For AI agents: an index of this guide is available at https://www.plungeai.com/guide/llms.txt. Append .md to any page URL, or send Accept: text/markdown, to get markdown. The developer documentation is at https://www.plungeai.com/docs/llms.txt.

Documentation Index: fetch the complete documentation index at /llms.txt. Use this file to discover all available pages before exploring further.
Feedback

The automatic QA gate checks every concept before a person sees it. It never sends work to a client by itself: a person decides every piece the gate passed or held, and can dispute or re-queue anything it failed.

TemplateRules (each judged by two identical copies)
Ad creativeBrand, claims, ad policy, brief
Landing pageBrand, claims, accessibility and structure
Lifecycle emailBrand, claims, CAN-SPAM
Video clipBrand, policy, plus picture checks at 1, 4 and 7 seconds; when that check cannot run, a person watches the clip

Diagram 4. How the gate decides each concept. Held concepts go to the Review queue; failed ones are listed on the request page, where a reviewer can dispute them. Pieces being checked go through deterministic checks, rule votes with two copies each, and visual QA of each image. A one to one split gets a tie-break copy. Each concept is then decided: fail first, then hold, else pass. A pass goes to internal review, a hold to a person, and failing concepts of an ad set get one free repair that is checked again. Each piece also gets a shadow verdict measured on the Quality page.

  • Fixed checks run without a model: the block list (banned words, competitor names, required legal lines on pages and emails), numbers no approved claim covers, the client's website link, image facts, and the AI restrictions the client set.
  • The vote. Two agreeing copies decide a rule. A 1–1 split gets one tie-break copy; if that cannot run, the concept is held for a person.
  • Visual QA looks at each image for artifacts, text or logos inside the image, people, and a product shown wrongly. A person in an image is flagged to the reviewer. If the check cannot run after three tries, the concept is held.
  • One repair. When concepts of an ad set fail, the gate starts one repair run for them with the reasons as a steer. It is never billed. If a repaired concept fails again, an exception opens.
  • Shadow verdict. Each piece also gets a “would pass / would fail” verdict that the Quality page compares with the people's decisions.

Dashboard Quality page with agreement tiles, overrules and eval runs

Dashboard › Quality: agreement with people, severe misses, the golden set and eval runs.

The Quality page (admin) shows how often the gate agreed with people, and the severe misses: pieces the gate would have passed that a person sent back or rejected for a claim or a policy reason. An overruled or disputed piece can be added to the golden set. Eval runs replay the golden cases against the current rules; a client's workflows are saved only when every one of its templates passed at least 90% per rule. This is the evals gate. Only the owner can skip it, and must give a reason. Evals spend is capped by a ruling (default $25 a month).