NivareQ · Standalone laboratory

RAM Intelligence Lab

Semantic Spike S3 · RAM-Aware Retrieval Architecture

Idle
Isolation contractS3 is not RAM. RAM v0.8 Build 16 is a read-only architecture reference only. This lab cannot mutate RAM canonical storage. S2 remains frozen; S3 adds a separately versioned held-out judge.

1 · Model

Not loaded.

2 · Evaluation suite

—queries
—memories
—semantic Hit@1
—RAM-aware Hit@1

Architecture comparison

Every query is scored three ways using the same embeddings: raw semantic ranking → conservative lexical/fuzzy fusion → bounded deterministic RAM-aware reranking.

—fusion Hit@1
—aware nDCG@5
—aware Δ Hit@1
—rescued / regressed
CaseFamilyQueryExpectedSemantic topFusion topRAM-aware topSemAwareWhy / note
Load a model and run a suite.

What S3 is testing

S3 does not ask an embedding model to become canonical intelligence. It tests whether a good local retriever can provide candidates while deterministic RAM-style evidence handles spelling, exact identifiers, exclusions, temporal/date-space distinctions, state/currentness, recency, and answer form. The S3-H1 benchmark is frozen separately from S2 and should be treated as the held-out judge for this architecture revision.

Device / execution telemetry