Context Evals
Test which context produces the best AI result
Parallel testing tracks. One fixed task; four lanes — Baseline, Full Context, Compressed Packet, Custom Route — race through Accuracy, Completeness, Grounding, Latency, and Cost checkpoints.
Parallel eval tracks
Same task. Four contexts. One winner.
Fixed task
Draft a refund reply that cites policy R-12 and the confirmed duplicate charge.
Draft a refund reply that cites policy R-12 and the confirmed duplicate charge.
Baseline
Acc 0.71Comp 0.64
Ground 0.58Lat 1.2sCost $0.04
Full Context
Acc 0.89Comp 0.92
Ground 0.80Lat 3.8sCost $0.31
Compressed Packet
Acc 0.93Comp 0.90
Ground 0.94Lat 1.1sCost $0.06
Custom Route
Acc 0.88Comp 0.86
Ground 0.91Lat 1.4sCost $0.08
Click a track to inspect input, result, and the basis for every metric — not just a final score.
Inspect · Compressed Packet
Input: 1,140-token packet · Output grounded on billing.get + policy R-12 · Latency win from dropped thread noise · Cost 5× below Full Context.
Pass The Right Context To Every AI Step
Context that arrives intact — not more context, the necessary context.
RelayLum sits between agents, models, tools, sessions, and memory. It turns raw inputs into Context Packets optimized for the next hop — compact, reliable, and fully traceable.