Core concepts

Tracira sits between your AI automation and your customers. Your automation produces an output, you send that output to Tracira, Tracira checks it against your rules, and anything risky is held back for a human to approve or reject. This page explains the pieces involved and how they fit together.

The building blocks

Everything in Tracira lives inside a few nested objects:

Workspace
└─ Project  "support-replies"
   ├─ Task  "draft"
   │  ├─ Log
   │  └─ Log
   └─ Task  "summary"
      └─ Log

• Workspace is the top-level boundary. It holds your members, billing, and integration settings (webhook token, API key, Slack, model keys). Every other object belongs to exactly one workspace.
• Project groups everything for one automation or use case. It collects logs and can carry its own rules. A project has a name and an optional icon.
• Task is an optional layer inside a project. Use it to separate different steps or output types of the same automation, for example a "draft" step and a "summary" step.
• Log is a single AI output you sent for evaluation. It carries the output text (and optionally the input that produced it), the model, cost, latency, any metadata you attach, and the evaluation results described below.

Rules

A rule is one condition that every matching output is checked against. Rules come in two families:

• Deterministic rules run in code: exact, instant, and free. They cover keyword required or forbidden, regular expression matches, length limits, and checks on a field inside JSON output (a value exists, equals something, falls in a numeric range, is a boolean, or carries a date inside a window). No AI is involved, so they cost no credits.
• LLM judge rules are written in plain English, for example "Flag if the reply promises a refund". An AI model reads the output and decides whether the rule passed or failed. These handle nuance that exact matching cannot, and each evaluation costs credits.

Every rule applies at one of three scopes, which decides the logs it runs against:

When a log is evaluated, every rule whose scope matches runs against it.

What you get back

For each log, every matching rule produces an outcome: passed, failed, or skipped (skipped when the rule could not apply, for example a JSON field it needed was missing). Those individual outcomes roll up into three separate fields on the log. They are easy to confuse, so it is worth keeping them straight:

Each log also gets a short plain-English explanation summarizing why it landed where it did.

Human review

A flagged log waits for a person. Reviewers see it in the dashboard, and you can also be notified by Slack or email. The reviewer makes one of three decisions:

• Approve: the output is fine, let it through.
• Reject: the output is bad, block it.
• Change: send the output back to the AI with a written instruction. Tracira records the new attempt as a revision linked to the original.

The decision is stored on the log, including who made it and when. If a step in your workflow is waiting on the result, Tracira can call back once the decision is made.

The lifecycle

Putting it together, a single output moves through four stages:

Next steps

• Connect your automation: Make, n8n, or any HTTP client.
• Read the full API reference for every endpoint and field.
• Review rate limits before you go to production.