RESOURCES · EXPERTINI ATS

AI & Automation — Avoiding Hallucination with CMS

Why asking a language model for a score is structurally different from asking it to read.

3 min read · Updated July 2026 · Expertini Editorial

Hallucination — a language model confidently producing an answer with no basis in its actual input — is often described as a bug that better prompting fixes. It's better understood as an intrinsic property of how these models generate text: a next-token predictor asked a question will produce a fluent, confident-sounding answer whether or not the underlying facts support it, because fluency and factual grounding are two different things the model was never specifically trained to keep aligned. This page explains why that matters specifically for hiring scores, and the architectural choice that makes hallucinated scores structurally impossible in Expertini's implementation rather than merely less likely.

▶ Watch the 30-second overview — no sign-up needed

0final scores generated directly by the AI model

1schema the model's output must validate against or be rejected

100%reproducible — deterministic math, run outside the model

2retry attempts before an extraction failure is surfaced honestly

01The specific failure mode in hiring

Ask an ungoverned language model to "rate this candidate 1 to 100 against this job" and it will produce a number — confidently, with a plausible-sounding justification attached. Ask it again, with the identical CV and identical job description, and there's no guarantee you get the same number back; the model's output depends on sampling randomness in how it generates text, not solely on the input content. For a decision with real consequences for a real person, an unreproducible number with a confident-sounding explanation is a specific, serious problem — not a minor accuracy gap that improves with a better model version.

02Bounded extraction: constraining what the model is allowed to do

Expertini's extraction step gives the model exactly one task: read an anonymised CV against a job description and return a structured list of competency dimensions, each with a JRIS weight, a CSS evidence score, and a short citation of the evidence found — validated against a fixed schema before it's accepted. Output that doesn't match the schema is rejected and the model is asked to retry, meaning it is structurally unable to smuggle a final score, a ranking, or a hiring recommendation into its output — those fields don't exist in the schema it's constrained to produce.

03Deterministic aggregation: computed outside the model entirely

The actual score — CMS = Σ(CSS×JRIS)/Σ(JRIS), detailed in full on the dedicated CMS page — is arithmetic, run as ordinary code completely separate from the model. The model has already finished its job by the time this calculation happens; it has no ability to influence, bias, or vary the arithmetic, because the arithmetic isn't part of what it generates. This is what makes identical inputs produce identical outputs on every run, indefinitely — not a property of a particularly well-tuned model, but a property of separating extraction from computation entirely.

04Where this leaves genuine limitations

The extraction step can still misjudge or miss evidence in a CV — a genuine capability limitation of the underlying model, not a hallucination in the technical sense, since a schema-validated citation is either grounded in real CV text or the output is rejected. This is a real limitation worth stating plainly: an extraction that undercounts genuine evidence produces a lower CMS score than the evidence actually supports, which is why the audit report showing every dimension's cited rationale exists — so a recruiter can independently check the model's reading rather than trust it blindly.

Frequently asked questions

What is AI hallucination, technically?⌄

A language model generating fluent, confident output that isn't grounded in its actual input — an intrinsic property of how these models predict text, not a fixable bug in any one implementation.

How does CMS make hallucinated scores impossible rather than just unlikely?⌄

The AI model is constrained to a fixed extraction schema that has no field for a final score, ranking, or recommendation — those simply cannot appear in its output. The score itself is computed by separate deterministic code the model never touches.

Can the AI still make mistakes?⌄

Yes — it can misjudge or miss genuine evidence in a CV, which is a real capability limitation. This differs from hallucination because every citation is checked against a schema and grounded in actual CV text; the audit report lets a recruiter independently verify the model's reading.

Does this approach eliminate the need to review AI-assisted scores?⌄

No — reviewing the dimension-level evidence and rationale on the audit report remains part of a responsible hiring process, precisely because the extraction step, while constrained, is not infallible.

At a glance

AI performs extraction only, never scoring
Schema-validated model output, rejected and retried if malformed
Score computed by deterministic code outside the model
Reproducible across every re-run, indefinitely
Every citation grounded in real CV text, not invented
Published methodology and stated limitations

See ai & automation — avoiding hallucination with cms on your own hiring.

Bring a real job description to a 30-minute demo — free trial included.

Book a demo