Two practical routes were run over the same English passages. A simple comparison-word search is precise but misses most relevant expressions. The two-model workflow searches more broadly, then orders what people inspect. Both analyses were added by this project; neither was part of the original #ReframeCovid dataset release.
| How candidates are found | candidates | worth keeping | strong | known-good items found |
|---|---|---|---|---|
| Look for comparison words "feels like", "as if", "like a" — a regular expression | 286 | 24.5% | 31 | 3 of 17 |
| Broad proposal, then second-model verification one model proposes; another checks each phrase | 1,665 | 24.6% | 123 | 17 of 17 |
Candidate recovery only: 17 of 17 means that every known-good item produced a candidate somewhere in the pool. Four Menu entries named in the Menu-likeness prompt were excluded from all ranking-based reference results, so this does not show that all 17 ranked highly.
Searching for comparison words is remarkably precise: judged one at a time, an expression found this way is as likely to be worth keeping as one from the full search — 24.5% against 24.6%. The problem is reach. Only 12% of English passages contain such a marker, and under 4% of Danish ones, because people describing illness rarely announce the comparison. It finds three of the seventeen known-good items and cannot reach the rest at any effort.
One practical note: the comparison-word route needs no model, no GPU and no paid service. A team with no computing resources at all can run it and obtain 70 candidates that pass every check, 31 of them strong — provided everyone understands it will miss four fifths of the material.