How do I read a critique result?

A critique result is short on purpose. It tells the agent whether the output met the rule, which words to change if it did not, and what is left to spend. Read it the same way every time.

Pass or fail, then the quoted passage

Start with result. On a pass, the agent hands the output over. On a fail, the result quotes the passage that broke the rule. The agent rewrites that passage, keeps the rest, and runs the check again. There is no score, so there is nothing to negotiate: the rule is met or it is not.

For a rule set

You get one pass or fail for the whole set, plus failed_rules, each with its ID, name, and version, and rules_run, every rule that was judged with its own result. Fix the failed rules first. The set passes only when every rule passes.

The loop fields

• attempt and attempts_left show where the agent is in its loop. The agent should warn the person before the last attempt.
• version_id says which version of the rule or set ran, so the result can be traced after the output is gone.
• credits_left is the balance, and low_balance appears when it is low. Company admins get an email when the balance crosses the threshold.
• validation_id is the reference to quote to support if a result is disputed.

When a call is refused

A refusal comes back before anything is charged and says what to do. The ones you will meet first:

Two outcomes happen after the credit is held: ENGINE_FAILED and TIMEOUT. Both say nothing was charged, and whether to retry. The critique reference lists every code.

Common questions

What does pass or fail mean?

A pass means the output met the rule as written. A fail means it did not, and the result quotes the passage that broke the rule. A pass is not a guarantee that the output is correct, only that it met this rule.

What is the quoted passage?

The exact words from the output that broke the rule. It is what the agent should rewrite, and nothing else.

What do attempt and attempts_left tell me?

Where the agent is in its loop. When attempts_left reaches 0 the next call is refused and nothing is charged, so the agent should stop and report.

What is the validation ID?

The identifier of that critique. If a result looks wrong, give it to support with the output, and support can find the exact rule versions that ran.

What does a rule set result contain?

One pass or fail for the whole set, the rules that failed with their versions, and every rule that was judged with its own result. The set passes only if every rule passes.