The letter arrived on a Tuesday. Not by email — by registered courier, which Beatriz said was the cruelest part. 'They want me to sign for the thing that threatens me,' she said when she called. 'There is a poetry to it that I do not appreciate.'
I went to her apartment after my last shoot. The Vila Madalena gallerist — Fernanda — had just confirmed the exhibition dates, and I should have been thinking about that, about the process documentation as installation art, about the booking spreadsheet that would hang in a gallery alongside the photographs it tracked. Row 147, the day my ranking dropped from 312 to 847. That row would be on a wall in Vila Madalena in six weeks. Instead I was on the metro to Pinheiros with my laptop and a bag of pão de queijo that Jjang would eventually sit on.
Beatriz's apartment is the kind of place that looks temporary even after three years. She moved in after they fired her and never quite committed to staying. The kitchen table is her office, her dining room, and now her legal defense headquarters. When I arrived, the cease-and-desist was open on the laptop screen, flanked by two terminal windows showing the Selo de Processo repository on one side and her old employer's public API documentation on the other.
'Read the specific clause,' she said. She had not said hello.
I read it. I read it twice. The third time I could not help myself — I laughed.
Detection methodology concepts developed during the course of employment. That was the claim. That the Selo de Processo — our Selo de Processo, the thing we built to document creative process as an alternative to algorithmic classification — used 'proprietary detection methodology concepts' from the company that fired her for proving their detection tools were biased.
The Selo does not detect anything. The entire point of the Selo is that detection is the wrong paradigm. We do not ask 'what is this?' We ask 'how was this made, and is the maker willing to show you?' Process documentation is not detection. It is the opposite of detection. Detection looks at the output and guesses. Documentation looks at the process and records.
'They are suing you for building a window,' I said, 'because they built walls.'
Beatriz did not laugh. She picked up her notebook — the one with the graph paper pages that she uses for everything from architecture sketches to grocery lists — and wrote it down. 'That goes in the public response,' she said.
I opened my laptop. Jjang, who had been investigating the pão de queijo bag, abandoned it and climbed onto my keyboard with the precision of a cat who has done this before. I moved him to the corner of the table where the afternoon light used to be — it was 10 PM now, the overhead fluorescent casting the flat shadowless light that Beatriz works in because she says it reminds her of the detection lab. She does not say this with affection.
'I need a technical comparison,' Beatriz said. 'Line by line. What their detection tools actually do versus what the Selo infrastructure actually does. Every function call. Every data flow. Every architectural decision.'
'You want me to write it?'
'You understand both sides. You are a photographer who uses AI tools and discloses everything. You are the use case. And you read code better than I write prose.'
This was true. I learned to read code from my grandmother, who never touched a computer but who could read a darkroom exposure sheet the way other people read recipes — each chemical concentration, each timing interval, each temperature adjustment as a sentence in a paragraph that told you exactly what the printer intended. The AI tools I use for color grading have configuration files. I read them the way she read exposure sheets. The parameters are different. The attention is the same.
I started at 10:30. By midnight I had the document laid out: two columns, left side their system, right side ours. Their system ingests an image, extracts statistical features (pixel distribution patterns, compression artifacts, metadata signatures), runs them through a classifier trained on labeled datasets, and outputs a probability score between 0.0 and 1.0. The score determines placement: above 0.7, AI-generated tier. Between 0.3 and 0.7, contested middle. Below 0.3, presumed human-made.
Our system does none of this. The Selo de Processo does not look at the image at all. It looks at the process log that the artist publishes alongside the image. It verifies that the log is internally consistent — that the timestamps are sequential, that the tools referenced exist, that the parameters described produce results consistent with the final output. It does not classify. It documents. The output is not a score. It is a chain of evidence.
Zero shared code. Zero shared methodology. Zero conceptual overlap. I wrote this in the comparison document, then deleted 'zero conceptual overlap' because it sounded like a legal brief, and wrote instead: 'The two systems address the same problem — the question of how creative work was made — from opposite directions. One looks at the product and guesses backward. The other looks at the process and records forward. A forensic analyst and a diarist are both interested in what happened. They share no methodology.'
Beatriz read it over my shoulder. 'The forensic analyst and the diarist,' she said. 'Yes. That is it.'
She went to make coffee. The instant kind — Beatriz does not own a coffee maker, which in São Paulo is a form of protest. I kept writing. The section on data flows was straightforward: their system sends image data to a classification model; our system receives process logs from the artist. At no point does our system touch image data. At no point does their system touch process logs. The data flows do not intersect.
The section on architectural decisions was where it got interesting. Their detection system is closed — proprietary classifier, proprietary training data, proprietary threshold calibration. The output is a number. You do not get to ask why. The Selo de Processo is open — open source code, public process logs, peer review by the community. The output is a documented chain that anyone can read and verify.
I wrote: 'The architectural philosophies are not merely different but opposed. One system derives authority from opacity — the classifier's judgment is final because the classifier's internals are hidden. The other derives authority from transparency — the process log is trustworthy because it is visible. To claim that the second system uses concepts from the first is to claim that a window uses concepts from a wall. Both exist in buildings. Both relate to what is inside and what is outside. But they solve the problem of the boundary in opposite ways.'
That was the title. I typed it at the top of the document: 'A Window Is Not a Wall: Technical Comparison of Detection-Based Classification and Process-Based Documentation in AI-Assisted Creative Work.'
Beatriz came back with two cups of terrible coffee. She read the title. She read the first section. She read the forensic analyst and diarist paragraph. She read the window and wall paragraph. She put her cup down on the graph-paper notebook, leaving a ring that she did not wipe away.
'We publish this tomorrow,' she said.
'With the letter?'
'On the same page. The letter above, the comparison below. Let anyone who reads the legal claim read the technical evidence in the same scroll.'
This was the thing I had learned from the Selo de Processo — from building it, from using it, from being one of its 400 practitioners. Transparency is not a defense strategy. It is a practice. You make the process visible not because someone is threatening you but because visible process is how you do honest work. The C&D letter was a legal action. The technical comparison was not a legal response. It was a demonstration. Here is what we built. Here is what they built. Read them side by side. Reach your own conclusion.
Jjang woke up at 1:30 AM, stretched, and walked across my keyboard, adding a line of random characters to the document. I deleted them. Beatriz had fallen asleep on the couch with her graph-paper notebook open to a page of architectural sketches for the Selo's next version — the one that would include video process documentation, not just stills.
I finished the document at 2 AM. Fourteen pages. Every claim sourced to public code or public documentation. Every comparison verifiable. I emailed it to Beatriz and closed the laptop.
On the metro home I checked my booking dashboard. One new inquiry — not from the Vila Madalena gallery, not from a client. From a law student in Bom Retiro who wanted to use my process documentation as a case study in a paper about transparency and intellectual property in AI-assisted creative work. I said yes.
Row 148 in the spreadsheet. Not a booking. Something else. The spreadsheet was developing categories I had not planned for.
Jjang was asleep in my bag. I carried him home like that, on the metro at 2 AM, a cat in a bag and a document called 'A Window Is Not a Wall' on a laptop that smelled faintly of pão de queijo. Fernanda's exhibition was in six weeks. The C&D response would be live by morning. The booking dashboard would still show 847 when I woke up.
But row 148 was not a booking. And the spreadsheet, I was beginning to understand, had never been only about bookings. It was a process log of something I had not yet named — the cost of transparency, yes, but also the things transparency attracted. Not clients, always. Sometimes law students. Sometimes gallerists. Sometimes a friend who needed you to prove that a window is not a wall.
The metro lights flickered. Jjang shifted in the bag. I thought about my grandmother's exposure sheets — the ones she kept in a shoebox in Incheon, each one a record of a specific intention executed in specific chemicals at specific temperatures. She never called them process documentation. She called them 'what I did and why.' She would have understood the Selo. She would not have understood why anyone needed to be told that showing your work is different from hiding it.
But here we are. Row 148. The spreadsheet grows.