When I check a draft against its source, I ask a lot of questions at the same time. Has the scope drifted? Has the emphasis shifted? Did a qualifier disappear, or appear? Did the meaning change? What was added or missed?
It's a sentence-by-sentence, section-by-section job. And it's painstaking.
The hard part is fidelity
Drafting from source mixes rules and judgement. You pick the key points. You apply a defined template. You decide what's essential and what's noise.
It makes little difference whether a human or an LLM writes the draft. AI wins on speed, grammar, and spelling. The hard part is fidelity. Does every sentence still say what the source says?
Two kinds of drift
No one holds all of that in their head at once. My mental rules engine isn't foolproof. It drops a check here and skips one there. It's sharper on page one than on page twelve. By the third review, I've lost sight of the source and keep refining the draft.
AI has a different problem. It reads the source, then drifts from it anyway. It smooths, generalises, and fills gaps with what sounds right.
Check every claim the same way
That's what FDC is for. It runs the same set of fidelity checks on every claim, every time, against the source and nothing else. Then it sorts each claim into one of ten categories, from rephrased to invented to missing.
Compare meaning, not words
It isn't a diff. A diff compares words, and a good draft never matches its source word for word. FDC compares meaning. That's where it gets subtle.
The source says 70%. The draft says "the majority." True, but no longer what the source says. Or the source includes a warning the draft never mentions. No diff shows you what isn't there.
FDC doesn't sample. If a claim is true but the source doesn't say it, FDC flags it anyway. True isn't the same as sourced.
Run it again at every draft
FDC doesn't create content. It catches drift and shows me where to look. The judgement stays mine.
Every run produces a report, so I can show exactly what changed between drafts.
It's also iterative. I rerun FDC at each major draft, because every round of edits can drift again. So by the third review, the source is back in view.
As I said, it's painstaking. Catch my drift?
Curious how I built the fidelity checks? Get in touch and I'll walk you through it.