Why Novus Learn keeps the source in view

The problem with black-box answers

When a tool summarises a topic for you, the useful question is not just “is this a good summary?” but “where did this come from, and can I check it?” Study material that can't be traced back to a source is impossible to verify and easy to quietly get wrong.

The failure mode is specific and it is nasty. A confident paragraph that is 90% right is more dangerous than one that is obviously wrong, because nothing in it signals which 10% to doubt. You revise from it, you repeat it in an exam or a meeting, and the error travels with you. By the time anyone catches it, you have already built other ideas on top.

Novus Learn was built around the opposite default: every claim, key point, flashcard, and quiz question is assembled from a real source's own sentences, and every one links back to the exact evidence it came from. There is no step in the pipeline where a model is asked to write something plausible.

What “source-grounded” means in practice

Start from a public Wikipedia article or a file you upload. Novus Learn pins the exact revision, splits the text into stable sections and source blocks, and extracts the direct claims those blocks actually make. Your study set is built from that structure — not from a model's memory of the topic.

That structure is what everything else hangs off. A key point is a claim with the concepts it mentions attached. A summary section is made of real sentences from one part of the source. A diagram node is an entity that some claim named. Nothing is a free-floating assertion, because there is nowhere in the data model for a free-floating assertion to live.

Open any flashcard's answer and you'll see the sentence that supports it, with a link to the exact section of the source. Nothing is paraphrased into something the source never said.

Every card carries its evidence

Flashcards are the clearest example, because a card is exactly the kind of thing that is easy to fake. Novus Learn builds five kinds, and each one is pinned to at least one claim it came from:

• Definition cards — a concept the source names, answered with what the source says about it.
• Cloze cards — a real source sentence with one term blanked out, so the answer is verbatim by construction.
• Date cards — a year or event the source states, never a date inferred from context.
• Relationship cards — only connections the source supports directly, and never the vague “related to” link.
• Source-evidence cards — the quotation itself, for when the wording is the thing worth remembering.

The validator that would rather say nothing

Before a study set is shown to you, it is checked against the project it claims to come from. Every claim ID, entity ID, and evidence reference has to resolve. Anything that doesn't is dropped — not repaired, not approximated, dropped.

Cloze cards are the neatest illustration of why this holds. The card is made by taking a real sentence and blanking out one term in it, keeping the source's own capitalisation for the answer. There is no step at which the sentence could be smoothed, shortened or improved, because the card is the sentence. Whatever you memorise from it, you memorise in the words the source used.

The practical consequence is that Novus Learn sometimes gives you less than you hoped for. A thin source article produces a short summary and few cards. A topic with no clearly-stated relationships produces no relationship cards at all. We think an honest gap is more useful than a confident invention, and the code is written so that filling the gap is not an option that exists.

It is worth being blunt about what this rules out. There is no fallback model that steps in when extraction comes up empty, no cache of general knowledge to borrow from, and no setting that loosens the check in exchange for more material. The refusal is the feature; everything else in the product depends on it being unconditional.

Asking a question without leaving the source

You can also ask a topic a question in your own words. The question is parsed into an intent — define, who, when, where, why, how, compare, list, summarise — and then answered by retrieving and ranking claims that already exist in the project. The answer you read is assembled from real source sentences, with evidence links attached.

There is no language model in that loop. That is a real constraint: phrase a question in a way the retrieval can't match and you will get an honest “the source doesn't answer this” rather than a fluent guess. When the source genuinely doesn't cover something, saying so is the correct answer.

What you give up, and what you get

Being honest about the trade: source-grounding makes Novus Learn worse at some things. It cannot explain a concept in simpler words than the source used. It cannot bridge two sources that don't reference each other. It cannot answer a question the source never addresses, however reasonable that question is. If you want a tool that will always have something to say, this is not it.

What you get in exchange is that checking a claim against its source becomes a one-click habit rather than a research project. That habit is itself a study skill — the one that separates people who know a subject from people who have read about it. Keeping the evidence in view turns “trust me” into “see for yourself”, which is exactly the discipline good research depends on.