qproj wikiReading notes
While reading in-world texts, the player highlights passages and writes notes in their own words. Non-player characters are meant to respond when the player has engaged with particular ideas. To do that, the game needs concept ids, short stable names such as slave_morality, derived from those highlights and notes.1
Contents
Three tiers of interpretation
The design compares interpreters in order of increasing cost:
| Tier | Approach | Output |
|---|---|---|
| 0 | Authored annotations: which tagged ranges does the highlight cover? | score per concept, no model |
| 1 | Embedding similarity between the note and each concept, using a small sentence-embedding model | similarity per concept |
| 2 | Schema-driven extraction models of the GLiNER family | concepts, stance and slots |
Tier 0 uses only the text’s authored tags. It has a clear limit: a broad annotation gives the same answer for any highlight inside it, and an untagged passage gives nothing. The note’s own wording has to make up the difference, which is what tiers 1 and 2 are for. Fine-tuning is deferred until there is a corpus large enough to justify it.
Unlike argument typing, note interpretation cannot be done in advance, because notes are written during play. A model would therefore run on the player’s machine. In the first draft of the debate plan, this was the only model used at runtime. The current draft adds free-text argument parsing as a second candidate.2 The lab therefore plans to measure latency as well as accuracy, under CPU-throttling profiles that imitate slower machines.
Annotation data
A data set is a directory of plain texts, exactly as the player reads them. Each text can have an annotation file, along with a list of concepts and a file of example notes labelled with their correct concepts. Annotations follow a subset of the W3C Web Annotation model. This lets spans overlap freely and gives the Rust side a stable specification. Offsets are counted in Unicode code points, so a Rust port must convert them to and from byte offsets. Optional quotations of the annotated text are checked when a data set is loaded, so annotations that drift after an edit fail loudly.
Annotations carry an extra layer field for the kind of structure being marked, such as argument structure or cross-references. Relations between annotations are part of the model, but are not read yet. The player’s own highlights and notes would fit the same format, so authored tags and the player’s notebook could eventually share one representation.
Current state
note-lab has the data format, a check command that validates a data set and reports every problem, and tier 0 coverage. Its only committed data is a small invented fixture: the real texts are book excerpts and stay local. No tier 1 or tier 2 interpreter has been measured, and the statistical method is to be designed once real data exists.
Annotations come from a labelling tool through a converter, not by hand. The lab’s labelling guide is written for INCEpTION, after Label Studio was also considered. Example notes and their correct labels are still written by hand.
See also
- Debate — where reading and notes fit in the player’s loop.
- Argument maps — the statement-level structure of the same texts.
- Research labs — note-lab’s working method.
External links
- W3C Web Annotation Data Model — the annotation format the data follows.
What links here: Argument maps, Debate, Embeddings and cosine similarity, GLiNER and GLiNER2, Glossary, Research labs