Most AI hiring tools hand you a number. This piece shows the machinery behind our Fit Score end to end, with a worked example, because a score you can't reconstruct is a score you can't trust. (That principle has its own manifesto; this is the engineering-room tour.)
A Fit Score is a 0–100% score that blends a candidate's AI CV evaluation and AI interview results into a single ranking number, weighted the way the hiring team chooses, and traceable at every step back to written, evidence-citing justifications.
The pipeline, in four stages
Stage 1: evidence becomes criterion scores. Two engines evaluate the candidate independently against the same role rubric: one reads the CV, one conducts and scores the async interview. Each produces a 0–10 score per criterion, and every score carries a written justification citing what was found. Nothing downstream exists except as an aggregation of these.
Stage 2: weight tiers shape each engine's rollup. Rubric criteria carry tiers (Must Have, Very Important, Important, Good to Have), so a 4/10 on a Must-Have drags the rollup far more than a 4/10 on a nice-to-have. Each engine's criterion scores combine through those tiers into a weighted 0–10: a CV score and an interview score.
Stage 3: the project's blend combines the engines. Every hiring project sets how much the CV and the interview each count: 50/50 by default, but a role where the interview is the real signal might run 30/70. The system normalizes whatever you set to a total of 100%.
Stage 4: the blend becomes the Fit Score. The weighted combination lands on 0–100%, displayed with its full decomposition. One deliberate consequence: because the blend is project-scoped, the same candidate assessment can produce different Fit Scores on different roles. Fit is relative to the role, and the score is honest about that.
The worked example
Meet Sarah Collins (a fictional candidate from our demo data), applying to an agentic-AI developer role with a six-criterion rubric.
| Criterion | Tier | Interview score |
|---|---|---|
| Proficiency in AI development | Must Have | 10/10 |
| Programming skills | Must Have | 10/10 |
| Understanding of ML algorithms | Very Important | 9/10 |
| Experience with AI ethics & safety | Important | 8/10 |
| Research & continuous learning | Important | 9/10 |
| Collaboration & communication | Good to Have | 7/10 |
Notice the shape: perfect on both Must-Haves, softer on a Good-to-Have. The tier weighting rolls this up to an interview score of 9.2/10: the 7 costs her little because the rubric says collaboration is a tiebreaker for this role, not a pillar. A rubric with Collaboration at Must-Have would punish the same answers hard. Same evidence, different role definition, different score. That's the rubric doing its job, visibly.
Her CV evaluation, run against the same criteria by the CV engine, lands at 9.2/10 with its own justifications. The project blends CV and interview 50/50: each contributes 46 of its possible 50 points, and Sarah's Fit Score is 92%, ranked first in the pipeline, with every layer inspectable underneath: click the 92%, see the components; click a component, see the six criteria; click a criterion, read the justification; click through, and the transcript itself is there.
The rule that makes the number trustworthy
One scenario reveals more about a scoring system than any accuracy claim: what happens when evidence is missing? Suppose a candidate's CV never touches one criterion. Three designs are possible: score it 0 (punishes the candidate for the document, not the skill), guess from context (fabrication with confidence), or say "Cannot evaluate": score nothing, exclude the criterion from the weighted math, and show the gap in the report.
Rubrily does the third, everywhere, without exception. A missing answer is treated as missing, not as bad. This is why we say the product never fabricates a score: every number in a Fit Score exists because the AI found evidence and wrote down what it was.
How is a Fit Score different from a "match score"?
A match score is typically a single opaque percentage from unstated criteria; you can't see its parts, so you can't contest them. A Fit Score is a stated aggregation: your criteria, your tiers, your blend, per-criterion justifications, and full decomposition on demand. The difference isn't the number; it's whether the number can be interrogated.
What this means for how you rank
Because the Fit Score is explainable, ranking by it is a starting posture, not an abdication: sort the pipeline by Fit, then read the justifications behind any decision that matters. Two candidates at 84% and 82% are a coin toss until you read why. One may be weak on a criterion you privately know matters more than the rubric admits (which is a signal to fix the rubric, not to trust vibes). The score compresses the evidence; the report preserves it. Use both.
FAQ
What is a Fit Score? A Fit Score is a single 0–100% score blending a candidate's CV evaluation and AI-interview results, weighted per project. It's fully traceable: components, per-criterion 0–10 scores, written justifications, and the underlying transcript and CV are all inspectable behind the number.
Can I change how the Fit Score is weighted? Yes, per project: each component (CV eval, interview) has a weight you control, normalized to 100%. Criterion-level influence comes from the rubric's four weight tiers. Because weighting is project-scoped, the same assessment can legitimately produce different Fit Scores on different roles.
What are the exact tier multipliers? The tier mechanics (Must Have > Very Important > Important > Good to Have) are as described; the precise internal constants are implementation detail we don't publish. What we guarantee is the property that matters: every rollup decomposes back into visible per-criterion scores and justifications, so any candidate's number can be reconstructed and audited from the report.
Does a low Fit Score reject a candidate automatically? No. Nothing in Rubrily auto-rejects anyone. Scores rank the pipeline; humans read reports, add reviews, and decide. The pipeline makes the ranked shortlist actionable; it doesn't act on your behalf.
See a Fit Score decompose in your own pipeline: define a rubric, screen your applicants, click through every number. Join the waitlist →