reading 150,000 vectors from the arXiv-150k fixture…

oneground · the choose door of oneproof

This is the ground. Every dot is a real vector.

150,000 arXiv abstracts embedded with bge-base-en-v1.5, cut into 256 regions by k-means. Semantic sharding copies a vector into every region within ε of its nearest. Move ε and the copies are recounted from four real distances per vector — not from a lookup table.

2-D placement is UMAP and is illustrative. Regions, distances and copy counts are computed in the full 768 dimensions.
what ε does

—

— vectors copied
— storage
— p99 copies

copies per vector

colour by

—

—

one query, every hop

Route it to one region and watch the neighbours you lose.

2,000 held-out queries, each with its exact top-10 neighbours from a brute-force scan of all 150,000 vectors. Follow one: score, route, and see where its neighbours actually live.

12 curated queries

nine worst-recall ambiguous queries, three that route well

drawn from all 2,000

pick a query

—

—

  1. 1score against 256 centroids—
  2. 2route to the nearest region—
  3. 3the ten true neighbours—
  4. 4what the region actually returns—

the verdict

—

—

—

—

the receipt

Three builds, four environments. Values agree, bytes don't.

—

the sampling receipt is identical

the same records, chosen by the same seeded rule, in every build

the vectors are not

embedding.vectors_sha256 after each canonical build

the published values agree anyway

—

check it yourself

—

verifying the full arXiv fixture is heavier — a checkout, the pinned environment and the release asset: how the checks work.

oneground 0.1.0-preview on GitHub — the release these numbers were produced under.

—

what is coming

oneground 0.1.0-preview — the advisor — installable today.

pip install oneground · characterize · simulate · verify (Qdrant) · report.

Next, undated — it ships when it is ready and tested: pgvector as a second engine, measured in turn on one host; a second fixture where the ground looks different; a latency verdict that says couldn't-check when three runs straddle the threshold.

Every number carries its receipt; every report cites the calibration run it was generated under.

v0.2: more engine adapters behind the same protocol, and chunking measured rather than asserted. Dated when a witnessed run makes it true.

Prevent · Detect · Prove · Choose.