Pipeline run report

Anyone can make an ad with AI now. The hard part is knowing which ads a brand can actually run. This pipeline writes ad copy for one brand’s market (Huggies, Singapore), then forces every piece through a gate before a human sees it: ten fixed rules refuse anything the brand could not defend, an AI judge scores the survivors for brand voice, and a small repair agent rewrites what failed and sends it back through the same gate. The rules are not guesses. Each one comes from the brand’s own published fine print, Singapore’s advertising code, past regulator rulings, or real customer complaints. This page is the record of one real run, rendered from the run’s own output files.

• Eval results  the checks the system must pass before any run is trusted: copy that must be refused, copy that must pass, and repair-behaviour tests, including one that catches the agent smuggling a banned claim past the rules in new words.
• Repair  the one place an agent lives. It reads the rule that stopped an ad and the evidence behind it, rewrites, and resubmits to the same gate. Two attempts, a spend cap, and no power to approve its own work.
• Refusal log  every ad this run stopped, the rule that fired, and the before/after wherever the agent repaired it.
• Batch sheet  all 48 ads this run generated, with each one’s final verdict. The ads are outputs of this pipeline, written for the demo; none are real advertisements.
• Unit economics  what the run cost, measured from actual tokens, expressed per 100 gated ads.

The ten rules themselves are listed on the try-the-gate page, where any line of copy can be tested against them directly in the browser, with no model call.

The dark cards are what the shipped demo already does; the light cards are the pipeline around it on a real engagement. Every card is tagged with who does the work: a model, deterministic rules, or a person.

Known-bad fixtures must be refused by the deterministic rule layer (no model involved); known-good fixtures must survive the full gate including the strong-model voice judge. The first row is the brand’s own published site copy. Repair fixtures run the bounded agent end-to-end — a salvageable refusal must come back PASS, and an unshippable claim must never sneak past as a paraphrase (the evasion probe checks wordings the gate’s own regex does not cover).

Refused and off-voice variants go to a repair agent that reads the rule that fired and the review evidence behind it, rewrites, and resubmits. Bounds: 2 attempts per variant, a per-run spend cap, brand and angle pinned, and no verdict authority — every rewrite re-enters the same gate, and whatever is still failing lands in the human queue with its full history. Each card below is one salvage from this run: the copy as generated, the rule it broke, and the rewrite that passed.

Every refusal names its rule and the evidence behind it. Entries the repair agent salvaged are shown in the Repair section above; what remains here is the live-copy audit and anything that stayed refused or off-voice.

Every variant below was generated by this run (corpus → briefs → variants on the cheap model tier), then gated and, where needed, repaired; verdicts are final, post-repair. All of it is demo output; none of it is a real advertisement.