AI readiness toolkit

/assess: the truth-pressure model, turned on itself

This release closes #34 — a gap surfaced by /assess's own feedback loop. Layer 0 could stay Present on a good grade from one committed file even while the repo's advertised instruction surface was broken. Now the framework scores its own "advertised-but-broken" failure the way it scores everyone else's: a broken map is worth at or below absent.

Instruction-surface integrity (Layer 0)

• Untracked instruction files are surfaced, not silently credited or dropped. An on-disk-but-uncommitted CLAUDE.md is reported ("present locally but won't reach contributors") and never raises the grade — only git-tracked files are graded.
• Broken instruction references now penalise Layer 0. A committed .cursorrules that is a dangling symlink, or an entry doc that links to a missing CLAUDE.md/AGENTS.md, caps the instruction half at Partial (Missing if it's the primary surface) — regardless of an unrelated file's grade — is named as a Top-3 fix, and surfaces the within-band regression the coarse Present/Partial/Missing bucket otherwise hides.

Also in this release

• Broken links to the same missing file now collapse into a single "ghost" node on the doc-navigability graph (less clutter when many docs reference one missing target).

Both the Claude Code plugin and the standalone Desktop/Cowork ZIPs are updated (ZIPs on the standalone-skills-latest rolling release).

What's Changed

Features

• feat: Instruction-surface integrity in Layer 0 (closes #34) by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/36

Fixes

• fix: Collapse broken links to the same missing file into one ghost node by @bjcoombs in https://github.com/bjcoombs/ai-native-toolkit/pull/35

Full Changelog: https://github.com/bjcoombs/ai-native-toolkit/compare/v1.9.0...v1.10.0