Kumamoto earthquake · August 2026 · a working example
33 bulletins. Which towns have water back?
Vera reads the actual ministry PDFs and answers with sources — or refuses, with a typed reason. No LLM in the answer path. Deterministic. Offline. Below is its real output on the real documents.
Input: MLIT bulletins #3, #13, #23, #33 (Jul 29 – Aug 7) · 61,083 characters
Read: 2,319 / 2,701 sentences (85.9%) · 6 comparable pairs · 6 findings, 6 true
This is the engine’s verbatim output rendered as HTML — not a screenshot and not an illustration. Every line names the file it came from, so you can disagree with the engine, which is the only way to find out it is wrong.
How it works
The name is the architecture: knowledge lives on stereo crosses, and disagreement is a geometric event.
1 · Crosses, not embeddings
Each named thing gets one cross: a core and accumulating facets. A state word is stored as aspect:value — 復旧:断水 — so two poles on one aspect surface as a contradiction by structure, not by similarity score.
2 · A subject gate, measured in
A pole lands only when the named thing is the subject of the sentence that carries it. Added after measuring 0-of-4 precision without it; with it, 14 of 14 findings on five disaster corpora were true.
3 · Typed refusal, never a guess
No model anywhere in the answer path. When evidence is missing the answer is UNKNOWN_NO_EVIDENCE — a name for what is missing, not a fluent sentence about it. Same input, same output, offline.
4 · It repairs its own reader
It reads the same documents twice through transforms that cannot change meaning; a claim that appears in only one reading is provably spurious and gets repaired unattended. What a new word means still requires you.
What would you point it at?
The boundary is measured, not guessed: it works where named things change state across disagreeing sources.
Disaster information desks
Bulletin #3 says the water is out; bulletin #33 says restored. Which municipalities changed, which are still contested, and which report said what — without reading 250,000 characters by hand.
Ledgers: contracts, permits, assets
Anything named that flips state — valid/expired, running/stopped, open/closed — across documents that disagree. The vocabulary is 12 oppositions and grows by approved proposal.
Auditing agent declarations
An agent that says "sandbox on" while requesting "sandbox off" is two sources contradicting each other about a named thing — structurally the same detection, measured to work on typed declarations.
Not: wikis, meeting notes, prose
Measured on 93 technical documents: 5 findings, 0 true. Abstract nouns recur across unrelated contexts, so comparing two mentions manufactures contradictions. We publish that number on purpose.
Measured, including the failure
| corpus | findings | true | precision |
|---|---|---|---|
| Government disaster reports (5 corpora, 4 blind) | 14 | 14 | 100% |
| Naive keyword baseline, same documents | 38 | 6 | 16% |
| Technical prose, 93 documents | 5 | 0 | 0% |
The 0% row is why the page above says “not for wikis”. Publishing where a tool fails is cheaper than an afternoon of your time finding out.