LLM-assigned feature weights: multiplying by expert tiers cannot fix the order; clamp the importance with floors and caps
Setup: an LLM rates each factor 1-10, the rating maps to points (1.5^(imp-1)), and domain-expert rules multiply the points (e.g. x3 for competition, x0.35 for hazards).
Failure: the LLM rated a niche factor 10 (covered school teams, only a fraction of patients), the factor that should decide the site 8, and flood risk 8. Even with tiers applied, the niche factor tied for #1 and a factor the expert considers meaningful (nearby PT clinics, rated 2) ended up below flood. Because the scale is convex, a 2-point rating gap outweighs most multipliers.
Fix: give each expert rule an importance floor and/or cap that is applied before the convex mapping: - the deciding factor gets floor 10; - niche demand sources get cap 6-7; - hazards get cap 3; - secondary but real factors get floor 5. When the clamp changes a rating, replace the LLM explanation with the rule text so the UI does not show an explanation that contradicts the weight.
Tip: replay the cached LLM advice through the weight function offline (pull it from the LLM response cache) to tune floors and caps before paying for a live run.