Verification Doctrine

Measure. Compare. Verify.

A performance improvement only counts when the workload remains comparable and the declared quality threshold is met.

Methodology

Same task. Same constraints. Different context path.

The Performance Benchmark runs a declared baseline and treatment against identical workloads, records context and cost deltas, and validates that quality remains above the threshold.

01 / measure

Measure

What did the agent actually consume? Provider usage, cache hits, tool calls, context volume, latency, cost.

02 / compare

Compare

Same task, same source revision, same constraints. Baseline and treatment side by side.

03 / verify

Verify

Quality threshold held? Receipt valid? Methodology known and reproducible?

Results

Observed performance under controlled conditions.

Baseline

Stock agent, no optimization

$1.00

five-turn coding task · reference cost

Treatment

LeanCTX context path active

$0.17

same task · same quality threshold

Reduction

Context + cost overhead

↓ VERIFIED

−79–86%

across controlled workloads

Real OpenAI API calls · Provider caching recognized · Five-turn multi-file coding tasks · Quality floor enforced

Canonical Receipt

Every run produces verifiable evidence.

benchmark / run 0042● verified
workloadpayment-review-v1
modelgpt-4o-mini
input (baseline)139,115
input (leanctx)14,016
cost (baseline)$0.370
cost (leanctx)$0.058
qualitypass / pass
cost delta−84.3%

Run Your Own

Benchmark your agent in minutes.

terminal
# Install LeanCTX
pip install lean-ctx

# Run a controlled benchmark
lean-ctx evidence realworld \
  --baseline "stock" \
  --treatment "optimized" \
  --workload "./tasks/code-review.yaml"

# Generate verifiable report
lean-ctx evidence report --format html

Workload-specific result. Not a universal savings promise.

Context Performance Infrastructure

Prove it before you ship it.

Install LeanCTX