LLM Explainability Patterns

1. Token-level confidence via logprobs

The cheapest form of explainability is asking the model for its own probability distribution. Pass logprobs: true and a small top_logprobs value to every chat completion you intend to audit. Meridian forwards the field to Azure's router transparently, so the cost overhead is a few percent on output tokens.

// Pattern: token-level attribution via logprobs
import { Meridian } from '@meridian/sdk';

const m = new Meridian({ apiKey: process.env.MERIDIAN_KEY });

const result = await m.chat.completions.create({
  model: 'azure/model-router',
  messages: [{ role: 'user', content: 'Why is the sky blue?' }],
  logprobs: true,
  top_logprobs: 5,
});

// Each token carries the model's confidence + alternates.
for (const tok of result.choices[0].logprobs.content) {
  console.log(tok.token, Math.exp(tok.logprob).toFixed(3));
  for (const alt of tok.top_logprobs) {
    console.log('  alt:', alt.token, Math.exp(alt.logprob).toFixed(3));
  }
}

2. Counterfactual prompt probing

For higher-stakes outputs, run the same prompt with one input field ablated and compare. A material divergence tells you the field was load-bearing; a near-identical answer tells you the model ignored it. Meridian's model: azure/model-routeralias makes this cheap because routine probes route to small models while the canonical run goes to the full one.

3. Structured reasoning capture

For reasoning models (gpt-5 family, o4-mini, grok reasoning), set reasoning_effort to low or mediumand persist the visible chain alongside the final answer. Pair the stored trace with the logprobs from pattern 1 and you have a two-layer audit record: what the model considered, and how confident it was at each emitted token.