wiki-boundary

Deterministic ranking of the pages everybody links to that have little in them. Borrowing C10 of [[roadmap-emprunts]] §2.17. Read-only, no LLM: one query over the persisted knowledge graph.

The score proposes attention. It does not establish importance. A high score says exactly one thing: many pages point here, and there is not much here when you arrive. That is a reason to look, never a verdict on the page, its author, or its priority. Say this out loud when you present results — it is the whole contract.

Pre-condition

The knowledge graph must exist AND be recent enough to carry substance measurements: wiki-meta/graph/knowledge-graph.json. If the tool refuses with "no substance measurements", the graph predates this feature — offer to run /wiki-graph, then retry.

Always look at graphAnalyzedAt in the result. The graph is a snapshot. A month-old graph ranks pages that may have been rewritten, moved or deleted since. If it is more than a couple of weeks old, say so before presenting the list, and offer to rebuild.

Steps

1. Call the tool

mcp__obsidian-router__find_boundary_pages({ vault, limit: 10 })

Arguments, all optional:

• --limit N → limit (default 10, ceiling 100).
• --min-inbound N → minInbound (default 1). Raise it on a big vault to keep only real crossroads.
• --exempt-types a,b,c → exemptTypes, replacing the default list. Pass the defaults plus the additions — ["redirect","source","answer","index"], not just ["index"] — or you will silently un-exempt the rest. See step 1-bis.
• --exempt-statuses a,b,c → exemptStatuses. No default — closed pages (status: superseded, retired…) stay visible and labelled unless the user asks to hide them. Exact match, case-insensitive; a page with no status: is never exempted; superseded-in-part is not swept up by superseded. This asymmetry with types is a decision, not an oversight: the candidate default ["superseded"] was rejected after adversarial review, because the ADR contract that standardises the token covers decision pages only, and a global filter would silently erase legitimate historical hubs.
• --all-types → exemptTypes: [], scoring every page including the ones held out by default. Use only when the user explicitly asks to see everything; warn them that migration stubs and capture records will dominate.
• --as-of YYYY-MM-DD → asOf. By default recency is measured against the graph's own build stamp, which makes the ranking a pure function of the graph file.

1-bis. On a vault you have not run this on before: check the exemptions FIT

Do this before presenting anything. The default exemptTypes — redirect / source / answer — are not universal truths; they are one vault's vocabulary. Every vault names "a page whose job is to point elsewhere" differently, and that list is the single thing keeping the ranking meaningful (see the limitation note below: on the router's own vault, without exemptions, 12 of the top 20 were migration stubs).

So look at the top few results and ask of each: is this page thin because it was neglected, or thin because that is its job? The page's own frontmatter usually says. If a type: value means "stub / pointer / summary / moved", re-run with it added.

This is not hypothetical. The first run on a second vault (DEDIBOX) put a page at 2.68 — more than 1.6× anything the router's vault produces, and a clean outlier. It was type: index, kind: folder-index, status: redirect-summary, and its body said the real documentation had been migrated to another vault. A migration stub in all but name, which the default list does not catch because that vault calls it index. Re-run with --exempt-types redirect,source,answer,index and the ranking became sane.

Do not silently widen the defaults instead. index means "deliberate curated map" in some vaults and "leftover pointer" in others — growing the built-in list to cover every vocabulary is exactly the unfalsifiable creep the simple word count exists to avoid. Calibrate per vault, and say in your report which exemptions you applied and why.

2. Read the numbers before presenting them

The result carries everything needed to audit itself:

3. Present it honestly

Render a short table — path, score, inbound, words, age — then three sentences of interpretation, not more. Good interpretation names the specific reason a page might deserve work ("11 pages point at graphify, it holds 719 words, and nobody has touched it in 77 days"). Bad interpretation restates the score.

Then say what the list cannot tell you, in plain words:

• Index and hub pages will legitimately appear near the top, unless their type: is exempted (step 1-bis). A page whose job is to point elsewhere is thin by design, and the score cannot tell that from a page that is thin by neglect. Expect to dismiss one or two at a glance — that is the tool working as intended, not failing.
• A high score is not a promise that the page is thin. It can also come from a page that is merely well-linked: on DEDIBOX the top-ranked page after calibration held 986 words — not thin at all — and rose because ten pages cite it. It was still worth opening (a binding, critical: true gate, untouched for 82 days, its completion table still blank), but the reason was not the one the score implied. Say which of the two it is when you present a page.
• Read the status column before recommending research. A superseded/retired page in the top spots is a closed subject — say so, and if the user wants it hidden, re-run with --exempt-statuses and name the filter in the report. Do NOT hide it on your own initiative: a retired page with many inbound links is itself information (those links should probably point at the successor — a link-hygiene observation worth passing along, not a research topic). And the inverse trap is pinned in the tests: a topically closed page still marked status: active (KIVIRI's genesis page) passes every metadata filter — only reading the page catches it.
• A thin page is very often fine. A definition, a deliberate index, a disambiguation page. Read before acting.
• Low scores across the board mean the vault is in good shape. Say so rather than manufacturing urgency from the top of a flat list.

4. Offer the next step, don't take it

If a page genuinely looks under-served, the natural follow-up is /autoresearch to open a research programme on that page's topic — see the wiring note below. Propose it; never launch it. Autoresearch costs real tokens and needs its own confirmation.

How this relates to autoresearch

/autoresearch already picks its own questions: it reads ## Open Questions from a programme and takes the least-covered one, gauged with search_smart. This skill does not replace that and must not be wired into it. It sits one storey UP: autoresearch chooses which question inside a programme, this chooses which page deserves a programme at all. Feeding boundary scores into the question picker would break a mechanism that works.

How this relates to wiki-lint

wiki-lint reports boundary pages as an info section, never a warning. A crossroads that is thin is not a defect — nothing is broken, nothing needs fixing, and the lint report must not imply otherwise.

Note that the two count inbound links differently and that this is deliberate: lint Check A re-parses every page (and so sees wikilinks written in frontmatter), while this skill uses the graph (which parses page bodies only). The graph's set is a strict subset, so a page credited with inbound links here can never be called an orphan there — the two never contradict.

Anti-patterns

• Don't present the score as a priority list. It proposes attention; the human decides importance.
• Don't hide the exemptions. A ranking that silently dropped 31 pages reads as "I looked at everything" when it did not.
• Don't skip the calibration on a new vault. The default exemptions come from one vault's conventions; on another they can miss the very pages they exist to hold out, and the top of the list becomes noise that looks like signal.
• Don't add a type to exemptTypes without saying so. Every exemption is a page you chose not to show; that choice belongs in the report, not in your head.
• Don't rank a stale graph without saying so. Check graphAnalyzedAt first.
• Don't rewrite a page because it scored high. Read it. Thin is frequently correct.
• Don't chain into /autoresearch automatically.

Output format

🧭 Frontier pages — vault `<name>` — graph built <date>

| # | Page | Score | Inbound | Words | Age | Status |
|---|---|---|---|---|---|---|
| 1 | wiki/…/graphify.md | 1.63 | 11 | 719 | 77d | — |
| 2 | wiki/…/adr-modes-ecriture.md | 0.90 | 8 | 808 | 6d | superseded ⚠️ |

The Status column is not optional decoration: the API annotating closed pages
while the table hides the annotation is exactly the failure this feature fixed.

Held out: 31 pages (29 redirect, 2 source) — thin by design.
Exemptions applied: types = the defaults; statuses = none. [Name every addition
and its reason — e.g. "+ type `index`, this vault labels its migration stubs
that way" or "+ status `superseded` at the user's request".]
Ranked 103 of 140 articles; 6 have no inbound links (orphans are Check A's subject).

Score = inbound links damped by length (`inbound / (1 + words/100)`: full weight on an empty page, halved at 100 words, a tenth at 900), ×1 to ×2 for staleness.
It proposes attention, not importance.