Apparatus-first.
Mechanism-made.

The middle of the creative economy has collapsed. In 2023, two hundred and forty-five agentic AI startups were founded in the United States; year-to-date in 2026, the count is one. The most-cited study of Upwork writers after ChatGPT's launch found earnings down about five percent overall, with the top-earning writers losing the most; a Harvard analysis bundling writing and coding put the decline at twenty-one percent. Ramp's *Payrolls to Prompts* tracked labor-marketplace share of corporate spend falling from 0.66 percent in Q4 2021 to 0.14 percent in Q3 2025; AI-model spend for the same firms rose from zero to nearly three percent. Commodity content writing now trades at three to eight cents per word, where it used to trade at twenty to forty. If a language model can produce something from a one-line prompt, the market has decided that something is worth nothing - and the market is correct.

What replaced the middle is not a smaller market. It is a different market. When consumers suspect content is AI-generated, trust drops nearly fifty percent and roughly half disengage entirely; unedited AI content shows a thirty-five percent conversion decline after three months. Seventy-nine percent of Americans say they prefer a human over an AI agent in service contexts. Slop is a Merriam-Webster word of the year. Brands you would not expect are doing the same: Aerie with its Pamela Anderson "no AI bodies, no AI people" campaign, iHeartMedia with its "Guaranteed Human" tagline across all stations, the Authors Guild rolling out its Human Authored certification, the Human Made Mark launching at the Lascaux caves this April. Multiple organizations are racing to establish a global certification mark - modeled on Fair Trade for craft.

I am not joining that race. A certification only names what the work isn't. The mark is one step from being faked, gamed, or diluted; and the moment a Human-Made badge can be slapped on any output a person clicked into existence, the badge has lost its meaning.

I want to make a different kind of claim - one that AI-generated work structurally cannot make. Mechanism-made is a positive claim with technical bite: every piece names its own mechanism. You can point at the gear that made the sound. Diffusion models cannot do this; they predict pixels statistically. Language models cannot do this; they predict the next token. Generative video cannot exit zero only when every frame is causally bound - generative video cannot exit at all. It has no compiler.

Slop, by definition, has no apparatus.

A piece is apparatus-first when the visible artifact is causally bound to the named mechanism that produced it. The gate question is one I borrowed from a private audit I ran on my own work: can you name what physical element crosses what fiducial mark to fire each pluck? If you cannot name it, the piece is not apparatus-first. If you can, it is.

The principle generalizes. In music, every bar names its chord function; the director declares meaning, form, motif role, and energy arc, and the generator obeys within those constraints. In code, every render exits zero only when every step has a smoke test that passes. In data work, every claim cites the path that produced it and the failure modes that path can hide. In design, every element is anchored to a geometric or semantic structure that precedes the atmosphere - flow fields without that structure are weather, not work.

None of this is the same as "good design." Good design is downstream of apparatus-first; it is not the principle itself. The principle is causality between the depicted thing and the experienced thing. Not aesthetics. Not polish. Causality.

Some work passes this test by accident. Some passes by deliberate construction. Almost no work passes it that was generated by a system whose only obligation is statistical plausibility. The image that "looks AI-generated" is one that floats - the figure stands but no shadow falls, the light has no source you can name, the wall behind has no shadow of the figure either. The video that "looks AI-generated" is one where the second-by-second causality of motion does not bind to the second-by-second causality of light. These are not aesthetic failures. They are apparatus failures. The gear is not there.

I distrust magic. I distrust decoration that floats. I distrust elegant abstractions with no shown derivation. The provenance is the product.

The work lives at projectlavos.com. Each piece - plate, specimen, formation, study, note - runs the technology it depends on inside its own page. The catalog tiles are not screenshots of working software. They are working software. Open one and the demo is running because you opened it; click into it and the demo grows; close the tab and the demo halts because the apparatus that ran it has been deallocated.

What the notes section adds is the mechanism walk. Each entry follows the apparatus that produced one piece: the file, the function, the constraint, the failure modes, a quoted code excerpt where it matters. I will show my gears. The fear that doing so trains a model to copy me is real and worth naming - my answer is that a model can copy the surface, not the causality. The semicolons in this sentence are forgeable. The mechanism that produces them, and the reason it produces them in this specific Louisville hand, are not.

The standard governs publication; the calendar does not. A piece appears when it passes the gate question. The body of work that already exists is enough for many months of mechanism walks at any cadence I choose; manufacturing schedule pressure on top of the standard would be the kind of decoration the standard refuses.

The same goes for citations. Every number I cite in this manifesto is hyperlinked above. If a claim does not earn a link, it does not earn the manifesto either. I tested the first draft against this rule and four of seven citations failed it; the failures are corrected here, and the rule is now load-bearing for everything that follows under this name.

There is no co-author. There is no AI credit. The tools I use are tools, the way Logic Pro is a tool and a chord chart is a tool. The author is named. The author is me.

I do not know whether any of this works commercially. I know no one else is making this exact kind of claim. I know what the data says about what readers want, what foundations are funding, what brands are paying premiums to support. I know what I have built - a working surface of fifty-three specimens, thirty-two plates, and nine formations, each shown alongside the mechanism that produced it. The rest is the kind of guess one has to make in public to find out if one is right. I am betting that the work is the case.

I do not think the AI economy is something to fight publicly. Being anti-something is defensive, exhausting, and finally beside the point. I would rather define what I am for than what I am against. The middle of the market is gone; the high end is paying premiums for work that is grounded, slow-built, and structurally sound; the openings are visible; the surface already exists.

This is its name. Apparatus-first. Mechanism-made.

If this resonates, you will find the work at projectlavos.com and the mechanism walks under notes. If not, the gears are still shown - I have lost nothing in showing them, and you have lost nothing in declining the read.

Either way, the work is mine. Either way, the gears are there.