NivareQ · Standalone laboratory

RAM Intelligence Lab

Semantic Spike S3-R3 · Targeted Regression Closure

Idle

Isolation + holdout contractS3-R3 is not RAM and cannot mutate RAM canonical storage. Candidate generation remains frozen to S3-H1. H1/H2/H3 are exposed historical evidence. The S3-R3 ranker was hashed before H4 was created. H4 is a targeted fresh held-out regression/neighbor judge; do not tune R3 after seeing H4 results.

1 · Model

Candidate

BackendPrecision

Not loaded.

2 · Evaluation suite

Suite

—queries

—memories

—semantic Hit@1

—S3-R3 Hit@1

Architecture comparison

Same embeddings and exact same frozen Top-12 candidate union: semantic → fusion → frozen H1 aware → frozen R1 → frozen R2 → S3-R3.

—fusion Hit@1

—H1 aware Hit@1

—R1 Hit@1

—R2 Hit@1

—R3 nDCG@5

—R3 Δ vs R2

—R3 rescue / regress

Case	Family	Query	Expected	Semantic top	Fusion top	H1 top	R1 top	R2 top	R3 top	R2	R3	R3 why / note
Load a model and run a suite.

What S3-R3 changes

R3 keeps semantic embeddings, fusion, candidate generation, H1, R1, and R2 frozen. It is deliberately narrow: hyphenated and spaced measurements produce equivalent typed roles; neutral measurement queries can penalize conflicting values for the same unit without affecting causal/transition explanations; and temporal/date evidence owned by an explicitly excluded contrast branch remains suppressed instead of being reintroduced downstream. H4 was created only after the R3 ranker hash was frozen. H4 is targeted, so its aggregate percentage is not directly comparable to the 100-query H1/H2/H3 suites.

1 · Model

2 · Evaluation suite

Architecture comparison

What S3-R3 changes

Device / execution telemetry