Your reader can see every tree and still miss the forest

We keep a rule when we build reading tools: never trust a demo you wrote yourself. So for the first real stress test we handed the engine a story we did not write, a sprawling, real, hundreds-of-chapters fantasy epic, read start to finish, and asked the same question a good editor asks: what is this story actually about, and is it holding together?

The engine did beautifully at the small scale. It tracked every subplot. It caught who was in the room. It followed reveals from the moment they were planted to the moment they paid off. Tree after tree after tree. All correct.

And it missed the forest.

The tell

When we read the output back, the grand themes of the story, the ones a human reader would name if you asked them what the book was about, weren't showing up in the plot ledger at all. They'd landed on the lore shelf: filed as worldbuilding facts, not tracked as living threads that tension and release across the whole arc.

That's a specific, diagnosable failure, not a vibe. A theme that only ever appears as a static fact ("the world runs on X") reads as set dressing. A theme that appears as a thread (introduced, complicated, threatened, resolved) reads as what the book is about. Same material; the difference is whether the engine watched it move.

The cause was boring, which is the good news

It wasn't a reasoning failure. It was a starvation failure.

To read something this long without cost exploding, the engine summarizes each chapter down to a digest before the whole-story pass ever sees it. That cap was set conservatively low. The result: by the time the big-picture reader looked at the book, each chapter had been compressed past the point where a slow-burning theme was still visible. Local detail survived the squeeze: a reveal in chapter 40 is a sharp, compact fact. A theme that only exists as the pressure between chapters 40 and 240 does not survive being described 1,500 characters at a time.

The forest isn't in any one tree. It's in the spacing between them, and we'd thrown the spacing away to save room.

Why we're telling you this instead of quietly fixing it

Because the fix is uninteresting and the lesson isn't.

The lesson is that "the engine read the whole book" and "the engine understood the book" are different claims, and the gap between them is exactly the stuff that lives at long range: theme, tone drift, the slow inversion of who's right. Those are also the things a human editor is most valuable for and least able to hold in working memory across 300 chapters. Which is to say: it's the part worth getting right.

So the next pass isn't "raise the cap and move on." It's a dedicated whole-book reader whose entire job is the long-range structure the per-chapter digests can't carry, reading the forest on purpose, because we now know the trees will take care of themselves.

This is an internal draft. Numbers and specifics are from one real analysis run and will be tightened before anything ships publicly.