J-WORKS blames jworks_fit_score mismatch for off-topic AI picks

The article walks through the candidate-selection pipeline in order, from search to score to approval, and reports that the skew was not a failure of AI judgment alone. In the search stage, seasonal keywords forcibly overrode the choice of search keywords, so the candidate pool itself leaned toward seasonal general consumer goods. That matters because even a later approval step judging brand relevance would have received mostly off-theme inputs.

At the scoring stage, the name jworks_fit_score suggested a fit score, but the score actually evaluated demand, reliability and profitability. A high score therefore did not mean the candidate fit J-WORKS's themes. The approval prompt also lacked criteria for judging brand fit, so J-WORKS changed the design to insert a mechanical relevance filter before the AI compares candidates, rather than leaving the relevance call to the approval stage.

J-WORKS also re-evaluated existing data under the new logic and adjusted its operating state, disabling the related periodic candidate-judgment runs so the old path did not keep running. The stakes here hinge on whether the front-end filter can keep candidates on-theme on its own, since the article argues that fixing only one stage leaves the same problem at the others.