Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mehen.ophi.dev/llms.txt

Use this file to discover all available pages before exploring further.

Evidence Coverage measures structural support for each section. Where Repository Grounding is a document-level signal, Evidence Coverage is per-section, so one well-linked section cannot hide many unsupported ones.

​
Formula

anchor_density_s   = evidence_anchors_s / max(1, W_s / 250)
section_evidence_s = sat(anchor_density_s; 0.2, 1.5)

EvidenceCoverageScore =
    0.5 · mean(section_evidence_s)
  + 0.5 · p25(section_evidence_s)

​
What is an “evidence anchor”

For each section, mehen counts:
  • Resolved internal anchors.
  • Resolved relative repository links.
  • Code fences with explicit language tags.
  • Diagram blocks (Mermaid, GraphViz, etc.).
  • Table headers.
  • Footnote references.
These are normalized against the section’s word count (per 250-word window).

​
Why the 25th-percentile term matters

The mean alone hides skew: one section with 10 anchors and four sections with 0 anchors averages 2, which looks fine. The 25th-percentile term ensures the score reflects the worst supported sections.

​
How downstream metrics use it

Evidence Coverage feeds DMI indirectly via Filler / Lazy Risk and is reported in the PR comment drill-down.

​
References

  • Daugherty, S. R. & Krueger, K. R. (1991). The 90-10 rule of unequal coverage and its ramifications for measurement. Psychological Reports 68(3) — motivation for percentile-based reporting (the 25th-percentile term that prevents one well-anchored section from hiding many unsupported ones). DOI.
  • Pirolli, P. & Card, S. (1999). Information Foraging. Psychological Review 106(4): 643–675 — source for the “evidence anchor” concept (each link, code block, or table is a foraging cue). DOI.

​
See also