Research method
The properties of a character's rationale are not all the same kind of thing; treating them all as “predict labels from corpora” is a category mistake. Our research therefore works in three layers along two tracks, and replaces an impossible “accuracy” with falsifiable tests. Method
Three layers and what can be identified
| Layer | Properties | Identifiable? | Ground truth | Method | Evaluation |
|---|---|---|---|---|---|
| Form | Composition, nine-square layout, strokes, radicals | Decidable | Yes (IDS / Unihan) | Deterministic extraction | Standard accuracy |
| Meaning (common reading) | Original meaning, classification, families, distributional semantics | Inducible | Partly (Shuowen, glosses) | Supervised learning + distributional semantics + LLM drafts | Accuracy + human review |
| Rationale (design reading) | Design readings, combined stroke meanings, yin-yang values | Generable, not inducible | None | Rotor composition + expert validation + falsifiable tests | Expert acceptance + compositional consistency + leave-one-out |
The rationale row is the epistemological crux: a design reading is a reason written in by design, not a distributional signal, and no corpus is large enough to induce it; but it can be generated compositionally by the rotor, and it can be tested against corpora.
Two tracks: existence and generation
The meaning layer belongs to the existence track (distribution, induction, bottom-up), the rationale layer to the generation track (rotor, composition, top-down). The task is not for one to replace the other but for both to calibrate each other on one database: agreement is strong support; disagreement is a research signal, not an error — either the framework needs refining or it is exactly the split between common and design reading the framework predicts (borrowing, transfer and loans pull everyday meaning away from the designed one).
Expert in the loop
Compositional consistency
Leave-one-out
Ablation (minimal pairs)
Measured The leave-one-out figures are machine-side measurements. About 241 atomic cores cover the CJK basic block, and annotating the first 50 or so covers about 84% of component occurrences, which keeps manual annotation manageable.
The limits of the method
- AI produces no ground truth for design readings, only candidates; their authority comes from expert review, never from model confidence.
- The form layer is highly reliable; classification and families in the meaning layer are contested and shown with probabilities; the rationale layer is generation plus testing and must never be described as “discovering rationales from data”.
- Framework properties (the zhen-shan-ren code, stroke meanings) are input axioms; algorithms and experiments test whether characters compose consistently under them and whether their predictions fit the corpus — not whether the axioms are true.
- Corpora can falsify or support the framework's observable predictions but cannot establish its metaphysical claims. Keeping to that line is what makes this science rather than apologetics.
Work in progress Planned
- Chapter 5 of the book: entries for 600-plus components (currently a framework draft with sample entries); consistent terminology and style across the book; replacing some references with original sources.
- The character-rationale database: form-layer pipeline, fine annotation of atomic cores, compositional generation with leave-one-out tests, ablation of database-driven composition (several rounds reported).
- AI tutor: retrieval-based, strictly grounded in reviewed database entries, always saying whether a point is reviewed, pending or a candidate.
- Trial teaching of the first prototype unit (the 田 / 由 / 甲 / 申 / 神 family) with 10–30 learners, assessed on acquisition, retention, transfer and insight, with an explicit go/no-go gate.
- Long term (about 36 months): final theory and database foundation → courses L0–L4 and the platform → advanced courses, learnability experiments, AI companions, an open database and scholarly dialogue.
