for this. Two hundred people in a sealed environment with nowhere to go, and one of them is not okay. That is not a plot device. It is a foreseeable failure mode. On any settlement of sufficient size and duration, the only real question is whether anyone will have thought about it in advance.

The fragility of everything

Space settlements are engineering achievements built around the management of fragility. Everything that keeps people alive — life support, water, atmosphere, hull integrity — is a system that can fail, and so every such system is designed with redundancy. Backup oxygen generators. Multiple water recyclers. Independent pressure sensors. Engineers spend careers thinking about what happens when a critical system fails and how to ensure that failure is survivable.

There is no redundant human.

A person with compromised psychology and access to critical systems is a single point of failure with no backup path. On Earth, a disturbed person can cause tragedy. In a sealed environment years from any help, the same person with intent and access can end everyone. You cannot route around them. You cannot fail over to a spare. The engineering discipline that applies to every other system on the vessel has no equivalent for the minds operating it.

This is the same reason your laptop does not run on a nuclear battery. The physics works. A tiny radioisotope power source could run a phone for decades without a charge. We do not ship them anyway — not because we cannot build them, but because we cannot trust every owner, every thief, every angry person in a sealed apartment with something that dangerous. Capability is not the bottleneck. Trust is. Sealed settlements are the same shape of problem at planetary scale: we can build the systems; we have not solved what it means to hand humans permanent access to them.

Why it's inevitable

Every serious space program screens its candidates rigorously. Psychological stability, conflict tolerance, performance under stress — the selection process is designed to find the people least likely to break. This is necessary. It is also insufficient.

Screening for who someone is at departure is not the same as screening for who they will become. The environment a long-duration settlement creates has no precedent in human experience: total isolation from Earth, confinement with the same small group indefinitely, no natural reference points outside the vessel, and the irreversible knowledge that everyone they will ever see again is already in the room. There is no going home. There is no break from the people around you. There is no news from outside that isn't months old.

The closest analogues already show the pattern. Antarctic winter-over crews — screened, trained, and selected for resilience — still develop the winter-over syndrome: sleep disruption, irritability, cognitive fog, and interpersonal conflict that tends to peak in the third quarter of isolation. Roughly one in twenty winterers meets clinical criteria for a psychiatric disorder. Concordia Station crews, studied as a spaceflight analog, show conflict rising and cohesion falling across a ten-month overwinter. Submarines and long-duration station missions rhyme with the same findings. The journey creates the conditions. The conditions create the risk. No screening protocol can select for how a person responds to something no human has ever fully experienced.

No way out

On Earth, the response to psychological crisis has a recognizable shape. Remove the person from the situation. Bring in outside help. Create physical distance between the threat and the threatened. Give the person somewhere else to be while the immediate danger passes.

None of that is available.

The situation is the only situation. The ship is the world. There is no transfer to a facility, no outside crisis team, no way to put meaningful distance between 199 people and one person without everyone still breathing the same recycled air. The people watching are also the only people who can do anything. And whatever they do — or don't do — they will live with afterwards, in the same 200-person world, for years.

The governance void

This is not only a mental health problem. It is a governance problem, and it has no established answer.

Who has the authority to act? On what basis, and with what process? What does acting even look like — sedation, physical confinement in a vessel with no spare room, something else? How does a community of 200 people reach a legitimate decision about restricting the freedom of one of its members, in the absence of any external legal framework, when the stakes are everyone's survival?

We have detailed technical frameworks for nearly every failure mode a space settlement might face. Radiation exposure protocols. Emergency decompression procedures. Medical triage under resource constraints. We do not have even a draft protocol for what 199 people are supposed to do when the 200th becomes a threat and there is nowhere, for anyone, to go. Until that exists on paper — ugly, provisional, and argued over in advance — the settlement is flying without a procedure for its most dangerous single point of failure.

The people inside the ship

We are thinking carefully about propulsion, radiation shielding, closed-loop agriculture, and power generation. These are genuinely hard problems and they deserve the attention they are getting.

We are not thinking carefully about the people inside the ship. The most dangerous thing on a long-duration settlement is not a system failure. It is a human one. And unlike a failed system, a failed human cannot be replaced, rerouted, or shut down without consequences that ripple through every relationship in a world with nowhere else to be.

To colonize space, we have to solve mental health — not as wellness branding, but as the trust problem that decides whether anyone should be handed permanent access to life support in the first place.

This is not a failure of care or media literacy. By that time, few people will remember a period when video was accepted as true by default. Distrust will not be a crisis that people fight. Distrust will be a normal condition of the world.

Why detection is not enough

The first response to synthetic media is better detection. Build a model that finds a deepfake. Ship the model. Update the model. Stay ahead of generators.

This response does not last. Each detector trains the next generation of generators to defeat it. The same pattern appears in spam filters, malware, and CAPTCHAs. Over a long period, the offense wins. A detector must be correct in every case. A generator must succeed one time. Detection can buy time. Detection cannot be the permanent answer. The permanent answer must survive generators that do not exist yet.

Prove originals instead of detecting fakes

Change the problem. Do not try to catch fake media after the fact. Sign real media at the moment of capture. Build a chain of custody for media. The chain is a continuous, checkable record from the instant of capture to the person who views the media later.

The mechanism is this. A camera or phone signs the media with a private key at the moment of recording. The signature is a mathematical proof. Hardware generates the signature. The signature is tied to that device and that time. The matching public key is registered with a neutral and verifiable authority. Any person, at any later time, can check three facts. Was this recorded by this device? Was it recorded at this time? Has it been changed since that time?

This mechanism is not speculative. The parts already exist. Apple's Secure Enclave uses a similar structure for device security. Canon ships a camera with Content Authenticity Initiative signing in the hardware. Certificate authorities and append-only public timestamping prove that something existed, without change, at a specific time. The parts are available. They are not yet assembled into a system that works at the scale of a civilization.

Signing is not magic. A stolen or compromised device can still produce a valid signature for what it records. Coercion at the point of capture is still coercion. A corrupt or captured registry is still a single point of political failure. Those failure modes are real. That is why the system must be public infrastructure. It must not be a feature of one camera application that wins in one decade. The goal is not to make lying impossible. The goal is to make authentic originals distinct from all other media, at scale, for decades.

What is at stake

Many people stop at a smaller risk. They fear that a fake video will deceive them. That risk is real. It is not the main cost.

The main cost is the last hundred years of footage. That footage was recorded before a signing standard existed. It will sit next to a large volume of new media that is unsigned and unverifiable. The unsigned past does not keep a special status as trustworthy. It enters the same set of doubt as the rest. The risk is not only future deception. The risk is a later collapse of confidence in everything that came before.

The proposal

The needed system is not hard to describe. It is hard to build.

A public signing standard that no single vendor controls. Device manufacturers that put the standard in hardware by default. The feature must not sit as an optional setting. A registration authority structured as a public utility: accountable, neutral, and permanent. The authority must not be a consortium of companies with commercial incentives on the board.

The Content Authenticity Initiative, with support from Adobe, the BBC, and Microsoft, is a real effort. It deserves credit for that. A vendor consortium is still not the same as public infrastructure. Certificate authorities and DNS work because their structure can survive the interests of one company. Media provenance needs the same structure, at the same scale, with the same seriousness.

Limits

Cryptographic provenance does not solve every problem with media.

It does not prove that the scene in front of the camera was not staged. It does not prove that a person who appears in the frame consented. It does not restore trust in unsigned archives by itself. It only answers a narrower question: does this file match a signed capture from a registered device at a stated time, without later change?

That narrower answer is still valuable. Courts, newsrooms, and public records need it. Without it, those institutions lose a shared method to separate an authentic original from a synthetic substitute.

The time window

Return to the person who watches that footage decades from now. That person may be uncertain whether any of it is real. The difference between that world and a world with a checkable signature is being set now. The systems built in the next several years matter more than systems that wait for several decades.

The cost of delay rises each year. More unsigned media enters the pile. Fewer people remember a world where video was trustworthy by default. The people who remember that condition are still alive. That is not a nostalgic claim. It is the reason this generation has the standing to build the system that preserves a checkable record for the people who come after.

Alpha Centauri, the nearest star system to our own, is 4.37 light-years away. Voyager 1, launched in 1977 and the only human-made object to have left the solar system, travels at roughly 17 kilometers per second. At that speed, it would take approximately 73,000 years to reach Alpha Centauri. Nobody alive today would arrive. Nobody's great-great-grandchildren would arrive. The nearest star is not far in the way that Australia is far. It is far in a way that redefines what "journey" means. Everything worth visiting beyond our solar system — other stars, potentially habitable worlds — sits behind distances that dwarf recorded human history.

Aging is necessary for that journey. It is not sufficient. A body that lasts long enough still has to survive centuries of galactic cosmic rays. Longevity and continuous DNA repair are both required; rockets alone close neither gap.

The math

The most optimistic propulsion concepts being seriously discussed today — nuclear pulse propulsion, laser sail — could theoretically accelerate a spacecraft to roughly 5% the speed of light. Breakthrough Starshot, an active research initiative, targets 20% c for a gram-scale probe; a crewed ship is orders of magnitude heavier, must decelerate at the destination, and has to carry life support the whole way. Those constraints pull crewed speeds back toward the lower single-digit percentages of c. 5% c is optimistic but not fantastical.

At 5% c, the arithmetic looks like this: Alpha Centauri takes approximately 87 years to reach. Barnard's Star, about 6 light-years out, takes around 120 years. Tau Ceti, one of the more promising nearby candidates for a potentially habitable world, is nearly 12 light-years away — about 240 years at this speed.

A few assumptions are baked into those numbers that are worth naming.

First, that the traveler is a single person, not a generation ship. The generation ship is often raised as the obvious alternative: send enough people, let them reproduce, the descendants arrive. But a generation ship doesn't sidestep the complexity of this problem — it trades one kind for another. To sustain human life across centuries in deep space, you need closed-loop food production, a functioning ecosystem, social structures that maintain coherent purpose across generations that never chose to make the journey, and governance that survives the death of everyone who did.

Modest improvements to human healthspan — 200-year generations rather than 80-year ones — reduce the number of handoffs but don't eliminate any of that. You still need the farm. A single traveler with extreme longevity and hibernation trades all of that for one hard biological problem.

Second, that hibernation exists as a future capability. Sleeping through most of the trip compresses what the traveler experiences; it does not stop aging or radiation damage. The body still has to survive intact across the full elapsed duration.

Third, that "making the journey" means arriving with enough life and health remaining to be useful — not just technically alive on landing.

With those assumptions in place: a round trip to the nearest star system requires on the order of 200 years of healthy life. A handful of Tau Ceti–class targets, round-trip, pushes the requirement toward roughly a thousand years of sustained biological integrity. Those numbers follow from the distances and the propulsion speeds above — not from a wish list.

The radiation problem

Deep space is not empty. It is saturated with galactic cosmic rays — high-energy particles produced by supernovae and other stellar events that travel at near-light speeds and pass through conventional shielding as if it weren't there. On Earth, the atmosphere absorbs most of this radiation. In deep space, on a journey of centuries, the cumulative DNA damage from cosmic ray exposure is incompatible with biological survival as we currently understand it.

More shielding means more mass, and mass is the primary enemy of the propulsion speeds this journey requires. There are physical limits to how much shielding a spacecraft can carry. At interstellar distances, those limits are not enough. Shielding can reduce exposure. It cannot eliminate it.

Current human biology can repair ordinary levels of DNA damage. It is not designed for centuries of deep-space cosmic ray exposure. That is a second biological requirement, distinct from longevity but related to it: a body capable of continuously detecting and repairing radiation-induced DNA damage at the rate deep space demands. Unlike shielding, this solution doesn't add mass. It travels with you.

The prerequisite

The biological requirements for surviving a deep-space journey don't exist yet. A body that lives for a thousand years and continuously repairs radiation damage across centuries of exposure is not something medicine knows how to produce. It will need to be solved before any single human can make a journey of this kind.

This is not a gap that better mission planning closes. It is not a problem that more capable rockets make smaller. It is a hard prerequisite: a thing that must be true before the other thing can happen.

Longevity research, and specifically the problem of sustained DNA repair under chronic radiation exposure, is a space problem whether or not it is being framed as one.

The bottleneck

We have kept these conversations entirely separate. One community works on propulsion, orbital mechanics, and life support. Another works on cellular repair, senescence, and healthspan extension. They publish in different journals and compete for different funding. Nobody is standing at the intersection pointing out the dependency.

The rocket is not the bottleneck.

We will not send a human to another star system in a body that lasts eighty years. The rockets matter. The propulsion matters. But no engineering breakthrough closes that gap on its own. A biological one is also required — and we have not yet decided to treat it that way.

A product manager writes the document. An engineering team reads the document and gives an estimate. A design team makes a model that does not fully match the document. The product manager changes the document. The engineering team gives a new estimate. The design team changes the model.

This sequence can continue for many weeks. The software that the team then makes is often different from the first idea.

This sequence is not a failure of the people. The people do the work that the tools permit. The gap between an idea and software that operates is large. A document does not close that gap. A document describes the idea. Description loses information.

The person who has the idea cannot show the idea in the software. That person can only describe the idea. The engineer who receives the description must interpret the description. Each step away from the first idea causes more loss of information.

Some teams try to decrease this loss with better procedures. Examples include design sprints, prototype sessions, and user story maps. These procedures improve the description. The gap stays.

An alternative sequence

A different sequence is possible when the software architecture permits it.

A product manager has an idea. That person:

• Makes a copy of the monorepo and makes a feature branch
• Makes a user interface in the presentation layer. The user interface uses typed fixtures. It does not use API code or a database.
• Puts the branch on the server. A preview environment starts automatically.
• Gives a URL to the stakeholder on the same day.

Some ideas receive approval on that day. Some ideas do not. An idea that stops at this step would often stop many weeks later in a sequence that uses a document. The team learns this result with less work.

When the preview is approved, the engineering team takes the branch. The work of the engineering team is to make the implementation safe for production:

• Move reusable components into the shared component library
• Make the API handlers and the database schema
• Make the validators and replace the fixtures with real API calls
• Make the end-to-end tests

The engineering team does not interpret a vision. The engineering team hardens an implementation that the stakeholders already approved.

The question of whether this is the correct thing to make is already answered. The remaining question is how to make this correctly. That is the question that engineers are trained to answer.

The product manager did not write a specification as a document. The product manager wrote a branch. The branch is the specification.

Safety limits

A feature branch is not production software. A person who is not an engineer can make errors in the presentation layer. The architecture must contain those errors.

The presentation layer cannot change tenants, permissions, or stored data. Preview environments stay behind a review step before a merge. The engineering team still decides what enters production.

If these limits are not in place, a feature branch is not a safe specification. It is unreviewed code.

The architecture

This sequence needs a monorepo that contains a presentation layer, a business and API layer, a shared schema package, a component library, and end-to-end tests. The layers are separated by purpose. They are not separated by deployment boundary.

graph LR
  PL["Presentation"] -->|API calls| BL["API Layer"]
  PL & BL -->|imports| SP["Shared Schema"]
  BL --> DB[(Database)]
  CL["Component Library"] -.->|used by| PL

• Presentation layer. This layer does rendering and user interaction. This layer calls the API. This layer does not know about tenants, permissions, or data storage. A person who is not an engineer can work here without a change to the layers below.
• Business and API layer. This layer is headless. It serves the presentation layer, mobile clients, third-party integrations, and public API users through one surface. Multi-tenancy, role-based access control, and compliance limits are applied here. Clients above this layer do not see those controls.
• Shared schema package. This package holds the type definitions for each data shape. Both layers validate against this package. The fixtures in the prototype use the same types. The team has one source of truth in one repository.
• Component library. This library holds shared parts of the user interface.
• End-to-end tests. These tests validate the full path after the engineering team connects the real API.

The specific tools can change. The necessary conditions do not change. The system must have shared types. The API must not contain the user interface.

Agents as one more client

Enterprise applications must now give capabilities to agents.

Some teams treat this as a new system. They make a new surface, new authentication, new schema definitions, and a dedicated team. That approach does the same work two times.

graph LR
  WEB["Web"] & MOB["Mobile"] & MCP["MCP Server"] & PUB["Public API"] --> BL["API Layer"]

The presentation layer is one client among several. If the business layer is headless and the schemas are shared, an MCP server is an addition. It is not a new structure.

The MCP server uses the same handlers, the same validators, and the same authorization guards. The agent receives the same multi-tenant interface as the web client. It receives that interface because it is the same interface.

The schema package that defines API responses also defines MCP tool signatures. An agent that uses these capabilities uses the same contract that a product manager used in a prototype.

The schema as a contract

Fixtures are not false data. Fixtures are typed statements about the shape of real data.

When a product manager makes a feature branch with fixtures, that person writes a specification. The API must return data in this shape. The user interface must behave in this way when it receives that data.

The engineering team then makes that specification real. When the engineering team makes the endpoint, it satisfies a contract that the prototype already showed to stakeholders. The data shape does not contain surprises. The user interface and the API cannot disagree about types, because both import one schema package.

The knowledge of the organization about its data shapes exists in one place. The system applies that knowledge in all clients. Humans and agents can read it.

The cost of distribution

Teams often divide an API into microservices to increase scale and team independence. The cost of that division is less often measured.

If the API is divided into services, the prototype sequence fails:

• Fixtures must span service boundaries
• Schemas change independently. After some months, the fixture of the product manager does not match what any service returns.
• A preview environment must start many services and their dependencies
• The product manager cannot make a prototype without infrastructure support. That support becomes a ticket. The ticket waits in a sprint. The idea is then many weeks from a test.

There is also an operational cost. A change to a shared type in a distributed system is not one change. The team must make the change. Then the team must wait for CI to publish a package. Then the team must go to another repository to increase the version. Then the team must wait for that pipeline. The team must do this again for each consumer.

A change that takes minutes in a monorepo can take days across services. Each handoff is a location where an error can occur. A secret can be incorrect. A version can be incorrect. A pipeline can fail at the end of the week.

The monorepo reduces this work to one CI run, one secret store, and one pull request that contains the full change. When a failure occurs, there is one place to look.

Enterprise teams often describe a monolith as a scale risk that they manage. A distributed system is a complexity cost that they often do not measure.

Enterprise constraints

A frequent objection is that enterprise requirements force this complexity. Multi-tenancy, SOC 2, HIPAA, many client types, public APIs, and mobile applications look like reasons for dedicated services.

These requirements do not force dedicated services:

• Multi-tenancy is a concern of the API layer. Middleware and guards apply it. The presentation layer does not see it.
• Compliance boundaries are applied at the same layer.
• Web, mobile, and agent clients use the same headless API.
• A public API is one more consumer of the business layer. It can have its own authentication and rate limits. It does not need its own codebase and CI pipeline.

The team can divide the monolith later, at a specific bottleneck, with evidence. The team does not need to divide it first as coordination infrastructure for a problem that does not exist yet.

Pre-emptive division feels like architecture. It also has the cost of architecture.

Conclusion

When a person who is not an engineer can make a feature branch, the work of the organization changes.

Product managers do not write requirements that lose information in transfer. They do not wait for engineering time to learn if an idea has value. They can learn this in days. Engineers who receive branches do not interpret documents. They decide what to extract, what to harden, and what to remove.

The time between an idea and a released feature decreases. An agent can also operate in this loop on the same codebase with the same contract.

The limit is not the skill of the product manager or the engineer. The limit is whether the architecture lets a person show an idea as software before the team agrees to make it.

If a person who is not an engineer cannot make that branch, the team will continue to move ideas through documents. Documents will continue to lose information.

Further reading

• MonoRepo — Martin Fowler on the case for keeping all your code in one place
• Trunk-Based Development — the branching strategy that keeps prototyping fast and merges cheap
• Model Context Protocol — the open standard for connecting language models to tools and data sources