Core Concepts

A guide to the key objects and terms in Reliai.

Traces

A trace is the raw execution record of one request through your AI system.

Each trace captures:

• the input and output
• latency
• tool calls and retrieval steps
• guardrail and policy events
• errors

Traces are stored exactly as recorded. They are never modified, summarized, or filtered by Reliai.

Incidents

An incident is a detected regression — a change in system behavior that crosses a threshold.

Reliai groups into an incident:

• the failing traces that triggered it
• the metric signals that changed
• the root cause signals computed from the evidence

An incident has a lifecycle: open → acknowledged → resolved.

Root Cause

Root cause is the explanation for why an incident occurred.

It is computed deterministically from:

• trace comparisons (failing vs. baseline)
• prompt or model version changes
• clustering of failure signatures

Root cause is not generated by AI. It is a signal derived from the data.

AI may be used to explain root cause in plain language, but the underlying determination is always deterministic.

Evidence

Evidence is the set of signals attached to an incident that explain what happened.

Evidence includes:

• metric deltas (what changed and by how much)
• trace comparisons (what the failing traces look like vs. baseline)
• prompt or version diffs (what changed in the system)
• resolution impact (whether a fix worked)

All AI features in Reliai are grounded in evidence. AI does not generate signals — it explains them.

Resolution Impact

Resolution impact measures whether a fix worked.

After a fix is applied, Reliai tracks:

• whether the failure rate declined
• whether metric signals recovered
• how post-fix traces compare to pre-regression baseline

Resolution impact is shown in the incident command center after a fix is applied.

Use resolution impact — not AI summaries — to decide whether an incident is resolved.

Deployments

A deployment is a tracked change to your AI system — a new prompt version, model version, or configuration change.

Reliai uses deployment records to:

• identify what changed before a regression
• surface version diffs in root cause analysis
• gate deployments based on reliability signals

Guardrails

Guardrails are runtime policies applied to your AI system's inputs and outputs.

Reliai tracks:

• which guardrail policies are active
• when they trigger
• what they blocked or flagged

Guardrail events are captured in traces and surfaced in incident analysis.