What Scholaris understands about each work, and how it proves it

A layer on top of the text

Reading a work is not understanding it. Besides the faithful text, Scholaris can write down what someone who had studied the whole work would note: its parts, what happens or is argued in each, who appears and where, what the whole is about, and which other works in your library it relates to. This layer never touches the text: it is stored separately, with the model and version that produced it, and can be redone.

Every claim carries its evidence. A card saying a scene is shot in a medium shot points to the exact second and, when visible, to the region of the frame; one quoting a book points to the page and the characters of the quote and, on a scan, to the box on the page where it is. Anything without evidence is marked as inferred.

Specialist lenses

Each work is analysed through its discipline, not with a generic summary. There are sixteen lenses: narrative (order, duration and frequency, focalisation, discourse), poetry, drama, film (shot scale, angle, camera movement, composition, editing, diegetic or non-diegetic sound), documentary, journalism, interview, podcast, lecture, essay (question, thesis, arguments, concepts, works it argues with), scientific article, archive, legal, data, slides and a general one. The lens is chosen automatically from the kind of work and its content, asking a fast model when unsure; it can also be set by hand.

Structure, cards and entities

The understanding of a work has four pieces:

• The structure: parts, chapters, scenes, sequences or shots, as a tree. In audio and video it feeds the player's chapters.
• A card for each part and a global card for the whole work: synopsis, thesis or meaning, themes and motifs with their occurrences, narrative order against chronology, voice and context. Anything from outside the work is marked as external context.
• People, characters, places, works and motifs, linked to your library's graph: the same character is the same entity in a story and in the film that adapts it. In film, each character carries the actor who plays them.
• Appearances of each, with their anchor: named on a page, on screen in a shot, speaking in a span.

Cards are searched like the rest of the library, by words and by meaning.

Films, lectures and audio, actually watched and heard

A light version of each video is made, and a model watches it whole in windows of about eight minutes: scenes, shots with scale, angle, movement and composition, who is on screen and in which region of the frame, significant objects, on-screen text, music (diegetic or not), songs and silences.

The transcript is made in pieces of two and a half minutes and compared with a map of where there is voice: spans that sound but have no text are listened to again and filled, and who says each line is corrected by looking at who speaks on screen. Wordless spans that are music do not count as lost speech. On a two-hour film, speaker accuracy went from 52 % to 96 %.

Works that adapt, translate or answer each other

Scholaris looks for related works in your library: a film adapting a story, a translation, another edition, an essay discussing a book. It infers this from metadata, Wikidata, what the work itself states (a film's credits, for example), the citation graph and shared characters. You can confirm, reject or add relations by hand.

Each pair is aligned span by span: this scene against this passage, this translated paragraph against the original, with the kind of correspondence (equivalent, condensed, expanded, transformed, reordered, added or suppressed) and a note on what changes and to what effect. A card also analyses the whole transformation: structure, characters added or merged, point of view, time and ending. They are shown side by side, with both works open.

Across languages, a question in English, French or Italian finds the passage in Spanish, and quotes are given in the original with their translation alongside.

Research mode

For hard questions there is an agent that first looks at what you have already researched, then plans (which works are involved, what evidence is needed, which sub-questions) and then investigates in several steps with tools: searching the text and the understanding layer, reading whole units, opening a work's map, following an entity through the whole library, looking at alignments and watching or listening to the original span of a video or a page.

The answer is as long as the question needs, in an academic register, with tables when comparing. Each claim sits in its own sentence with its evidence, and a verifier checks sentence by sentence that the evidence says so: anything named or quoted must be literally in what is cited, and visual claims are looked at again. What cannot be proved is said. It never concludes that something is missing just because a search did not find it. Each investigation stores the question in your history and its evidence in a notebook, as verified cards.

On a bench of 36 hard questions about a film and the story it adapts (in four languages, with traps), the usual quick answer scored 47 out of 100 and covered 43 % of the required points; Research mode scored 84 and covered 91 %.

What it costs and who decides

Understanding uses paid models, so it never starts on its own: you ask for it with «Comprender» on each document, which first tells you roughly what it will cost. Almost every step reads the whole work, so the cost grows with its length. As a guide, a three-hundred-page essay costs between half a dollar and eighty cents in economy mode, which uses the fast model. A dense five-hundred-page book costs about three and a half dollars, and with the best model, close to three times that.

A strict cap can be set per document. Before each step, what it will cost is added up and enough is kept for the global card; whatever does not fit is skipped and reported. If the cap does not cover even the structure and the global card, understanding does not start and nothing is spent. A running understanding can be stopped: what is done is kept and, when you ask again, it resumes where it left off without paying for it twice. An investigation costs a few cents.

In the open file

When exporting to SPDF 5.0, the understanding layer travels as an optional extension (scholaris_comprension): structure, cards, entities with their Wikidata identifier, appearances, relations and alignments, with the standard's anchors. A reader that does not know it ignores it. The notebooks of an investigation are exported as W3C annotations. The proposal to make this layer part of the standard in version 5.1 is under public discussion.