The MCPs

Retina rewrites the first 10 to 15 seconds of a video, draws them and checks the fix before the creator publishes. The problem: on Facebook, retention curves fall hardest in the first seconds. Most tools read a video and hand back a grade. A grade does not tell an editor what to say first, and an AI rewrite that nobody checked can invent facts or add the engagement bait Meta penalizes. We built it at Badabun, a large Spanish-language media company in Mexico, with the retention curves of its own archive. What Retina does: you drop a draft and its title. Muse Spark 1.3 watches it and Muse Voice Transcribe listens. Retina decides whether the video is within its scope, then Muse Spark Open, Meta's open-weight model, writes a new opening that gives the first proof of what the title promises and leaves the rest to the video. The title is never touched. Fixed code rules check the words: no greeting, no engagement bait, nothing that touches logos or legal notices, numbers and names must come from the video. Muse Spark 1.3 reads the opening back against the video. Muse Image draws three frames and SAM 3.1 looks in the first one for what the title promises. If anything fails, Spark Open gets the exact list of failures and corrects, up to twice. What does not pass never reaches the person. Retina also gives a modest reference for how many of every 100 people who get past the first 3 seconds will still be watching at 15 seconds, learned from 23,710 retention curves, and says how far off it usually is (about 11 people out of 100). It warns when the video does not deliver what the title promises, and checks what is seen and said against 18 of Facebook's public rules. Measured, not promised: every pass/fail rule was written before its run, and sealed in a commit from 29-Oct on. The first test failed (10 of 18) and we report it. The fifth test measured the published version on 50 new videos and met all nine of its conditions: 43 new openings passed every automatic check, SAM rejected 10 of 10 trap frames of another kind of object, and a model judge found 42 of 43 faithful to the video. We have not measured whether the new opening keeps more people watching, and we say so. Try it in one minute with no key: PYTHONPATH=src python -m retina.demo. With your own Meta key, a review costs about USD 0.10 and a fix USD 0.05 to 0.25. Built during the event by Jossafat Ruvalcaba; Marco Zuno represents Badabun. Badabun's archive of retention curves existed before the event and stays private. Everything in the repository was built from 26-Oct-2026 on, with one small overlap the README declares.