The case for comprehension.

Reading is not the same as receiving. When a careful reader gets to the end of a long piece, what they carry away is not a copy of the sentences. It is a structure they built while moving through the text. Cognitive scientists have a good account of this. Walter Kintsch’s construction-integration model describes comprehension as three nested representations the reader assembles as they go.

The surface representation is the literal words and syntax: the substance a screen-reader would see. The textbase is the propositional meaning: what the sentences claim, stripped of phrasing. The situation model is the deeper picture: what the text is about in the world, how its argument hangs together, how it connects to everything the reader already knows. Understanding, in any usable sense, lives in the situation model. A reader who has only the textbase can answer questions; a reader who has built a situation model can do something with the material a week later.

This matters for product design. A summarizer can hand back a clean textbase. It cannot hand back a situation model, because the situation model has to be built by the person doing the reading. The work of selecting, organizing, connecting, and integrating new material with prior knowledge is not a bonus feature of comprehension. It is the thing itself.

Anything that bypasses that work produces fluency (the feeling of having understood) without the underlying structure. The reader feels done; a week later they cannot reconstruct what they read. The clean output of a model and the durable understanding of a person are two different artifacts.

There is a useful hierarchy that runs underneath the Kintsch picture. First-order thinking is recall: the surface and the textbase, what was literally said. Second-order thinking is inference: what follows from what was said, where the argument turns, what is being assumed. Third-order thinking is connection and application: how this piece relates to other pieces, to your own prior thinking, to a decision you have to make.

Tools that compete on speed tend to optimize for the first order. They get faster and faster at answering the literal question. But the questions worth keeping a library for are almost never first-order. They are second- and third-order questions: how does this argument actually work; what does it imply for the thing I’m working on; does this contradict what I read in March; what part of my prior model just got destabilized.

These questions cannot be answered from the textbase alone. They require the reader to hold structure, ambiguity, and prior context in their head while the new material moves through it. A product that takes that work over instead of supporting it gets faster output and slower readers. A product that supports that work makes the work easier to enter, and leaves the thinking with the person who needs to keep doing it.

Summarization is genuinely useful. It is also a specific cognitive trade. A good summary compresses by removing contour: the caveat that changed the meaning, the example that grounded the claim, the part of the argument where the author hesitated, the place where the evidence thinned. The result is short, and short is valuable when the goal is to decide whether to engage at all.

But compression has a cost the user often does not feel. When the contour goes, so does the structure that lets a reader evaluate the piece. Two paragraphs of bullet points can give someone the confident sense of having understood a position while quietly hiding the parts of the position that should have made them less confident. This is what cognitive scientists call false fluency: present clarity without future ownership. It is the dominant failure mode of summary-first workflows.

The deeper issue is that summarizers optimize for completion. They give the user a feeling of being done. For some content that is exactly right. You wanted the gist, you got the gist, you move on. For the content you actually saved on purpose, completion is the wrong success metric. The success metric is can you still use this in three months. A summary optimized for completion will reliably fail that test, no matter how good the model behind it.

Omphalis takes the opposite stance on summarization. We use it to help readers navigate, prioritize, and return, never to replace the act of reading itself. A summary that helps you find your way back into the piece is doing comprehension work. A summary that lets you skip the piece is making you fluent in something you do not actually know.

If summaries flatten contour, what replaces them? The answer is not a longer summary. It is structure that preserves the shape of the piece. A two-and-a-half-hour podcast has a shape: where the conversation opens, where the speakers shift into a deeper section, where a single insight does the load-bearing work, where the closing pulls things together. A skilled reader builds that shape in their head while listening. Most people, most of the time, do not.

Omphalis renders the shape directly. We call it Strata: a typed timeline of 8 to 12 moments, each labeled with what kind of moment it is and how dense it is. Not a summary of the content. A map of the content. You can see at a glance where the dense sections live, where the argument turns, where the story breaks for a definition. The map is what a careful reader normally has to build for themselves before they can do anything else.

The structural move is what makes the rest of the system possible. Once you can see the shape, you can decide where to spend attention, what to mark, what to come back to, what to skip without losing the thread. Structure is not decoration on top of content. It is the precondition for reading the content well, and it is the part summarizers cannot give you because the moment they compress it, it stops being a map.

Most products that promise to help you with content put their help off to the side. You read the article, then you switch modes, ask a question, read the answer, and try to return. The flow is broken on purpose: the help lives in a separate context, and you have to leave the material to reach it.

For occasional, one-off questions this is fine. For sustained reading, it is corrosive. Every mode-switch costs working memory. The reader loses the line of the argument, has to re-enter the piece, has to re-find where they were. Worse, answering from the side trains the user to ask discrete questions (first-order questions, the literal ones) because those are the questions you can phrase in a single sentence.

Annotation belongs in the margin instead. When clarification appears exactly where comprehension would otherwise break (a quiet definition, a source, an expansion), the reader stays in the material. Omphalis annotations are restrained on purpose: dotted underlines that reveal only when the reader pauses, with a bar set high enough that one wrong annotation does more damage to trust than ten missing ones. Low recall, high precision. The point is to keep the reader inside the piece, not to send them somewhere else for the answer. When Omphalis does answer a question, the answer is anchored to the passage and points back into it, so it deepens the reading rather than standing in for it.

Help that lives where the reading is assists comprehension. Help that answers from somewhere else can replace it. The difference is not cosmetic. It is the difference between a tool that supports the reader and a tool that takes the reader’s place.

Most reading apps have a highlighter. Most highlights die. A decade of user research on highlighting shows the same pattern across products: people highlight, they intend to come back, they almost never do. The highlight becomes a ceremonial act of intention, a way of telling yourself the passage mattered, without any of the work that would make it actually matter later.

The shift Omphalis makes is small and consequential. A mark is not a highlight. A mark is a user-owned understanding object: a span of text plus a short note in your own words. The note is the point. The act of putting an idea back into your own language is what moves it from the textbase to the situation model. It is the part of comprehension that has to happen inside the reader’s head, and the part that highlighters quietly skip.

The category around us has spent four years building ambient capture: systems that watch what you do and try to infer what mattered. The bet is that, with enough behavioral signal, the model can recover intent. The structural problem is that intent is not a behavior. Intent is a declaration. A single deliberate save is worth more than a week of screen-recording, because the save contains something passive observation cannot reproduce: your own judgment that the thing matters. Marks are that judgment, sharpened and attached to a specific span.

We treat marks as the most information-dense moment in the entire interaction. Everything downstream (the connections, the resurfacing, the way the library compounds) is built on top of those declarations. Clean inputs compound. Noisy inputs cancel out.

A single piece of content, however well structured, however well marked, is a single piece of content. The interesting thing happens between pieces. Two articles that contradict each other. A podcast and a paper that turn out to be describing the same idea from different angles. A voice note you recorded six months ago that suddenly belongs beside what you read this morning.

These cross-piece patterns are not visible from inside a single piece. They emerge once a library has accumulated enough deliberate saves and marks to have shape of its own. The threshold is not small. Somewhere around two hundred items is when the library becomes the kind of place where new content lands beside something earlier and the system can honestly tell you the two might belong together.

Connections also have to be earned. The product can propose edges; the user has to confirm them, because a connection becomes durable only when the reader can articulate why it matters. Possible connection. Not related. Add why. These are not interaction details. They are how a model- proposed edge becomes user-validated meaning, the thing the graph is actually for.

The same logic governs return. We borrow from spaced repetition but reject the flashcard. Flashcards optimize for recall of units. The Omphalis question is different: can you re-enter this meaning? You marked this two weeks ago. Do you still remember why? This new article contradicts something you saved in March. Want to revisit both? Spaced re-entry, not spaced rehearsal. The library is a place you return to, not a deck you grind through.

Most content products are betting that the future of reading is shorter content, faster summaries, and smoother synthesis. That is a reasonable bet. There is real demand for it, and the tools that win on those axes will be excellent at being fast assistants for casual use.

Omphalis is betting on a different shape of user, and a different shape of moment. We are not for the person who needs to skim a hundred articles before lunch. We are for the person who chose this article on purpose, expects to come back to it, and wants what they read, listened to, and noted to still be there in six months, structured, connected, and usable. The category words for that user are dwindling. Reader. Researcher. Pastor. Teacher. Engineer who keeps a long view. Anyone whose work requires understanding to compound rather than refresh.

The bet you are being asked to make is small in the moment and load-bearing over time. It is the bet that the things you chose to keep are worth scaffolding, not summarizing. That your own marks matter more than any model’s extraction. That a library you returned to ten times is worth more than one you accumulated faster. That comprehension, the slow kind, still has a place in a world that is mostly optimizing for everything but.

We think it does. That is what we built.