FeedbackCollector
Stores user comments like “this is not a longship” with the prompt, output file, and failure reason.
Rev4.8
Feedback loop, not blind waiting.
Running agents longer only helps if failures are saved, analyzed, converted into rules, and tested against future outputs. Rev4.8 adds a local feedback/back-propagation-style loop for ANNplot prompt quality.
Build rev4.8-1782171939This is not neural-network training. It is practical rule/prompt/template back-propagation: bad outputs become rejection rules and better generation constraints.
FailureAnalyzer
Turns failure feedback into structured causes: wrong subject, wrong silhouette, random geometry, missing critical features, bad composition.
RuleBackpropAgent
Updates learned subject/material/composition rules so the same mistake gets rejected next time.
VariantMutator
Generates multiple revised candidates using changed rules instead of repeating the same bad geometry.
RegressionJudge
Runs test prompts such as goat, solar system, periodic table, sailboat, and longship to prevent old failures returning.
PromotionGate
Only promotes SVG/HPGL outputs that pass intent, line density, HPGL bounds, and user-feedback checks.
Core rule
If the prompt says longship, the result must include longship signals: low hull, oars, dragon/prow/stern character, and no generic tall-ship rigging. If those are missing, the output fails no matter how clean the lines are.