Random

Overview

random picks one model uniformly from a decision's eligible candidates. The shared eligibility filter runs first, so the draw only ever sees models that are already valid for the request; the selector itself adds no filtering, fallback, or recovery of its own. Every candidate has the same chance, and no state carries between requests.

Key Advantages

• Produces an unbiased traffic split with no metrics, artifacts, or training.
• Has no selector model, storage, or external dependency.
• Gives other selectors a control group to be measured against.

What Problem Does It Solve?

Comparing two selection policies is only meaningful against a baseline that carries no preference of its own. Any ranked selector — even a deliberately naive one — encodes an assumption about which model should win. random encodes none, so differences you measure downstream belong to the policy under test rather than to the baseline.

It is also the simplest way to spread load evenly across interchangeable candidates when you have no signal that would justify preferring one.

When to Use

Use Random when running a routing experiment or evaluation that needs a control arm, when collecting unbiased samples across several models, or when candidates are genuinely interchangeable and any choice is acceptable. Prefer static when you want a deterministic, auditable choice, and multi_factor or latency_aware when you have quality, cost, or latency signals worth acting on.

Configuration

algorithm:
  type: random

The type takes no configuration block. Setting minimum_candidates is worthwhile — it makes the decision fail in a typed, visible way when the eligibility filter leaves too few models to draw from, rather than quietly degrading to a single choice. See a complete example: config/fragments/algorithm/selection/random.yaml.

Dependencies and Limitations

• No external dependencies and no request-content processing beyond ordinary decision matching.
• There is deliberately no seed setting. Routing stays non-deterministic in production; reproducing a specific sequence of choices is not supported. If you need to replay routing decisions, record them rather than trying to regenerate them.
• Every modelRefs entry is one equal slot. weight is not read, so this is not weighted traffic splitting, and two LoRA variants of the same base model count as two independent candidates.
• It does not fail over to another candidate, react to load, or learn from outcomes. Backend availability remains the responsibility of the normal provider path.