Not a language model. No weights, no sampling, no training: 1,200 Japanese statutes and 2,591 encyclopedia articles are read into a federation of cross-shaped nodes, and the same question always produces the same answer.

The artifact is one SQLite file. Nothing is unpickled, and every number below can be checked against the file itself.

Answers with the source it read, and when it cannot answer, says which kind of not-knowing it is — because the kinds need different things done about them.

Reproduce all of it with python3 -m verantyx.card_numbers --db vera.db.

Closure guarantees the store never emits a symbol it does not hold. It guarantees nothing about the subject, and the first release showed it: on 200 invented compounds, 77% were answered about a recognised substring — ヒュペリオン数人 answered about 数人 — with the unknown element dropped without a word.

The current release gates the seed on the asked subject. A seed passes only when it is the subject, contains it, or holds it on its own cross's faces; a held subject the staircase overlooked replaces the seed; anything else refuses by name:

ヴォルフガング粒子とは → UNKNOWN_NOT_PRESENT
                        subject: ヴォルフガング粒子, nearest held: ヴォルフガング

Measured: invented compounds answered 77% → 4% (the residue enters through direct Latin-substring retrieval, not the staircase); wrong-subject answers on suffixed questions (〜の要件は, 〜について教えて) 43–71% → 0–7%; correct-subject answers rose at every suffix (57%→86% bare, 29%→36% worst case) because a held subject now displaces a worse seed. The gate is conservative and it does lose borderline coarsenings — 不法行為とは now refuses with a pointer to 不法 instead of answering from it.

Still true and unchanged: it cannot explain a word it never read (0.0% — the same closure that produces the 60/60), summarise, or chain inferences past one step.

Also absent: explaining a word it never read, summarising, chaining inferences past one step, and fluent prose. The first follows from the same closure that produces the 60/60; the others are stated with their measurements in the module docstrings.

The code is MIT. The built structure in vera.db is derived from two corpora with different terms:

vera_edges.db (87MB, optional) holds same-sentence facet pairs — the edges of the cross, where a face holds an item and an edge holds the relation one sentence actually wrote. The engine answers identically without it; with it, evidence-tied cores may speak the pairs a sentence attested (25 of 97 silenced cores recover speech). Same licence as vera.db.

corpora/*.json pin name, url, sha256 and byte count for all 3,958 documents, with the selection rule recorded beside them. That is what makes the figures above checkable rather than quotable:

Full platform document (structure, signals, governed evolution, versioning, participation): https://github.com/Ag3497120/Verantyx-Vera-alpha/blob/main/docs/PLATFORM.md

Code: https://github.com/Ag3497120/Verantyx-Vera-alpha