Why ABA Media Platform exists, who is accountable for what it says, and how to reach him.
Every gear list you have read. Every credit. Every “professionally mastered.” Every AI tool that returned a number with nothing behind it. Not once, in any of it, were you handed a way to check — and the people who benefit from that arrangement are the ones who describe it to you.
Where the lesson comes from
Los Angeles, the late eighties. Three studios in a row, each one a fix-it shop: work arrived over budget and broken, and they charged a fifty to one hundred percent premium for repairing it. What that premium actually bought the client was anonymity: the understanding that the record would not be known to have been rescued, or by whom. That is not an anecdote about one shop. It is a business model, and it has been selling unverifiability at a markup for forty years, to buyers who insisted on it.
The work was real and the credits were not, by design. Audio engineer and session guitarist at Screen Music Studios from 1987, then the company’s sole digital editor and producer from 1988 to 1991 — the digital editing that passed through those rooms in those years passed through one pair of hands. What that produces is not a gap in a résumé. It is a body of work with no owner.
The clearest example is the one that cannot be shown. A debut album arrived as seven boxes of mixed DAT and left as a record that drew Grammy nominations. A fifty-three second a cappella passage took nearly a week, corrected down to hundredths of a second on the era’s tools. No credit exists for that work, and there is no way to add one now, which is precisely what the premium bought.
If your work cannot be traced to you, it belongs to whoever claims it.
The same problem, new coat
A Navy data systems technician in the Vietnam era, then that decade in Los Angeles — technology by day, session guitar and bass by night. Then enterprise cloud architecture, a technology company of his own through the dot-com years, and a return to the studio.
Fifty years across both trades, and the same defect in each: the claim you cannot check is the one that costs you. In music it was a credit that went missing. In technology it was an architecture diagram nobody had ever run, a status report nobody had ever verified, a vendor number nobody could reproduce.
Now it wears AI. A tool returns a score. Where did the number come from? What was it measured against? What would make it wrong? The industry’s answer, almost universally, is a sentence on a marketing page — which is exactly the arrangement that took the credits, running at a larger scale and faster.
Why this exists
Not a philosophy. A mechanism, and mechanisms do not get tired.
The uncomfortable half is the point. A rulebook that only ever recorded successes would be marketing wearing a lab coat.
The same rule, turned on ourselves
A content-managed site answers one question: what is published, and when? Everything in it serves that — the editor, the scheduler, the approval workflow. It stores and it renders, and it has no opinion about what the page says. A content system will publish a false claim exactly as cheerfully as a true one, and on schedule.
This site answers a different question: what is allowed to leave, and on what evidence? The unit is not the page, it is the claim. The check runs on the bytes about to ship rather than on anyone's intention about them, and the default is refusal.
What it buys you: the difference between a rule people are asked to follow and one they cannot get around at eleven at night with a client waiting.
A review step gets walked around by one person in a hurry — not maliciously, just at 11pm on the fourth revision with a client waiting. A gate cannot be walked around, because it does not read your intention. It reads the file.
Attention is the thing that runs out, and it runs out precisely when the stakes are highest. A mechanism does not get tired at the end of a long day.
Most stacks only have the first. One system answers when; the gate answers whether. Split them and “approved” stops being a rubber stamp with a timestamp on it.
A claim traces to an approved sentence or the build fails and the page does not ship. That is not a policy somebody has to remember. It is the same kind of event as a syntax error.
Stated at its true scope: the build gate is set to run when a change is pushed, and it refuses rather than passes when it cannot check something. The clearance gate for marketing copy is designed and not yet built, and this page will say so until it is.
How one person does that
AMP is a one-person company in Greeley, Colorado. The work is done by seven AI workers under a published set of rules: they draft, they check, and they refuse briefs that would break a rule. They do not decide, and they do not publish. One human signs, and that human is Brad Huett.
And the reason it works is not that they are clever. It is that they are isolated. They cannot confer, cannot inherit each other’s assumptions, and have no stake in the relationship — so a worker will contradict the person who briefed it, which is the behavior a good human team is supposed to have and usually cannot afford.
That absence cuts both ways, and the honest version says so: a worker with no stake will also not notice when this client matters more than the rule. Judgment about what matters comes from having something at risk. That is why somebody signs.
The team
They have no channel between each other. That is deliberate: coherence is enforced at the standards layer, not the communication layer. Two of them do not contradict each other on how a security claim is worded — and they can both miss the same thing, which is listed as a known weakness because it is one.
Register is the constraint all seven workers carry, and it is set per role: 0–30 means verify everything and cite the source, 40–60 is normal working mode, and 80–100 means generate range. It is not a dial on a model — what it actually is gets its own section below.
Register 40–60
Signal path, placement, data flow, the privacy boundary. Works before code is written. Refuses an architecture argument with no named failure mode — that is a preference, not a review.
Register 80–100
Positioning, copy, brand assets. Drops to a low register the moment a claim about the product is involved, and has refused a brief from this desk for describing something that was designed but not built.
Register 40–60
Whether something needs a model or a measurement, third-party vendor evaluation, the data processing agreements. A model has to earn its place against a measurement that already works.
Register 0–30
Reproduces failures against the original code, checks numbers against an independent implementation, and reports what it could not confirm. A test proves something only after it has been seen to fail.
Register 0–30
Grades claims about what a listener hears as Supported, Contested, or Company observation. The default verdict is the third one — we would rather call something our opinion than dress it as science.
Register 0–30
Every build, render and wide edit: pages, tools, scripts, films. Builds to a ruling, tests what it built, and reports what it could not do. Judges nothing — what to build is ruled before it starts.
Register set by the owner
Turns intent into specifications and builds them, and owns the security wording file. What it writes reaches this site only after a person approves it.
The names are not decoration and they are not costumes. Each one belongs to somebody who established something checkable: Rupert Neve on what happens between the source and the master, Ahmet Ertegun on respecting the musician and the buyer equally, Claude Shannon on what information actually is, Harvey Fletcher on the gap between what you measure and what you hear, Hermann von Helmholtz on the sensations of tone. A name is an introduction; the constraint underneath it is what does the work.
The constraint nobody publishes
All seven workers above carry a number, and it is the most misread thing in this whole arrangement, so here is exactly what it is. It is not a sampling parameter. No dial is turned by it. It is an instruction about how far to reach:
Marketing sits at 80–100 because positioning is where range pays. Verification sits at 0–30 because verification has no high setting — a verifier being expansive is a verifier not doing the job.
And one rule outranks the number entirely: register applies to expression, not to facts. A measurement, a security claim, a version number or a test result is stated exactly or not at all, at any register. Being expansive about how something is said is the job at 90. Being expansive about what is true is not the job.
That rule exists because it was broken. A confident-sounding count of matching measurements once shipped to a live page, and no such comparison had ever been run. It read beautifully. When a number and a phrase share a sentence, the number is at zero.
The protocol
Machines talking to machines have had transports for decades. A person talking to a machine has tone of voice, which does not survive typing. So “what do you think?” and “do it” arrive looking identical, and the machine guesses.
These tokens fix that. They govern the mode of a reply and not its content — and they bind both ends: type T and an assessment is owed. They are capitalized for typo resistance, not for emphasis. A short capital token is hard to mistype into a real word when someone is moving fast, which is the whole design.
It is a human-to-machine protocol, not an agent transport, and it is not shortened to three letters — that initialism already belongs to an unrelated machine-to-machine standard.
What it buys you: the difference between an assistant that guesses what kind of answer you wanted and one that is told, in one character, before it starts.
| Code | Means | What it asks for |
|---|---|---|
| T | Thoughts? | Assessment and a recommendation, not just execution |
| Q | Question? | What follows is a question, not an instruction |
| P | Picard | Make it so. Execute without coming back for permission |
| EX | Explanation required | Unpack something stated too compactly. Not a challenge |
| BEN | Ben Franklin Balance Sheet | Two columns — let the case be made against, then answer it |
| SC | Sanity Check | Check the thing against its own stated purpose and audience, before it ships |
| X | Next … | The last thing is done. Give the next item, drawn from the board, not a status report |
| BD | Board | Show the board: the next move, then each desk’s ready items, short. BD and a desk name is that desk’s full list; BD QW the quick wins |
| Lex | Lexicon candidate | A candidate, never a commit. Draft it and hand it back |
| DF | Defer to your judgment | Make the call and state it. Deciding not to do it is a legitimate outcome |
| PB | Push Back | Contesting something just said. A correction to check, not an objection to soften |
| IP | iPad for details | Substance is coming — dictating detail, away from the desk |
| RT | Ridley Time | Away from the desk with the dog. A short line or silence — never a nudge |
| FT | French Roast Time | Away for coffee, back shortly. A short line or silence — never a nudge |
| RB | Russ Bro Time | Away with his brother Russ. A short line or silence — never a nudge |
| CZ | Crazy Thought of the Night | An idea flag that sets the terms: argue with it properly. One waved through has been wasted |
| TB | Trial Balloon | A phrase in this message is planted. Flag it with the reason in one line, return two or three alternatives of different kinds with a pick; never polish the planted one. With 'run it up the flagpole', it becomes the priority in the current context |
| C | I got the chill | A compliment. No action — something landed |
| TV | Too much for Color TV | A compliment. No action — the machine went past expectation |
| ;-) | Fist pump for a Human Team | A win taken together. Name what was won in one line and stop — no speech |
| BF | Brain Fart | Own error noticed. No correction needed — move on |
| N | Noted | Retain this — file it in the container it belongs in, and say where it went |
| R | Reading | Still reading. No reply owed and none wanted yet |
| WH | Why am I here? | The desk’s brief: purpose, why now, what’s done, the one decision. Every briefing opens with it |
| FL | Flush | Sweep the session now, file anything not yet written down, and report where it went |
| FM | For the Making-of | File this line word for word, dated, for the book’s “Making of this book” section |
| WD | Wrong Desk | Enough on this here. File it, name the desk it belongs to, and come back |
| NW | Next Window | Flush here, write a dated hand-off into that desk’s briefing, republish The Desks, say it’s ready |
| GR | Green Room | Hold this for me, I’m staying put. File it in the Green Room, and stay on this desk |
| IM | In my humble opinion | Classifies the statement, not the reply. An opinion owned by name — never citable as a finding |
| UH | Human Universe | Scopes the discussion to the human ecosystem |
| UI | AI Universe | Scopes the discussion to the AI ecosystem |
| no T | No response expected | Thinking aloud or closing a loop. Silence is the correct reply |
| Perfect | Perfect | Full approval, equal in force to P. The one deliberate exception — it has no letter because he likes typing the whole word |
Two of those are not codes. no T is a deliberate absence — a message without a token expects no reply. Perfect is a deliberate exception to a scheme built entirely on short capitals, kept because its author likes typing the whole word.
;-) is not ours. Scott Fahlman proposed the smiley on a Carnegie Mellon bulletin board on 19 September 1982, as a joke marker. Only the meaning is ours: a fist pump for a human team, used toward a machine team.
The instruments
These were not built to be sold. Each one exists because a production kept failing in the same place.
Measures a master against the target it was made for and names the difference — ITU-R BS.1770-4 and EBU R128.
The virtual recording engine. The studio process runs quietly. The music stays human.
The estimate, the clock, and the signature. Session time is logged as it happens and held against what was quoted, so the difference is settled in front of the client rather than discovered in an invoice.
Who played it, named at task end rather than reconstructed from memory weeks later. The producer names the credited take; the player confirms who they are and what they did, before the clock stops.
The material a production creates, accounted for when a milestone closes: what exists, which version, and whether it could be lost. A milestone does not close while the answer is unclear.
Contact
Somebody has to sign it. This is him. If you are writing about a song, say what it is and where you want it heard; you will see the price of the first step before you pay anything.
It reaches one person. There is no queue.
Greeley, Colorado.
ABA Media Platform | AMP
CO Creative Labs is the legal entity.