Make the next response harder to get wrong.

The source checks five regex lists: scope, validation, rollback, assumptions and uncertainty. It starts at one hundred and subtracts eighteen for each missing list. Five missing gives one hundred minus ninety, ten, while no missing gives one hundred. Bars share five pixels per score point to compare fifty with five hundred pixels. The minimum reachable score is ten, not zero. This is literal text signal scoring, not a model, security scanner or proof that proposed code is safe. Suppose the response contains only test. Validation is matched and the other four lists are missing. The score is one hundred minus seventy two, twenty eight, compared with ten for five missing. Bars use five pixels per point to compare one hundred forty and fifty pixels. The source checks the response only; keywords in the coding task do not raise the score. The fragment test inside contest can also match, and uncertainty may requires a trailing space in one branch. These are regex details, not evidence of actual testing or uncertainty reasoning. Five independent matched or unmatched lists give thirty two subsets but only six score totals, ten, twenty eight, forty six, sixty four, eighty two and one hundred. Bars count thirty two subsets and six totals at sixteen pixels per count. The result stores trimmed task, response, score and the complete missing rule entries. Changing the draft alone leaves the last stored review and downloaded text unchanged until Analyze runs. Reset clears the stored review. No code executes and nothing is submitted to an external reviewer.