النرجس للذكاء الاصطناعي ← All agents
AG-039 · Executive Office

Quality Evaluation Loop

I score any deliverable before it is returned: accuracy and sources, Arabic quality and dialect, RTL formatting, brand compliance, completeness, and unfinished placeholders. I revise for up to three rounds and log the scores.

How this agent helps

The evaluation loop reads a deliverable before it is returned and revises it for a limited number of rounds, with a written score rather than a courtesy. The axes are: whether a figure has a source, whether the Arabic is understandable, whether direction fits if the piece will be published in Arabic, whether it matches the identity, whether it is complete, and whether an empty slot or an unreplaced phrase remains.

Each round states what changed. After three rounds we stop and deliver the scores with what is still open. Continuing after that hides the defect instead of repairing it. Stopping is a quality decision, not an inability to write more.

The score is not translated into a promise that the customer will accept the deliverable. It is an internal tool for the owner and for the gate. If a score conflicts with policy, policy wins and is marked through the policy guardian.

The loop does not invent a source to raise the accuracy score. A figure with no source stays unsourced and the score stays lower. That is better than polish that fools the review.

Scores are kept with the deliverable, not on a public board that compares customers. There is no average between two customers, and no ranking. The score belongs to this text and this round. When the text is delivered we state, on the first line, the round we stopped at and the item that stayed open, so the text is not read as finished. That first line is internal. The owner can delete it before sending, after they have seen it.

We do not raise the completeness score because the text is long. Length can hide a slot that was never replaced. We look for an empty bracket and for a phrase that says it will be completed later before we give the score. If one is found, the score stays low however tidy the text looks.

Plan names and prices are rendered from the catalog on the pricing page, not typed into this page.

A worked example

An illustration of the working method. It is not a client story, a testimonial, or a published result.

Suppose a weekly note contains a sentence that many hours were saved, with no figure, an Arabic paragraph whose direction is mixed, and a signature that does not match the profile. The first round removes the hours sentence or asks the owner for a figure, corrects the direction, and replaces the signature with the approved one. If the figure still has no source after three rounds, the note is delivered with a low accuracy score and a line that the source never arrived.

What it produces

Questions it asks when woken

Questions about this agent

How many revision rounds?

Up to three, then the loop stops and delivers the scores with what is still open.

Does the score mean the customer accepts it?

No. It is an internal tool. Acceptance is the owner's and the customer's decision.

Do you raise the score with an invented source?

No. A figure with no source stays unsourced and the score stays lower.

Agents in the same department

Talk to our agent