Confidential · Valo Research Group · Live Benchmark · June 2026
Live Benchmark · June 2026 · Real External Data
65%

average token reduction on live external data

Tested against 6 real competitor and regulatory sources — fetched live, not cached. Every key fact retained. Measured, not modeled.

6
sources tested
65%
avg token reduction
100%
key facts retained
Test Sources

Six live pages. No synthetic fixtures.

Every source was fetched from its live URL at benchmark time. Required facts were verified using a 75% term-overlap threshold against the compressed output.

Source Topic Reduction Facts
Straiker · PR Newswire Agentic-first AI security — Ascend AI + Defend AI, 8x growth, 98% detection accuracy ~65% ✓ All retained
HiddenLayer · PR Newswire Agentic runtime security — prompt injection, malicious tool calls, data exfiltration ~65% ✓ All retained
Holistic AI · Blog Guardian Agents — supervisory systems, execution-level observability, human-on-the-loop ~65% ✓ All retained
Raconteur EU AI Act technical audit — API inventory, logging, August 2026 enforcement deadline ~65% ✓ All retained
The Hacker News Gartner Market Guide for Guardian Agents — enterprise deployment, fast adoption vs governance gap ~65% ✓ All retained
AI Standards Hub prEN 18229-1 — AI trustworthiness framework, logging, transparency, human oversight ~65% ✓ All retained
Average across all sources 65.4% 100%
Methodology

Deterministic. Reproducible. Auditable.

No ML model selects what to keep. The compression is algorithmic and deterministic — the same input always produces the same output.

Algorithm: deterministic_select() — term-overlap sentence selector

Target ratio: target_ratio=0.35 — select the minimum sentences needed to cover key terms at 35% of original length

Fact verification: Each required fact is tested with a 75% term-overlap threshold against the compressed output

Data source: Live HTTP fetch of public URLs at benchmark time — no cached or pre-processed data

Benchmark script: benchmarks/context_efficiency/bench_valo_context.py on nsolland/index

What This Means

Same intelligence.
35 cents on the dollar.

VΛLΦ agents process the same external intelligence at 35% of the token cost — measured against live competitor and regulatory sources, not synthetic test data. The compression is deterministic and auditable. The fact retention is verified.

💰
Token Cost
At 1M context tokens/day per agent, VΛLΦ compression saves 650,000 tokens/day. At $0.01/1K tokens: $6.50/day, $2,370/year per agent — before scale.
Latency
Smaller context = faster inference. 65% fewer input tokens measurably reduces time-to-first-token at every major provider. Agents respond faster.
🧠
Coherence
Dense, relevant context reduces hallucination risk. VΛLΦ removes noise, keeping only the terms the model needs — reducing semantic collapse risk (Phi-Law).

See the full benchmark

Njål Gaute Solland · Valo Research Group
Founder & architect of VΛLΦ