← Résumé generator · Finding categories · Evaluator scores vs human labels

Evaluator summaries vs human summary

One paragraph per case and prompt, beside my own summary of the same résumé. Select the evaluator, and whether the paragraph is the evaluator's own overall.evidence or a rollup model's summary of the same review.

Cases run left to right; scroll the grid sideways to reach the rest. Each cell is prose only, with no findings, severities, scores, or verdicts. — means nothing was stored. The top row is my own labeled assessment of the résumé, shown as written; it is not one of the prompt treatments.

About the summaries

The default source, Evaluator's own, is not a rollup: it is the overall.evidence text the evaluator itself wrote in the review, quoted from the stored record. The other two are rollups — a second-stage model read one evaluator review and wrote the summary field of its report, so the same review reads differently under Haiku and Flash-Lite, and both typically run longer than the evaluator's own line. The scores are the evaluator's own in every case and do not move when the source changes; the findings the rollups pulled out of the same reviews are counted on the finding-category grid.

Where a rollup stored no report, the page falls back to the summary in that rollup's raw output, if that output parses, and tags it unvalidated; otherwise the cell shows —. A stored report can still have failed citation validation. This page does not distinguish those, because validation applies to a report's citations rather than to its summary prose; the finding-category grid marks them as having unverified citations. The evaluator's own evidence goes through no such step — it is read straight from the review.

The labeled assessment row is mine, not a prompt treatment and not a model's output: my own reading of each résumé, written in the style of the rollups' summary field so it can sit in the same grid, and shown verbatim. These are assessments of the résumé, not of any review. The tag under each case name records where that résumé came from: organic generator output, one of the manipulated twins, or the retired case.

All 16 historical cases are included, in the order used by the other two grids, including Skills-spam trap (Xcel), temporarily restored to this comparison and marked retired. The four rows are Base (anchored), Mechanism (anchored-mechanism, human-reading passage), Criticism (anchored-pressure, evaluator-bias passage), and Combined (anchored-mixed, the integrated passage). Source reviews remain available in the evaluation playground.