How recommendations are ranked

Signals used

PavedIT 1.4.3 is the Intelligence & Trajectory layer inside PavedIT. It combines evidence-weighted skill similarity, transferable-skill adjacency, role and seniority progression, career trajectory, compensation alignment, work style, and industry fit. Résumé-backed skills require an exact quote selected by the candidate before they enter matching. Exact occupation titles and skill labels are checked against the active, attributed O*NET release, and each result identifies the source version when a label resolves.

Signals excluded

Protected personal attributes are not accepted or scored. An automated test verifies that adding age, gender, or ethnicity fields cannot change a ranking.

What the numbers mean

The match score is a directional, deterministic ranking aid produced by unvalidated heuristic parameters — never a probability of hiring. When a job or candidate record lacks a required input, that dimension is marked unavailable and excluded; PavedIT renormalizes only the observed dimensions instead of assigning neutral evidence. A job title is never substituted for missing source-provided skills. Seniority is inferred from a title only when the title explicitly names a recognized level; unrecognized titles receive no default level, and unsupported trajectory or growth signals display as unavailable. A target role or target seniority is treated only as an aspiration and never as evidence of the candidate’s current level. Signal coverage measures whether the inputs needed for each weighted ranking component were present; it does not judge their truth and is not statistical confidence or interview likelihood. Growth scores are directional comparisons, not earnings promises. Every result discloses its heuristic version, taxonomy provenance, fallback state, and missing signals.

User control and exploration

The three available curation preferences—leadership, compensation, and remote work—apply small, versioned ordering bonuses after modeled fit; they do not enter the matching model or change the displayed fit score. Controls are shown only when a corresponding live job field exists in the product contract. One in four visible curated positions may be reserved for an evidence-backed higher-growth adjacent role, reducing over-narrow personalization. If no candidate role has an observed growth score, PavedIT withholds the exploration slot instead of promoting an unsupported opportunity.

How quality evidence is collected

When a signed-in candidate labels a recommendation relevant or not relevant, PavedIT links that label to the exact server-authored ranking version, position, score, signal coverage, and taxonomy release. Ranking snapshots contain no résumé text or protected attributes and expire after 90 days. These self-selected labels can estimate labeled usefulness with a 95% Wilson interval only after minimum candidate, judgment, time-window, and concentration gates. They cannot prove nDCG, recall, hiring likelihood, or causal career benefit.

Recommendation-outcome research is separately off by default. When enabled, a future candidate-reported application-stage change may retain the prior model version, score, rank, signal coverage, and ranking timestamp for up to 365 days. Outcome analysis requires a complete exposure cohort, one frozen model version, a fixed follow-up horizon, adequate coverage across five score bands, and candidate-concentration limits. Right-censored or low-coverage evidence fails closed. Even a complete report is labeled observational only—never accuracy, probability calibration, causal benefit, or fairness evidence—because consent and stage reporting are self-selected.

A scientific release claim requires a separate frozen holdout study: complete candidate–job lists, at least two independent blinded assessments per pair, measured assessor agreement, time splits, confidence intervals, and pre-specified segment checks. Every model must beat the strongest of three frozen comparators—recency/popularity, exact skill overlap, and semantic retrieval—not an average or an easy hand-picked baseline. The lower bound of a paired query-level 95% bootstrap interval must clear the registered lift threshold against every comparator; a positive average with uncertain lift does not pass. Before collection, an externally timestamped plan must freeze the model, comparators, taxonomy, sampling frames, thresholds, and temporal boundary. Before analysis, a hashed freeze manifest binds the resulting judgment file. Local timestamps alone do not prove pre-registration. Changed artifacts, missing double-judgment coverage, or agreement below the registered gate display insufficient evidence before model metrics can pass; sparse samples are never converted into a passing grade.

Every evaluated query must reconcile exactly to its pre-declared frozen candidate-pool size and meet the registered minimum depth. Missing or extra jobs invalidate the study; a complete but shallow pool displays insufficient evidence. PavedIT never evaluates only a favorable top-ranked shortlist.

Fairness segments are pre-specified, consented audit cohorts—not ranking inputs. Raw protected-attribute values are excluded, but pseudonymous cohort assignments remain sensitive, access-restricted audit data. Each sufficiently sized cohort receives its own interval. The release ceiling applies to the 95% upper bound of the worst segment gap, so an apparently small disparity with wide uncertainty cannot pass.

Limitations

PavedIT has not yet been calibrated against independently judged candidate relevance outcomes. Its weights, thresholds, fallback synonyms, and adjacency values are expert choices, not validated estimates. Candidate confirmation proves authorship of a claim, not employer verification or proficiency. Unmapped labels use a disclosed lexical fallback. Sparse, non-standard, non-English, frontline, and career-break profiles remain known blind spots. Before general availability it requires segmented relevance evaluation, fairness monitoring, and drift detection.