I like the idea, though I think it would be difficult to apply given that the wild/captive dichotomy is fairly fundamental to the way iNat classifies observations as verifiable (and eligible to become RG) vs. non-verifiable (e.g. casual).
Maybe if the standing feature request to not lump non-wild observations with defective observations were ever implemented, it would become possible to refine the wild/non-wild criteria to allow for more nuance and better address some of the perennial issues (free-ranging non-wild individuals on one side of this coin; feral non-established individuals on the other). It would still be necessary to find some solution to determine which of these “neither wild nor non-wild” observations would be eligible for sharing with GBIF.