Total score
The system scores candidate evidence for each requirement, keeps the strongest support, and assigns a status from that best score.
total = sum(component_score * component_weight)Methodology
How the evidence score is calculated.
Each requirement from the AdventHealth Product Engineer posting is matched against evidence-backed claims, skills, projects, education, language, and credential records. The score is a ranking aid for interview review, not an automatic credential assertion. The numeric weights are local policy for this run; the scoring dimensions are tied to the research basis below.
43
Requirement rows scored
72%
Strong threshold
50%
Supported threshold
244
Reviewable claims
Formula
Weighted evidence fit
Status rules
strong
Score is at least 72%.
supported
Score is at least 50% and below strong.
weak
Some evidence exists, but the best score is below 50%.
blocked
The graph should not claim it without verified evidence, usually credentials.
Weights
What moves the score
These values are read from the generated evidence-map artifact for this run.
Requirement and evidence overlap
28%Measures how directly the selected evidence text matches the language of the requirement.
Deterministic token overlap in this evidence map; BM25 is implemented in the broader ranking layer for lexical retrieval.Evidence strength
26%Rewards evidence that is source-backed and marked usable in the graph instead of loose wording.
Local provenance scoring over graph evidence levels.Source backing
16%Rewards claims tied to a concrete experience, project, education, language, or credential record.
Local graph-lineage rule inspired by evidence-first fact verification.Domain match
10%Rewards matched domain concepts when the requirement calls for a specific environment or specialty.
Deterministic taxonomy and normalized-skill matching.Inspectable public source
10%Rewards evidence with a public URL such as a project page, GitHub artifact, or public demo.
Local provenance bonus for inspectable public artifacts.Curated claim
6%Rewards evidence that has been promoted from raw facts into a reviewed claim record.
Local claim-review policy; generated or raw facts do not get the same trust as reviewed claims.Deterministic match
4%Rewards direct structured matches such as explicit education, language, or credential facts.
Exact structured-field matching for fields that should not be inferred from work claims.Research Basis
Algorithms and papers behind the rubric
The papers justify the retrieval, similarity, diversity, and evidence-support methods. They do not set the local weights.
Lexical retrieval
Token overlap plus BM25-style lexical retrieval in the ranking layer
The current evidence map exposes a transparent token-overlap component. The broader workflow also computes BM25 scores for candidate ranking debug output.
Robertson and Zaragoza, The Probabilistic Relevance Framework: BM25 and Beyond, 2009Semantic similarity
Sentence embeddings with cosine similarity
The code has an embedding-provider boundary, but this artifact uses the deterministic fallback rather than a live embedding model.
Reimers and Gurevych, Sentence-BERT, 2019Diversity reranking
Maximum Marginal Relevance
MMR supports the broader optimizer's diversity logic. It is not one of the seven evidence-map weights shown on this page.
Goldstein and Carbonell, Using MMR for Diversity-Based Reranking, 1998Claim support framing
Natural-language inference labels used as a support vocabulary
The current runtime uses deterministic heuristic support labels, not a learned entailment model.
Bowman et al., A large annotated corpus for learning natural language inference, 2015Fact verification
Evidence-linked supported, refuted, and not-enough-information framing
The review keeps requirement fit separate from evidence support, and blocked rows prevent unsupported credential claims.
Thorne et al., FEVER: a Large-scale Dataset for Fact Extraction and VERification, 2018Vector retrieval infrastructure
Nearest-neighbor vector search
This supports the future adapter path for vector indexes. The current AdventHealth artifact does not depend on FAISS or a vector database.
Johnson, Douze, and Jegou, Billion-scale similarity search with GPUs, 2017Interpretation
How to read the score
Strong and supported rows are interview anchors: they identify claims worth discussing first.
Weak rows are honest edges: they may be true, but the graph needs cleaner wording or better lineage.
Blocked rows are intentionally conservative and should not be presented as claims.
Limits
What the score does not prove
It does not replace a reference check, credential verification, or code review.
It does not expose private corporate files; sensitive evidence is summarized.
It ranks fit against this posting, so another posting can produce different scores.