Reproducible benchmarks

The numbers, and how to reproduce them.

Every figure below comes out of the benchmark directory in the open source repo. Clone it, run the suite, and check our math. We keep two things separate on purpose: hard counts we measured, and scale scenarios we modeled. Both are reproducible. Only one is field telemetry, and neither is yet.

Measured

Hard counts and direct measurements.

These are counted or measured directly, not extrapolated.

100.0
Gauntlet v2 score, out of 100
1,185
Tests passing
76%
Smaller context than naive, full coverage
31ms
Render, 5 directives

Gauntlet v2: 100.0 / 100

The full directive and adversarial suite, scored on a fixed rubric. Source: benchmark/gauntlet/.

Context size: 76% smaller

On the compose fixture, naive whole-corpus context is 1,832 tokens. A Perseus compile is 434, and it still answers all four test questions where tuned retrieval drops facts. Source: benchmark/compose/.

1,185 tests passing

The product test suite on the current release. A plain count, run it yourself with pytest.

Scale scenarios (modeled)

What happens at scale, computed from the data.

These come from scripts that model larger workloads (tens of thousands of directives, a 500 developer week) using measured per-operation costs. They are reproducible, but they are projections, not production telemetry. We label them that way so you can weigh them honestly.

450×
Cold to warm, 50,000 directives
1.0
Semantic equivalence, 20/20 A/B pairs
611×
Cold to warm, 22,528 directives

450× cold to warm

At 50,000 directives a cold render takes 612.6s and a warm one takes 1.36s. The cache absorbs the cost. Source: benchmark/cold-vs-warm.json.

1.0 semantic equivalence

A live judge scored every A/B pair semantically equivalent: same answers from the resolved context as from the raw workspace state. Source: benchmark/ultimate_suite_results.json.

611× cold to warm

22,528 varied directives against the real Perseus repo: cold 619.1s, warm 1.014s. Source: benchmark/real_deltas.json.

Reproduce it

Run it yourself.

~/perseus/benchmark
# clone the repo and run a suite
git clone https://github.com/Perseus-Computing-LLC/perseus
cd perseus/benchmark

# compose: context size and answer coverage
python compose/run.py

# cold vs warm at scale
python run_coldwarm_heavy.py

# the gauntlet
python gauntlet/gauntlet_orchestrator.py

All data files and scripts live in the repo. If a number here does not match what you get, open an issue and we will fix it.