Skip to content
HireSalt
Research & Benchmarks

Does every HireSalt score come with a reason?

The HireSalt team 5 min read

In both counts we have made, yes. In 237 test screenings from our 7 October 2026 benchmark run, all 9,314 criterion scores came with a written reason. A second count on 16 October 2026, on the model we use in production today, found the same: 1,972 of 1,972.

That is a narrow result. It shows a reason was there each time, not that the reason is right. The same test’s check on that second question missed the bar we set for it.

What a reason looks like

HireSalt scores each resume against criteria grouped into categories, and each sub-criterion gets a score and a note of one or two sentences. One note from the test, on required skills for a data engineering job: “Resume cites SQL, ETL, data pipelines, and modeling, but not Snowflake, dbt, Airflow, Python, Git, or command-line use.” The test resumes are synthetic or benchmark files, not real applicants.

What we counted

Before counting, we wrote down the rule: the coverage claim passes if at least 95% of sub-criterion scores carry a non-empty note. We took the 237 stored scorings from the 7 October run (40 job descriptions, 225 resume files, one scoring per resume and job pair) and counted. No new model calls were made.

  • 9,314 of 9,314 sub-criterion scores had a note. The bar was 95%, so it passed.
  • 2,302 of 2,302 category scores also had a note.
  • The shortest note was 5 words. None was a blank placeholder.

The run notes for 7 October name one model. One summary file from that run names a different model, and we have not settled which is right. The count does not depend on it, because it only asks whether a note exists.

The 1,972 figure comes from the 48 successful scorings in our PDF-versus-Word test on 16 October, run on our production model. One attempt errored and produced no scores. We counted this after the fact, not under a pre-set rule.

Top panel, coverage: in the 7 October 2026 run, 9,314 of 9,314 sub-criterion scores across 237 scorings had a written note; in the 16 October 2026 check on our production model, 1,972 of 1,972 across 48 scorings did. Bottom panel, a separate measure on the 7 October notes only: 7.6% repeat a four-word phrase from the resume against a pre-set bar of 50% (chance rate 0.4%), and 49.3% pass a lenient rule against a bar of 70% (chance rate 10.4%). Both bars were missed. Criterion scores that came with a written note 7 Oct 2026 run, 237 scorings 9,314 of 9,314 (100%) 16 Oct 2026 check, 48 scorings 1,972 of 1,972 (100%) Share of those notes that trace to resume wording 7 Oct run only, 9,314 notes. A separate measure. Repeats a 4-word phrase from the resume 7.6% (chance 0.4%) pre-set bar 50% Lenient: shares a word pair or a number 49.3% (chance 10.4%) pre-set bar 70% 0% 25% 50% 75% 100%
Top: criterion scores with a written note out of all criterion scores, in two runs (9,314 scores from 237 scorings; 1,972 scores from 48 scorings). Bottom, a separate measure on the 9,314 notes from 7 October: percent that repeat resume wording, with the pre-set bar dashed.

What it does not show

The same test asked a second question, also with a bar set in advance: does the note trace to wording that is actually in the resume? We checked all 9,314 notes against the resume text.

  • 7.6% (709 notes) contain a run of four or more words copied from the resume. The bar was 50%.
  • 49.3% (4,596) pass a looser rule: a shared two-word phrase, or a number that is in the resume but not the job description. The fallback bar was 70%.

Both missed, so the test as a whole is marked not passed. Only the coverage result passed. As a chance check, we tested the same notes against a different resume for the same job: 0.4% and 10.4% matched. The real matches are well above chance, but well below what we asked for.

Part of the gap has a plain cause: 67.5% of notes describe something missing, such as “No professional certifications or credentials are listed.” A note about an absence has nothing to quote. Other notes summarise in their own words. Our test cannot tell a faithful summary from an invented one, so we do not claim the notes are accurate, and we do not say they quote or cite the resume. Read each note as the model’s summary of the resume, not a line taken from it.

Other limits: this is one model’s output on synthetic and benchmark resumes, one scoring per pair. No human panel checked the notes. Nothing here measures who turned out to be a good hire.

What you can do with a note

Read it against the resume. If a note names a skill, a title or a number, find it. If you cannot, give that sub-score less trust and ask about it in the interview.

Disagree in writing. You can add your own note to a candidate, shared with your team or kept private. The score is decision support; a person decides.

Re-weight. If a criterion matters more or less than you set, preview a different weighting on candidates already scored and watch the ranking reorder, computed from the stored sub-scores. Each slider moves within ten points of its baseline, and making a change permanent means scoring again. The preview reorders candidates; it does not change any note. See what happens if you weight skills higher.

All results are on the benchmark results hub. See also how consistent is AI resume scoring, can keyword stuffing game HireSalt scores and how fast HireSalt scores resumes.

Keep reading