Brand & claims gate for AI-written copy
Before an agent ships marketing copy, the gate checks every line against brand voice, approved claims, competitor rules and regulated wording. Each draft is auto-approved, sent for review or blocked, and the page shows why.
Adjust inputs ↓What the gate decides
Highlights mark each finding. Select one to jump to it.
Edit the draft; the verdict updates as you type
Each sample sets its own channel and market.
Each finding comes with a fix
What gets written to the ledger
The gate appends every decision to the ledger: which draft, which rule set, which rules fired. An auditor or legal reviews that record later.
Run an agent's drafts through the gate
An agent writes 20 variants. The gate sorts them before anyone looks. It uses the market and threshold set in the controls.
One variant per line. For multi-line variants, put a line with only --- between them.
Verdict per variant
Select a row to open that variant in the single-draft view.
| # | Variant | Verdict | Score |
|---|
Rules live in a file brand and legal can edit
Each rule has an ID, a category, a severity, a scope and a fix. Brand, legal and marketing ops change the file; the engine code stays the same.
Takes effect at once, on the draft and the batch. Rule TOV-CUSTOM, severity warn.
Approved claims
Any number, percentage or superlative in the copy must match a row here: same wording, same number, right market, not retired.
| ID | Approved claim | Markets | Status |
|---|
Rules that apply to this channel and market
Rule summaries are simplified.
How it works
The rules engine
Each rule is a row of config: an ID, a category, a severity (info, warn or block), the markets and channels it applies to, a matcher and a suggested fix. The engine parses the copy into fields (headline, description, subject and so on), runs every rule in scope, and returns findings with character positions.
Claims work differently. The engine finds every number that reads as a claim (a percentage, a currency, a "+", or a number next to words like hours, companies or faster) and matches its clause against the register by shared words and the number itself. Same topic and same number: approved, with the evidence ID. Same topic, different number: blocked as altered. No match: blocked as unsubstantiated. Retired claims and claims cleared for another market are blocked too.
The verdict is fixed logic. Any block finding blocks. Otherwise the warnings add up to a risk score, and the draft goes to a human if the score reaches the threshold. Info findings cost nothing.
In production
An LLM agent drafts the copy. This page does not run one; it runs only the deterministic layer that gates its output (same input, same answer, no model call). In production the gate sits between the agent and the ad platform or email tool, and runs on every draft:
- Blocked drafts go back to the agent with the findings as feedback, so it rewrites against the exact rule it broke.
- Review drafts land in a queue for a human, with the findings attached.
- Every decision is appended to a ledger: draft hash, rule set version, findings, outcome, and who overrode what.
- A weekly drift report shows which rules the agents keep tripping, so the prompt or the brief gets fixed at the source.
Limits
- Rules catch known patterns. They miss novel misleading framing, implied claims and context a reader would pick up.
- An LLM judge can add semantic checks on top, such as tone fit or implied promises. The deterministic layer stays the gate, because it is auditable and gives the same answer every time.
- Claim matching is word overlap plus the number. A clever paraphrase can slip past, and a rare false match is possible. Both are reasons to keep a human on the review queue.
- Readability uses the English Flesch formula. It means little for German copy.
- Legal owns the rule set. The regulatory rules here are simplified illustrations, not legal advice.