UNDER THE HOOD

Let code check.
Let Jev judge.

Not every audit step needs a model. Keep exact work deterministic and semantic questions small.

The deterministic layer

The collector reads selected HTML files or a validated inventory. It resolves local URLs, records link occurrences, checks fragments and keeps the exact extracted anchor/context. It does not browse your accounts, execute page scripts or fetch missing pages.

The semantic layer

For eligible links, an explicit Jev audit sends a bounded source context and destination excerpt to the official TypeSafe API. One request asks independent questions:

  1. Promise: does this destination satisfy what the anchor promises here? A Noul returns a yes/no probability.
  2. Relevance: is the destination relevant to the source paragraph’s immediate topic? A second Noul returns another probability.
  3. Anchor label: informative, generic, misleading, or insufficient context? A Choice returns the label, full distribution and confidence.

Noul does not have a separate confidence field. A probability near one-half means uncertainty, not “medium link quality.” Choice confidence is derived from its distribution. These are model judgments, not facts about search rankings.

The policy belongs to code

Conservative thresholds turn signals into findings or review. Conflicting evidence is not forced into a clean result. Bad API output, timeouts or exhausted coverage remain visible failures. The tool never asks Jev to invent a quote or explanation.

A report is an audit trail

JSON, Markdown and standalone HTML preserve source, target, original href, anchor and nearby words. Semantic decisions add requested/resolved model, usage, input truncation and signals. A baseline can identify existing findings, but cannot hide operational errors.

Where Jev is different

TypeSafe exposes Choice, Score and Noul as typed decision primitives. A generative model is still the right tool when you need new prose, code or extended reasoning. Jev’s own limitations include arithmetic, indirection, distracting context and adversarial input.

AnchorLint uses Jev as an advisory layer, not a correctness guarantee. No performance or ranking benefit is claimed from the interface alone.

Official Jev introduction ↗ · Documented limitations ↗ · Project research and comparisons ↗