/map-check

The read-only conformance front door for Map Your Knowledge (MYK) — the single checking engine. It validates a scope against the frozen contract's deterministic catalog and never invents rules of its own.

kernel-version: MYK v2.3 f2f96f2de49b4863bca55ee8f6004d24e00574a7db5e7e5ef0e3cb28c42510cf

The two iron rules

1. Read-only forever. map-check never writes, renames, or proposes-and-applies. Findings go to the Handler, and repairs run through map-this (one-off) or map-maintain (the tending loop).
2. Unqualified CLEAN is banned. The only positive verdict is: "clean for checks {list} over inventory {source}, mode {mode}; not checked: {list}." A crash can never read as a pass: the JSON envelope separates operational errors from findings by shape.

Running it

The bundled deterministic script does the mechanical work (no LLM in the validation path):

node <this-skill>/references/tools/map-check.mjs <target-path> [--json]

• Target classes: an explicit path inside a scope, or a scope root (a directory containing .myk/README.md or a declared root map). Ambiguous or contested scopes fail closed (scope-conflict).
• Modes: FULL (a .myk/README.md contract is present — its m1.root-map and m3 declarations govern) or DEGRADED (no contract — the agent supplies the declared root map from kernel markers; absence of any declaration is NEVER a finding: undeclared trees are outside jurisdiction, and the script says so and stops).
• Inventory honesty: the script walks the scope itself (never following links/junctions out of it; skipped boundaries are counted against completeness) and reports inventory: {source, complete, scanned}.
• Exit codes: 0 ran, zero findings (qualified-clean) · 1 ran, findings present · 2 operational error · 3 target/scope resolution error.

The checks (each pinned to its kernel rule; the script prints this table per run)

Id	Checks	Finding classes
C1 owner-declaration	every eligible Markdown member declares exactly one meta.map; the declared root map declares none	missing-owner · multiple-owners · root-declares-owner
C2 reciprocal spine	roster derived from member-side meta.map (authoritative): each declared owner map exists AND links the member; untyped map links are never membership claims	spine-drift-mapside · dangling-owner
C3 address form	duplicate-basename targets (incl. any README owner) use path form, never bare	bare-duplicate-basename
C4 meta container	frontmatter parses; single well-formed meta; one meta.map value; the exact legacy key meta.parent	parse-error · malformed-map-value · legacy-owner-key
C5 archive consistency	meta.archived implies the archive callout; archive/ folders reported as likely-unsignalled	missing-callout · unsignalled-archive
C7 case-fold collisions	sibling names colliding case-insensitively	case-fold-collision
C8 exclusion integrity	M3 entries well-formed; excluded trees generate zero other findings; managed artifacts reachable per their declared graph-entry	malformed-exclusion · managed-artifact-unreachable

Not checked, permanently, reported every run: prose quality; semantic-role presence in free-form maps; earned-scope judgment; whether an undeclared tree should be organized; tag semantics beyond syntax; positive cross-vault reference validity; frontmatter-contained links (resolution reads note BODIES only — links inside YAML frontmatter are invisible to R2 and to map-scan alike); lifecycle VALUE legality where no lifecycle layer is declared (C6 runs only when the root map declares the layer — v1 script reports it not-checked); snapshot exemptions (C10 — needs Handler-declared pairs; v1 reports not-checked); the 8-field frontmatter ceremony (C9 — vault-policy-dependent; v1 reports not-checked); anything that is a state of the work.

The v1 script boundary — declared, not implied

The accepted design (map-check design v3) is the FULL contract; the bundled v1 script implements a declared subset and says so in every run's not_checked list. Implemented: C1, C2, C4, C5, C7, C8 in full deterministic form; C3 limited to owner-declaration resolution, duplicate-basename/ambiguity, and the raw-cross-vault-wikilink prohibition; and — per the v2.4 link-resolution rider and its two amendments — R1 (declaration grammar, independently parsed: dialect + acceptance lists, fail-closed) and R2 (occurrence-level link resolution and closure, independently implemented: 11 terminal classes, fragments at exact heading/^block granularity, governed boundary targets, normalized comparison records pinned to the inventory/contract/governed-target fingerprints). The walk is bounded (--max-files / --max-ms); a cap that prevents canonical-inventory closure is a typed check-incomplete operational envelope (exit 2), never a clean verdict. The R implementation shares the rider's declared semantics with map-scan but no code — disagreement between the two tools' normalized records is evidence, never suppressed. Staged for the next script revision, by name: C6 (lifecycle values, where declared), C9 (frontmatter ceremony, vault-policy), C10 (snapshot pairs), scope-local-instruction discovery (marker 2 automation — today the caller supplies --root-map), Unicode/Obsidian heading-slug normalization and embed semantics (exact-match only today), and qualified-Lyt-vault targets with the typed resolver-unavailable error. This boundary is itself a reviewable claim — it stands only while the peer reviewer and the Handler accept it; narrowing it further requires the same acceptance.

Suppression sources — exactly two

M3 declarations in the scope contract, and Handler-confirmed rulings persisted in the style record. Raw prose suppresses nothing. Every suppression is listed in the output with its provenance.

Agent duties around the script

Run it; relay its output faithfully (verdict form intact — never upgrade "clean for these checks" into "all clear"); where the script reports DEGRADED or incomplete inventory, say so in the first sentence; route findings to the Handler with map-this/map-maintain as the repair paths. If the script cannot run (no Node), fall back to the behavioral checks in map-rules and label the result AGENT-CHECKED, SAMPLED — never script-equivalent.

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): if this run surfaced a concrete, blocking-or-recurring weakness in this skill, propose a specific fix for the handler to approve. Conservative — silent otherwise. Never auto-apply.