Make the model abstain

Every pipeline I've built that puts an LLM between messy input and a structured decision has had the same early failure: the model fills in fields it shouldn't. Given a thin candidate profile (a job title and two lines of prose), it will happily produce a confident list of "skills" that were never stated. Given a sparse lead, it invents a company size. The output looks clean. It is wrong in a way that is invisible downstream, because nothing about a fabricated field says "fabricated."

The fix is boring and it works: make abstention a first-class output, and treat it as normal. No field beats a guessed field. The extraction step returns null for anything it can't ground in the input, and every consumer of that step is written to expect null as the common case, not the error case.

Two things follow from that:

• Confidence has to be calibrated against abstention, not against being right. The question isn't "how sure are you this is Python." It's "are you sure enough to say anything at all."
• Downstream ranking has to degrade gracefully. A profile with three grounded skills and eight nulls should produce a weaker, clearly-marked match, not one that silently competes with fully-populated profiles.

The version of this that scales is deterministic rules for the high-frequency cases and the model only for the genuinely ambiguous tail, which is allowed to come back empty. You lose some recall. You stop shipping confident nonsense. On data entered by people who aren't thinking about machine readability, that's the right trade.