AI03 · VECTOR CONTRACT LAB

Same shape. Same ranking?

Change the query, server metric or axis fault. Advance the calculation to see transformations, eligibility and ranked results. These vectors are synthetic, not embeddings from a model.

Query xQuery yQuery zServer scoreServer transformUnsafe server includes Red document

Query vector

Local contract

Server contract

same first result?

same eligible set?

max query-component difference

cosine(q,d) = dot(q,d) / (norm(q) * norm(d))
eligible = public || document.tenant == trustedTenant
rank = score descending, then document ID

A, B, C are Blue; R is Red. Bars show absolute score magnitude within each path; ochre means negative. Local uses cosine. No ONNX, tokenizer, network or semantic quality is measured. Zero-vector cosine is refused.

Backward seeking reviews this calculation; applied model events are not undone. Change inputs or Reset for a fresh scenario.