What it's actually good for

Five stages of honest measurement had left me with a clear verdict on the thing I set out to build. As a predictor of how a story moves, the structure embedding is weak — the signal is faint, sentence-local, and meaning turned out to be a predictability mirage.

But the same measurements were quietly pointing at what the embedding is good at. Every time it won, it won by being characteristic — by describing a text so distinctively that you could tell what kind of text it was, or who wrote it. That's a discriminative tool, not a generative one. So I pointed it at two discriminative problems.

It fingerprints an author

Style is form: sentence rhythm, how you handle your cast, how heavily you modify, how deep your clauses run. All of which this embedding captures, with zero vocabulary. So I asked it to attribute authorship across fifteen public-domain authors, testing on held-out works so it can't just memorize a book.

Content-free form alone identified the author with 98.0% accuracy — against a 6.7% chance baseline. Adding meaning barely moved it (98.6%). A near-perfect authorial fingerprint, built from grammar and cast dynamics, that never once looked at what the author was writing about.

Why does the content-free part matter more than the slightly higher content score? Because the moment your author writes about something new — or someone imitates their subject matter but not their syntax — a content model is lost and a form model isn't. Style is the robust signal. This is the forensic-linguistics, plagiarism, ghostwriting use case, and form is exactly the part you want doing the work.

It catches machines it has never seen

The other problem is machine-text detection, and here the structure embedding targets a known blind spot. The usual detectors read content and word-likelihood; they get fooled by LLM prose whose form is the tell (the relentless rule-of-three, the antithesis, the tidy staccato). Form is content-free — which, this repo has found before, is exactly the property that lets a signal transfer to generators it never trained on.

So I ran the test that matters: train on some LLMs, then detect text from a different, unseen LLM. The numbers (AUC, held-out generator):

• content (meaning): 0.815 — the worst transferrer. It learns one generator's vocabulary and topics, which don't carry to the next.
• content-free form: 0.861 — it keys on syntax and cast dynamics, which do carry. It generalizes to unseen models better than meaning.
• content + form: 0.882 — best of both.

The hypothesis held: the content-free channel is the transfer-robust one. It's not usable alone — its false-positive rate on human text is too high (around a fifth), the same caution this repo has raised about every pure-form detector — so it belongs as a contributing feature group alongside the word-likelihood features that keep false positives near 5%, adding the cross-generator robustness the likelihood half lacks.

Where it landed

I started out trying to predict the shape of a story and found that the shape, as a forecast, is faint. But the honest negatives kept sharpening what the tool actually is: not a crystal ball for narrative, but a remarkably clean description of how a text is built. Point it at "who wrote this" or "is this a machine" and it earns its keep. Point it at "what happens next" and it will politely remind you that most of a story's next sentence is just the story staying on topic.

That's a smaller claim than "stories have a predictable shape." It's also true, which — around here — is the only kind of claim worth keeping.