Source-Grounded Verification: Why Every SiteNexis Finding Carries a Confidence Score
Audit findings are only as valuable as the confidence you can place in them. SiteNexis runs a verification pass on every issue, computing source reliability, extraction consistency, and adjusted severity — so you know which findings to act on first and which to investigate further.
An audit tool that reports findings without confidence scores is asking you to trust all findings equally. That is not how evidence works. A missing H1 detected by DOM parsing has near-certain confidence. A contradiction detection result from semantic analysis has meaningful but non-certain confidence. A synthetic entity pattern flagged with partial evidence has lower confidence. SiteNexis runs a verification pass on every issue that assigns confidence scores and adjusts severity based on evidence strength.
Deterministic vs. Probabilistic Findings
Some findings are deterministic: the H1 tag is either present or absent, the canonical URL either matches the page URL or it does not, the schema type is either valid or invalid. These findings carry DETERMINISTIC_CONFIDENCE (1.0 by default) — they are facts derived directly from the DOM, not inferences. Other findings are probabilistic: entity consistency scores derived from semantic comparison, contradiction detection from Claude API analysis, synthetic entity patterns from statistical signals. These carry confidence scores below 1.0 that reflect the uncertainty in the inference.
Source Reliability Scoring
The verification layer scores the reliability of each data source used to produce a finding. DOM-parsed data has the highest source reliability — it is taken directly from the rendered page with no inference. Schema-parsed data is slightly lower — it depends on correct implementation of structured data by the publisher. AI-generated analysis (contradiction detection, entity extraction quality) has the lowest source reliability because it introduces model uncertainty. Source reliability directly affects confidence scores.
◆Filter your Issues list by confidence score. A missing H1 at confidence 1.0 and a potential entity contradiction at confidence 0.61 are both real findings, but they warrant very different investigation approaches. The high-confidence deterministic issues should be fixed immediately. The lower-confidence inferred issues should be investigated before acting.
Adjusted Severity
Confidence scores affect how findings are presented. A critical issue at confidence 0.95 is reported as critical. A critical issue at confidence 0.55 has its severity adjusted: presented as "potential critical" with a confidence indicator and the specific evidence that triggered the finding. This prevents lower-confidence inferences from being treated with the same urgency as high-confidence deterministic findings — a problem common in audit tools that conflate all findings into a single undifferentiated list.
Extraction Consistency
Extraction consistency measures whether the same piece of data, extracted by different methods, produces the same result. An entity name extracted from schema that matches the entity name extracted from body text that matches the entity name extracted from meta tags has high extraction consistency — it is almost certainly correct. An entity name that differs across extraction methods has low consistency — the data is unreliable and the finding should be treated with lower confidence. Consistency checks run across schema, DOM, and AI-extracted data for all entity attributes.