Deterministic Mode

Reproducible evaluations with fixed temperature and seed.

Deterministic mode attempts to make model outputs reproducible by setting temperature to 0 and using a fixed random seed.

How It Works

When enabled on a suite, PeerLM checks each model's capabilities before sending the request:

• If the model supports temperature control, it's set to 0
• If the model supports seeding, a fixed seed value is included
• If a capability isn't supported, that parameter is omitted

What Gets Stored

Each response records what was actually used: temperature_used, seed_used, and deterministic_attempted. This provides a full audit trail regardless of whether the model honored the settings.

Limitations

• Not all models support deterministic output — some may still produce varying responses even with temperature=0
• Seeding support varies by provider. PeerLM records whether it was attempted so you can assess reliability.