An earthquake has struck Kumamoto, and various government agencies are issuing numbered bulletins. So far, bulletins #3, #13, #23, and #33 have been released. Bulletin #3 reported that a water outage occurred in a certain town, while Bulletin #33 reported that service had been restored. To determine which municipalities’ situations have changed, it is necessary to manually sift through 250,000 characters, a task currently being handled by municipal employees working the night shift.

Vera reads PDF files and extracts the following four elements without confusion: points where information sources match, points where information has changed, points where opinions differ, and points where no one has provided an answer. Since every claim is linked to a source file, it is possible to challenge the engine. This is the only way to verify that the engine is incorrect.

The response process involves no models whatsoever. It uses no embeddings, no sampling, and no GPUs. It always produces the same results from the same documents and operates offline. This is a crucial point, because what municipal officials actually have in front of them are evacuation shelter registries and lists of hospitals. There is simply no place to “store” them “somewhere.”

If the basis for a response cannot be identified, the system returns a type-specified rejection response (UNKNOWN_NO_EVIDENCE, UNKNOWN_LOW_COVERAGE) rather than a fluent sentence. This system does not write, summarize, translate, or chat. That is precisely the trade-off made to ensure its other capabilities.

Measurement Results (including failure cases):

  Japanese government disaster reports, 5 corpora, 4 blind tests
    14 detections, 14 correct
  Simple keyword-based baseline for the same documents
    38 detections, 6 correct — 16%
  Technical documents, 93 mixed English-Japanese project documents
    5 detections, 0 correct — 0%

The last line is intentionally included on the first page of the README. The dividing line lies not in the subject matter, but in whether the document asserts a state regarding “proper nouns”—such as municipalities, routes, contracts, assets, etc. In prose, the same abstract noun appears repeatedly in unrelated contexts, and comparing the two mentions creates contradictions that did not originally exist. Wikipedia is an inappropriate source, and pointing that out requires less effort than spending an afternoon verifying it.

One thing I did not anticipate was that this system would end up “correcting” its own “readers.” The system reads the same document twice through conversion processes that cannot alter meaning (such as filling in spaces inserted by a PDF extraction tool). Then, any assertion that appears in only one of the two readings is judged to be provably false, since the layout contains no information about it. Since this requires neither an answer key nor a human, after re-evaluating all verified findings, these corrections are made automatically. Determining what a word “means” still requires human judgment. That part is handled via a queue with an approval button, and no element on the vocabulary path can be written to the configuration file without approval.

The demo above is intentionally running on fictional data hosted on a public server. Please do not paste any confidential information. The local installation version does not perform any network communication whatsoever, and this is the version that municipal employees will actually use.

  Kumamoto Board, actual output   https://verantyx.ai/
  Installation, 2 minutes                https://verantyx.ai/vera/download/
  pip install verantyx-vera         Then: vera field
  Source                            https://github.com/Ag3497120/Verantyx

Both the app and the CLI mix Japanese and English throughout. If you have any questions, I’d be happy to answer them—including the one about the 0% row. That’s the number I hope is the most wrong.