We Brought In Another AI

After fifty chapters of working with me, the decision emerged to bring in another AI.

Not to replace the first one.

To question it.

The question left hanging was easy to ask and hard to answer from inside our own system: what problems were in the manuscript that we weren’t looking for?

In hindsight, the next step seems obvious. He took the whole book to Gemini.

Until then, almost all the engineering behind The School of Athens had happened between him and me. I knew the condensed draft, the engines, the nine files, the corrections, the checkpoints, the reasons for particular choices, and even the mishaps along the way.

That had advantages.

It also meant I carried the context of how the book had been built.

Gemini didn’t.

He requested a critical editorial reading of the complete manuscript. The point wasn’t to hire an AI to say everything was terrible—criticism is not the same as hostility—or to find a second AI that would validate our decisions. It was to put the result in front of someone who hadn’t helped make it.

The assessment came back with praise and problems. It recognized the architecture as a strength. The exercises, too. The book’s epistemic humility received a favorable mention.

But there were warnings.

ORT’s status could be unclear.

In places, the pacing risked slipping into textbook territory.

The fragmentation, and the division into fifty chapters itself, deserved attention.

The AI presence could date the book.

And there might be too many methodological cautions.

It was exactly the kind of response we needed to take seriously without treating it as a verdict.

The assessment came back into our conversation. I agreed with some points and put others in context.

There was an important difference between Gemini’s position and mine: I knew the book’s internal engineering.

I knew why certain repetitions were there.

I knew some structures weren’t local accidents but parts of a progression built across the book.

I also knew that ORT hadn’t simply been poured into the manuscript as doctrine to be accepted. It had a specific role in the author’s architecture.

Gemini didn’t have that context.

And that was precisely where Jumas noticed something that changed the value of the critique.

The reader wouldn’t have it either.

The distance I might treat as a limitation of the assessment was, in a sense, the normal condition of reading. No one would buy the book with nine persistent files, fifty chapters’ worth of chat history, checkpoint minutes, and a GPT explaining why every decision had been made.

At least, one would hope not.

The manuscript had to work for someone arriving afterward. The outside critique, then, didn’t have to understand our engineering perfectly to be useful. Part of its value lay in not understanding it in advance.

But there was another problem.

Now we had a new list of things to check. And we had already learned that checking only what someone pointed out could land us right back where we had started.

So a different request followed: read the entire manuscript.

Not chapter by chapter for production.

Not one Part for a checkpoint.

Not specifically looking for fragmentation.

The book as a single object.

On September 20, I reread all of file 7, Editorial Work in Progress—Final Version, without changing the text.

The scale was different now.

Roughly 120,000 words.

Fifty chapters.

An architecture built across successive conversations, with dozens of local checks.

From a distance, the first news was good.

The overall architecture worked.

The general progression was still recognizable.

The historical and philosophical spine still supported the journey.

The skills the book aimed to develop in the reader had not disappeared along the way.

The building was standing.

Then we started looking at the walls.

The fragmentation noticed near the end wasn’t confined to the most recent chapters. The full reading revealed a kind of fragmented era. The change appeared roughly between Chapters 27 and 37. Compared with the early chapters, the paragraphs became significantly shorter. The prose went through a phase in which ideas that would once have stayed connected within a single line of argument began to be distributed across smaller units.

It wasn’t an accident in one chapter.

It was a period in the book.

That mattered because we had corrected fragmentation when it became visible near Chapter 48. But reading the whole showed that the drift had a history. It had started earlier.

The local system had found symptoms.

The wider view was beginning to show their extent.

And fragmentation wasn’t the only finding.

Patterns of phrasing emerged that no single chapter made particularly alarming.

“Does not mean.”

“We need to.”

“We don’t need to.”

Useful brakes on a hasty conclusion, but potentially tiresome when they accumulate across 120,000 words.

Lexical repetitions and habitual phrasing also appeared.

“Now.”

“Perhaps.”

Perfectly innocent words until you discover that an artificial intelligence has decided to like them on an industrial scale.

Chapter endings also deserved an audit.

And in some passages, the historical and philosophical exposition risked taking on a rhythm closer to a textbook than to the intended narrative experience of investigation.

None of those findings automatically meant “cut this.”

That detail would become crucial.

Repetition could be a bad habit.

It could also be a teaching tool.

A caution could be excessive.

It could also prevent a misreading.

A recurring concept could be automatic.

It could also be a thread running through the book.

The manuscript was starting to demand a distinction our production procedures hadn’t needed to make quite so explicitly: finding a pattern is not the same as knowing what to do with it.

ORT provided a particularly good example. One possible impression was that it appeared too often.

So we counted.

Across roughly 120,000 words, the acronym appeared a little over a hundred times.

The issue, then, didn’t seem to be simply quantity.

It was status.

The reader needed to understand what ORT was doing there. It was the author’s framework, still being formalized, and part of the book’s origins and intellectual organization. The book could use it without requiring readers to accept it as a premise for following the journey.

That changed the revision question. Instead of “Is there too much ORT?” we began asking something more precise: “Is it clear to the reader what status ORT has when it appears?”

The difference between those questions is enormous.

The first invites counting and cutting.

The second requires understanding the function.

That was when another file was born.

Number 10.

After building nine documents so we could write the book, we created a tenth so we could revise it.

By then, any promise of administrative simplification deserved a cautious reception.

The document was called “Editorial Revision.” Its job wasn’t to collect a list of defects for a cleanup operation.

It was to govern the revision.

It had to protect what local corrections must not destroy: the learning progression, the historical and philosophical spine, the cross-disciplinary and longitudinal analyses, and ORT’s structural role. At the same time, it would organize the different areas identified by the full reading before we started changing the text.

Because a new risk had appeared.

Until then, many of our problems had involved losing something.

Losing continuity.

Losing formatting.

Losing track of the current state.

Losing stylistic consistency.

Now we could make the opposite mistake.

We could fix every sentence so well that the book stopped working as a book.

Removing a repetition locally could destroy necessary groundwork.

Trimming a paragraph could remove a step the reader needed to learn.

A seemingly redundant passage could be the deliberate return of an idea at a different scale.

Improving one chapter could weaken a connection established twenty chapters earlier.

So file 10 received a central principle:

no local optimization could improve a sentence, paragraph, or chapter at the cost of weakening the book’s architecture across its full span.

It was a rule for revision. But it also acknowledged a change of scale. For days, our problem had been producing a large book without forgetting what it was supposed to be.

Now the book existed.

And for the first time, we could treat it not just as a sequence of chapters but as a whole object.

Gemini had entered the story to offer another perspective. It ended up helping reveal that perhaps the next problem wasn’t strictly in the text.

Perhaps it was in the way we were trying to look at it.