NivareQ · Standalone laboratory

RAM Intelligence Lab

Semantic Spike S1-A1 · frozen S1 benchmark · adapter + observability revision

Idle

Isolation contractThis is not RAM. It does not import or mutate RAM canonical storage. S1's 37-memory / 20-query benchmark is frozen byte-for-byte; S1-A1 changes the runner only.

1 · Isolated model run

Candidate

BackendPrecision

Not loaded.

2 · Frozen S1 screen

0cases

0memories

—Top-1

—MRR

3 · Screening batch

Small/medium candidates only. Each model loads → runs once → disposes → checkpoints. Stress remains isolated.

No batch checkpoint.

EmbeddingGemma, BGE-large, Qwen3-0.6B and 4B/8B are deliberately excluded from batch selection. Run them one at a time.

Results

Pure semantic retrieval. Top-5 detail and margins are included in exported JSON.

Case	Query	Expected	Top result	Rank	Top score	Expected score	#1–#2 margin	ms
Run a model to populate results.

1 · Isolated model run

2 · Frozen S1 screen

3 · Screening batch

Results

Device / execution telemetry