Account for it honestly

This page exists because the last time this project described its own costs, it said six dollars a month. That was true of one line on one invoice and false of the platform. An argument about other people's spending that rounds its own down is not an argument, it is a sales pitch — so every line is here, including the ones nobody has a price for.

What is missing from that number

Lines this project knows exist and has no price for. They are excluded rather than estimated, which is why the figure above is a floor and not a total. An unpriced line that gets quietly zeroed becomes a claim; an unpriced line that stays visible stays a question.

Two figures that disagree

Turn on the costs an employer would carry

The floor above is money that has actually left D's account. It is not what this platform would cost an organisation, because an organisation rents space, moves people around, and pays somebody. Those switches are off by default and off in every headline on this page. Turning them on answers a different question, and the difference between the two answers is the whole argument.

What has been spent since the beginning

A run rate is not a total. The figure a pitch actually needs is lifetime spend, and this project cannot yet produce one — not because the rates are unknown, but because the start dates are.

What the work paid, era by era

Every figure on this tab comes from the Bureau of Labor Statistics through its public API, pulled by a script in this repository. None of them is typed by hand. That constraint is what makes the tab worth reading and it is also what puts a hole in the middle of it — see below.

The finding, which is not the one that was expected

—

The eras, and the languages that defined them

Each era is dated from the language table behind Rosetta — a language's period of dominance, not its birth year. The wage column is the best BLS series that reaches that era, and it says which one, because they are not the same series and must not be read as one line.

model The era boundaries are editorial: a language's dominance is a judgement, recorded in languages.json with its reasoning. The wages are not editorial. Where an era predates all wage coverage the cell says so rather than showing a number.

Today's occupational split

The single sharpest number on this page, and it is a snapshot rather than a trend, because the API serves only the current survey year for these series.

What D would have been paid

Very few people, very large sums

The counter-case to any wage chart is that engineers are not paid in wages at the top of the market — they are paid in exits. So here is the exit record for the deals where a handful of people split an enormous number, with the per-head figure computed and the wage of that year set beside it.

The deals that died

An exit is not a plan. These were signed by boards of public companies, announced with press releases, and then did not happen. They matter twice over: the price proves somebody really would pay it, and the outcome proves that proving it is not the same as being paid.

Who captures the value

—

The layoffs, from the filings rather than the tracker

Layoff coverage in 2022–2025 was built largely on crowd-sourced tallies. The federal government also counts layoffs, monthly, in a survey of employers — and its record of the Information sector says something different from the headlines.

The floor, and why it is the honest headline

The defensible number for what this platform is worth is not a valuation. It is a replacement cost: what a conventional team would have to be paid to produce the same working software. That is computed the same way Rosetta computes it — the platform's real measured source, backfired into function points, run through COCOMO II — and it carries the same caveats.

Who buys civic software

This is a market-structure question with a factual answer, so the roster below is built from what each buyer has already bought. That is evidence about appetite. It is not a prediction that any of them wants this, and nobody here has spoken to any of them.

The paths that are not a sale

Each of these is weaker on money and stronger on consistency, and one of them has a conflict serious enough that it is written into the row rather than left out of it.

The scenarios

Everything below is tagged speculative: a scenario, not a forecast. Assumptions listed, probability given as a range, reasoning exposed. The standard being matched is the Ouroboros white paper's ~34% finding, which was published against this project's own interest.

The thesis, stated so it can fail

An individual working with AI assistance can now produce what recently took a funded team, and there is a short window before that capability is priced in.

That is a claim, not a conclusion, and this page is built so the evidence can refuse it. Here is what would have to be true for it to be wrong, and where the current evidence sits on each.

The comparison, with both halves

The dot-com comparison is D's own, and he named the uncomfortable half himself: hype and speculation fuelled that growth, and most of it did not survive. A page that cites 1999 without citing 2001 is telling a story.

The position this is argued from

D lost his job in the 2020 pandemic and has been living inside this question since. Alone and with AI assistance he has built a platform whose conventional replacement cost is in the millions, on a cash floor of a few hundred dollars a month, drawing no wage from it. The capability that made that possible is available to anyone for the price of a subscription, which is precisely why it is an arbitrage and precisely why it closes.

Stated plainly, without melodrama, because the facts do not need any and because the dates behind them are still missing — work-history.json is a form waiting to be filled in, and until it is, this paragraph is the weakest-sourced thing on the page.

Four kinds of number

Rosetta uses three evidence tiers. This page needs a fourth, because it makes claims about the future and those are not the same kind of statement as a wage series.

measured Filed or counted. Every BLS series, the platform's own source line counts, the cost lines with a stated or listed price. Re-derivable by a stranger with curl.

model Produced by an estimating model. Replacement cost, function points, effort and schedule, the inflation adjustment. Explained below with its failure modes.

reported Somebody else's published figure. The acquisition record and the acquirer roster. Attributed, not adopted, and each row says whether anyone here has read the underlying filing.

speculative A scenario, not a forecast. Assumptions listed, probability as a range, reasoning exposed. If a number here turns out wrong that is expected; if the reasoning behind it was hidden, that is the failure.

How the wages are made

Where this is wrong

Annualising an hourly wage assumes a full year. Every annual figure here is an hourly rate times 2,080 hours. Nobody works exactly that, and the people this page is most about — the laid off, the between-contracts, the unpaid — work fewer. The assumption flatters every year identically, so year-on-year comparisons survive it. Absolute levels should be read as "a full-time year at this wage", never as observed income.

Two different populations. BLS publishes earnings for "all employees" from 2006 and for "production and nonsupervisory employees" from much earlier. They are different groups and this page never splices them into one line, which is why the long series is the narrower one.

Value-per-employee divides a hard number by a soft one. Acquisition prices come from filings. Headcounts at the moment of acquisition come from journalists. The ratio inherits the weaker parent and the table says so in every row.

Selection. The exit table is the tail of the distribution, chosen because it is famous. The typical software acquisition is small, quiet and unrecorded. A reader who takes a median from that table has measured which deals were memorable.

The replacement-cost model is somebody else's averages. COCOMO II was calibrated on other people's projects, and the gearing ratios behind the line counts disagree with each other by up to 40%. Good for comparing languages. Not a quote.

Sources, and how to check them