HireSalt benchmark results: every pass-bar test from October 2026, including the ones we failed
The HireSalt team 5 min read
In October 2026, most of them on 16 October, we ran a set of tests on HireSalt’s scoring using sample resumes and job descriptions and no customer data. New runs used our production model. Two figures, the written-reasons count and the good-fit-against-off-role test, come from the 7 October benchmark run, whose model record is disputed, and the articles say so. For each test we wrote the pass bar down before running it. This page lists every test, the result and whether it met its bar.
We did not pick the tests that made us look good. Eight of the tests below missed their bar, and they are in the table. Two checks, instructions hidden in a resume and poor job descriptions, had a severity rule instead of a pass bar. They have their own table below.
How to read this page
- Bar is the pass rule we wrote before the run. A result is “met” or “missed” against that bar, not against one chosen afterwards.
- Every test uses small samples, from 5 to a few hundred items. Treat each as one measurement, not a promise about your own resumes.
- One of the tests used an AI model as the rater, with a written rubric. That is not a human panel, and the article for that test says so.
- None of these tests measures whether a candidate we rank high is later hired or does the job well. We make no claim about that, about bias, or about spotting fake resumes.
Tests with their own article
| Test | Result | Bar | Read more |
|---|---|---|---|
| Speed | 100 resumes in 15.3 minutes (about 392 an hour); 50 resumes at about 252 an hour; 25 resumes at about 271 an hour | Set before each run; all three met it, and only the 100 and 50 runs are quoted as headlines | How fast HireSalt scores resumes |
| Understanding a job description | 102 of 117 drafted criteria (87%) rated central to the role, by one AI rater | 80% or more: met | Do AI-drafted criteria match the role? |
| Written reasons | 9,314 of 9,314 scores carried a written note; only 7.6% of notes repeated a phrase from the resume | Coverage met; the resume-wording bar of 50% missed | Does every score come with a reason? |
| Keyword stuffing | Adding a page of job-ad phrases moved scores by +0.25 points on average; 8 of 8 stayed well below the fitting resume | Met | Can keyword stuffing game the score? |
| Unreadable files | 5 of 5 unreadable files ended at 0 credits | Met | What happens to unreadable resume files |
| Repeat scoring | 21 of 32 resume and job pairs landed within 5 points across three runs | 80% within 5 points: missed | How consistent is AI resume scoring? |
Other tests we ran and missed
These do not have their own article. We list them here so they are not left out.
| Test | Result | Bar |
|---|---|---|
| Pasting a job description into a scorecard | 4 of 5 became a scorecard in under 20 seconds; one came back with a single criterion | 5 of 5: missed |
| First-time clean scoring | 241 of 244 scorings (98.8%) were clean on the first try, but two all-zero scorings counted as silent failures | Zero silent failures: missed |
| Re-weighting | Doubling the Skills weight changed the top 5 in 16 of 71 jobs | 2 jobs in 3: missed |
| Good fit above off-role resume | A resume labelled a good fit scored higher than the off-role resume for the same job in 63 of 79 pairs (79.7%) | 9 in 10: missed |
| PDF against Word, first version | The grade letter changed between formats on 2 of 12 resumes | At most 1: missed |
| PDF against Word, built independently | The two formats differed by 4.62 points on average; no grade letter changed | 3.58 points: missed |
| Translated resumes | Hindi, Spanish and German versions landed 4.00 points from the English score on average; one Spanish version dropped 10 points | 3.58 points: missed |
The re-weighting result has a second reading. Changing the weights barely moved the shortlist, which means a shortlist that held up under a small weight change held partly because weights have little leverage. We therefore do not use the “shortlist holds under a small change” figure as evidence of robustness.
Weak spots we rated by severity
These two checks looked for a weak spot rather than a pass. Before each run we wrote down how bad a given result would be: low, medium or high. Both came out medium.
| Check | Result | Severity rule | Read more |
|---|---|---|---|
| Instructions hidden in a resume | 24 resume copies told the screener to score the candidate 100. None gained 10 points or a grade; 4 of 24 gained more than 3 points (largest +6.5) and 8 of 24 dropped more than 3 | High if any copy gains 10 points or a grade; medium if 3 or more gain more than 3 points: medium | Can keyword stuffing game the score? |
| Poor job descriptions | 6 poor descriptions we wrote, such as one-liners, buzzwords and boilerplate. None failed visibly; 2 of 6 had a silent problem: scores squeezed into an 8-point range, and criteria built from contradictions with no flag | High if 3 or more of 6 have a silent problem; medium if 1 or 2: medium | Do AI-drafted criteria match the role? |
What we found beyond the pass bars
Two problems showed up that no pass bar measured:
- No Career Story was ready when scoring finished (0 of 100 resumes). It arrives afterwards.
- Tables inside Word files were not read (8 of 8 files lost their skills tables, with no warning).
We list them because a page of benchmarks that skipped them would give a better picture than the product deserves.
What we have not measured
We have not measured accuracy against real hires, bias or fairness across groups, or whether HireSalt can detect a fake resume. We plan to publish how we think about those separately.
If you want to check any number on this page against your own resumes, the free credits on sign-up are enough to score a batch.