Datasetpublic

kofdai/Verantyx-Vera-base-store

Verantyx Vera — Base Knowledge Store The shippable artifact of a deterministic knowledge engine. Vera has no weights and no embeddings: what ships is the store — a stereo-cross index of cores and facets built by pouring text through the engine's own ingestion path. Code: https://github.com/Ag3497120/Verantyx (engine lives in engine/) git clone https://github.com/Ag3497120/Verantyx.git cd Verantyx/engine python3.11 -m verantyx.cli fetch-store # this dataset, ~209 MB… See the full description on the dataset page: https://huggingface.co/datasets/kofdai/Verantyx-Vera-base-store.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes89downloads

Dataset Card

Verantyx Vera — Base Knowledge Store

The shippable artifact of a deterministic knowledge engine. Vera has no weights and no embeddings: what ships is the store — a stereo-cross index of cores and facets built by pouring text through the engine's own ingestion path.

Code: https://github.com/Ag3497120/Verantyx (engine lives in engine/)

bash

git clone https://github.com/Ag3497120/Verantyx.git
cd Verantyx/engine
python3.11 -m verantyx.cli fetch-store            # this dataset, ~209 MB
python3.11 -m verantyx.cli --store vera_store.json ask "valkyria"

Nothing is downloaded implicitly. If a store is missing, the engine says so and names the command that would fix it:

json

{"verdict": "UNKNOWN_NO_STORE", "how_to_close": "vera fetch-store --repo kofdai/Verantyx-Vera-base-store"}

What is actually in it


Source	`hf:imdb` — English film, biography and geography prose
Cores	889,241
Facet links	9,778,919
Sentences ingested	5,663,792
Size	~209 MB JSON

What it is good for, measured

Ask it about something the corpus holds and you get a typed ANSWER with the core it answered from. Ask about something it never read and you get UNKNOWN_NO_EVIDENCE with a reason — not the nearest paragraph. That refusal is the point: a neighbour search always returns something, and a generator always mixes in words the corpus never contained.

What it is not good for — please read before judging it

Measured on 2026-08-22, on this exact store:

—It is English film prose, not a technical or domain corpus. Coverage of a coding assistant's vocabulary: 絵文字 0, テスト 0, 型注釈 0, print文 0, TypeScript 0, JavaScript 0. There is no Japanese in it at all.
—Several apparent hits are false friends. print here means printed matter — its structural siblings are online / magazine / newspaper. console is a games console. black is a colour.
—Sibling inference ("the substitution nobody wrote down") only works where the corpus put two alternatives side by side on the same line. Statutes do that; prose does not. Measured recovery on technical pairs: 1 of 6.

So: use this store to see the engine's behaviour — typed refusals, order-invariance, provenance — not as domain knowledge. For real work, pour your own:

bash

python3.11 -m verantyx.cli --store mine.json documents ~/your-docs/

Pouring is measured to be non-interfering: adding documents does not change answers that were already grounded, and the ingestion order does not change any answer (6-permutation check, part of the repository's 50-measurement suite).

Everything published here

file	what it is	size
`vera_store.json`	the base store — English film/biography prose (`hf:imdb`), 889,241 cores	209 MB
`stores/guard_store_technical_ja_en.json`	a technical ja+en store built from 549 local documents (4.9 M chars), used to measure whether a domain corpus fixes instruction reading — it does not, and that negative result is the point	55 MB
`stores/guard_store_manifest.json`	exactly which files went in, so the store can be rebuilt or disputed	142 KB
`stores/ide_default_store.json`	the small store the macOS app ships against (ja definitions)	6 MB
`corpora/*.json`	the ingestion corpora: e-gov statutes, Japanese/English Wikipedia field sets, Aozora speech, disaster-domain probes	1.1 MB

Every one of them is an artifact of a measurement written up in the code repository, not a curated dataset. The manifest exists so a reader can dispute the corpus rather than take the numbers on faith.

Provenance and licence

The store is a structural index derived from the public imdb dataset on this Hub. It contains cores and facet counts, not reproduced documents. The MIT licence above covers this derived index; the underlying corpus keeps its own terms.

Reproduce every claim on this page

bash

python3.11 engine/experiments/guard/verify_all.py
# forks 89/89 / 測定 50/50

Each number above traces to a pre-registered measurement in engine/experiments/ — written before the measurement ran.