# UTF-4088 Interpretability Sampling This experiment moves beyond raw 96-bit capacity and samples for a conservative candidate glyph population. ## Candidate filter A sampled glyph is a candidate when: - it is 4-neighbor connected; - it contains 8–40 black pixels; - it has 8–90 horizontal/vertical transitions; - it has at least one graph edge. These are **engineering filters**, not linguistic truth criteria. They favor compact, connected, graph-expressive forms while rejecting empty, extremely dense, or trivially unstructured bitmaps. ## Sample Default run: `interpretability_sampler 5000000 utf4088-interpretability-sample.csv` The seed is deterministic (`0x40880812`) so subsequent runs are reproducible. The CSV contains per-sample topology measurements and candidate classification. The executable prints the observed rates. ## Interpretation The candidate rate is the quantity to compare against the 4-billion target after extrapolation. Because the sample is intentionally random over the 96-bit space, it does not establish semantic meaning. A later corpus stage must compare candidates against the 16,606 curated symbols, historical forms, graph structures, and language priors. The resulting model is therefore: `96-bit capacity -> graph-valid candidates -> visually constrained candidates -> historical/language matching -> interpreted namespace` A successful candidate-rate test establishes feasibility of the intermediate population; it does not claim that arbitrary generated glyphs are human language.