Audit rules

lumen’s audit engine ships 36 built-in rules across meta, content, links, technical, performance, accessibility and social categories. Rules are small, evidence-first, and plugin-shaped — the audit engine accepts additional rules via the plugin registry.

Severity model

Three severities, no vibes:

failThreshold (in lumen.config.json) picks the lowest severity that fails the run — error by default.

Scoring

Per-page score = 100 minus the weight of that page’s issues, floored at 0. Report score = the mean over audited pages. A site where nothing was actually audited scores 0 — honest, not generous.

The 36 rules

What lumen does and doesn't claim about AI search

The ai-search rules report facts: which AI crawlers your robots.txt allows (from a vendored, dated list — theasOf date and source are in the report), whether/llms.txt exists in the llmstxt.org shape, and whether your server HTML contains the primary content. They do notclaim any ranking effect: llms.txt is an unproven proposal, AI-crawler access is your editorial choice, and the JS-content signal is a labeled heuristic. lumen never recommends a training policy — blocking or allowing training bots is entirely yours.

Length thresholds (title, description) and the latency threshold are configurable; the thin-content word floor is configurable too; everything else is fixed behavior. Crawl rules (broken-internal-link, redirect-chain) only fire on evidence gathered during the crawl itself — a link target that was never fetched is reported as unknown, not invented.

Overrides

Per-rule severity overrides live in severityOverrides in lumen.config.json. An unknown rule id is a config error listing the valid ids — see configuration.

Issue + incomplete reports

An Issue looks like this (evidence snippets are escaped):

{
  "ruleId": "title-length",
  "severity": "warning",
  "message": "title is 72 characters (max 65)",
  "evidence": { "selector": "head > title", "snippet": "<title>Acme Garden Tools — hand-forged broadf…</title>" },
  "fixHint": "shorten the title to 65 characters or fewer"
}

Interrupted runs are labeled rather than dressed up as finished:

{
  "id": "example-2026-08-29t10-31-04z",
  "incomplete": true,
  "summary": { "pagesAudited": 37, "pagesSkipped": 3 },
  "stopReason": "time_budget"
}