Why we don't score resumes

Luminid exists to give people a fairer chance at work.

Over seven years of recruiting at companies ranging from Amazon and Grupo Bimbo to firms in the legal services space, I watched the same pattern play out hundreds of times. Candidates with the right skills filtered out by the wrong proxies. Resumes that told one story about a person's capability when the actual capability told a different story. Decisions made in seconds based on signals that did not predict job performance.

I started Luminid because the gap between what hiring software measures and what predicts on-the-job success is one of the most well-documented findings in industrial-organizational psychology, and few products are built around it.

The patterns

In legal services hiring, I would see paralegals with eight years of substantive case management experience filtered out of senior paralegal openings because they did not have a bachelor's degree. Their actual case work was more sophisticated than the work the role required. But the filter was not "can she do the job." The filter was "did she follow the credentialing path we expect." She never got the interview, so the question of whether she could do the work never got asked.

A different pattern. Paralegals, often women, returning after 2 to 4 years of caregiving leave, often for childcare or eldercare. Their skills were not stale. Their case management knowledge had not degraded. But the gap on their resume triggered screening signals that coded them as "out of practice," and the same person who would have been hired without the gap got passed over with it. They lost two to three years of seniority and pay band in the process of returning. Not because their capability changed. Because the resume told a story that the system interpreted as risk.

A third pattern. Career switchers from adjacent fields. The legal assistant who had done four years of contract review work in a corporate compliance department. The court clerk with deep procedural knowledge. The compliance specialist who had drafted hundreds of vendor agreements. All of them had substantively more relevant skill for certain paralegal openings than candidates coming from purely litigation discovery work. But the title-matching screen caught "paralegal" and missed "legal assistant" or "compliance specialist." The actual work was invisible. The title was what got evaluated.

These were not unusual cases. They were the daily texture of recruiting in that space. I would see five of them a week, sometimes ten. And the legal services space was not unique. The same patterns repeated at the consumer products companies I worked with. They repeated at the e-commerce companies. Different industries, same screening logic, same predictable filtering of people whose capability was real but whose paper trail was wrong.

The research

The thing that made this maddening, not just frustrating, is that the research has been clear for decades.

The most cited finding in industrial-organizational psychology comes from Schmidt and Hunter's meta-analysis of nearly a century of personnel selection research. They measured how well different hiring methods predict actual on-the-job performance. The findings were stable across industries, roles, and time periods.

Work-sample tests, where candidates do a piece of the actual job and the recruiter evaluates the output, predicted job performance at a correlation of 0.54. That is a strong signal.

Years of job experience predicted performance at 0.18. Mediocre.

Education level predicted performance at 0.10. Weak.

Resume screening, the thing that almost every hiring decision starts with, predicted job performance at 0.03. Statistically indistinguishable from random.

This is the dominant finding in the field. It has been replicated in subsequent meta-analyses, including more recent updates. The HR field knows this. The research is taught in graduate programs. And yet the entire applicant tracking software industry, a multi-billion-dollar category, is built around the signal that predicts performance the worst.

When I started Luminid, the architectural question was straightforward. What would hiring software look like if it built around the signal that actually works?

The choice

Luminid replaces resume screening with role-specific work simulations. A candidate applying for a paralegal role does not submit a resume and wait for someone to skim it. They complete a work sample tied to the actual work of that role. Contract review. Case analysis. Document drafting. Whatever the role actually requires. The simulation produces evidence. Recruiters evaluate the evidence.

The score that comes out of a Luminid simulation is called Role Match. It reflects the candidate's demonstrated capability for that specific role, weighted by the skills the recruiter has identified as required or preferred. It is not a verdict on the candidate as a person. It is a measure of fit between this person and this role.

Experience and education appear on candidate profiles. Recruiters see them. They can use them as context. They do not drive the Role Match score. The score is built from what the candidate has shown they can do, not from what their resume claims they have done. The full formula and bucket weights are documented at /methodology.

AI assists with scoring the simulations. It does not screen resumes, infer career trajectories, or evaluate candidates as people. Every simulation has a construct map that defines what the simulation measures, and the AI scores responses against that map with cited evidence from the candidate's work.

Recruiters see the evidence. They make the hiring decision.

We call this category verified hiring. Hiring decisions based on capability candidates have demonstrated, not credentials they have claimed. The category did not have a name before Luminid. We created one because the existing labels (ATS, assessment platform, talent acquisition software) describe how the work gets done, not what the work is. Verified hiring describes what the work is.

We publish how this works. The methodology is at /methodology, with the formula, the constraints on AI use, the bias-handling architecture, the calibration approach, and the known limitations. Hiring decisions have real consequences for people. The software making those decisions should be transparent about how it works.

The methodology will evolve. The version history at the bottom of the methodology page tracks every material change. When the formula adjusts, when the AI use changes, when calibration data shifts what we know, you can see it.

The honest stage

Luminid is in early development. We publish our methodology before we have a long customer track record because methodology should be public from day one, not after we have accumulated reasons to defend specific outcomes.

Calibration data lets us validate that our simulation scores actually predict on-the-job performance. It accumulates over time as customers use the platform and provide post-hire feedback. Early simulations carry less validated confidence than mature ones. We surface this in the product so recruiters can see how much weight to give different scores. The methodology improves as the data accumulates.

This is the work. Not a finished product, but a deliberate one.

The close

The paralegals I saw filtered out for the wrong reasons in legal services hiring are not the only people the current process fails. The patterns repeat across industries, roles, and seniority levels. People with capability filtered out by proxies. People with potential blocked by the way the system reads paper.

Luminid exists to make those people visible. Not by lowering standards. By measuring the right thing.

The methodology that does this is at /methodology. The patterns that drove it are above. The work continues.

---

The legal services patterns described here are composites from years of recruiting. Specific candidate details have been generalized to protect privacy. The patterns are real.