How We Made Definitive-Article Strength Measurable

I asked my agent what sounded like a presentation question: put the definitive pill on every definitive article and show how strong each one is from the number of meta-articles behind it. The useful part was not the pill. It was forcing the content system to prove which pages are actually canonical, which examples genuinely belong to them, and whether a reader can travel in both directions between the public instructions and the runnable task.

The maintained method for documenting a completed run is my Meta Article Prompt. This article is one such run: what we checked, what was wrong, what changed, and the rule the next agent inherits.

The loop we made visible

Why a backlink count was not good enough

A first pass found plenty of links, but many did not represent worked examples for the page they touched. The same Content Factory map appeared across many articles. Related-reading lists pointed to broad system pages. Some instructional pages carried an old Meta Article tag. Counting those links would reward templates and taxonomy mistakes instead of really real experience.

So the score now counts a post only when all of these are true:

1. It is published and publicly reachable.
2. The reviewed audit explicitly classifies it as a meta-article or worked example; taxonomy is supporting evidence, never proof by itself.
3. It documents a completed run, not instructions, an index, an archive, or a placeholder.
4. Its main deliverable materially executes the definitive hub’s task.
5. The task-specific narrative links to the exact canonical hub, not merely to a generic framework map or related-reading list.
6. The hub links back to the counted meta-article in its generated evidence footer.
7. The Task Library links to the hub, and the hub links to the exact filtered Task Library view.

The strength bands

The exact count remains visible. The label is only a compact reading aid; it measures documented-run volume, not writing quality, freshness, traffic, or semantic certification.

The canonical-page mistake the badge caught

This is why “definitive” must remain a semantic claim. A completed task list, a category, an old tag, or the word definitive in a title cannot certify a page. The canonical owner must agree with what the public page says.

We also repaired 43 stale guideline links across 36 published posts. Only three of those posts primarily documented an Article Guidelines run, so only those three strengthen that hub. The other 33 remain useful supporting links without inflating the evidence score.

How priority affects the rollout

Task importance decides where we work first; it does not change the evidence score. A hub with many high-importance mapped tasks gets inspected and repaired before a one-task supporting SOP. Its strength still comes only from verified meta-articles. Keeping those two facts separate prevents an important task from looking well documented when it is not.

What changed

• One generated manifest now owns the hub URL, mapped tasks, counted and held evidence, exact count, strength band, and audit time.
• The top of each certified hub shows its definitive status and derived meta-article count together.
• A collapsed footer lists every counted worked example and links to the hub’s filtered Task Library view.
• The Task Library exposes an article-hub index and stable ?article= and ?task= routes.
• Zero, partial, and unknown remain different states. An inaccessible corpus never becomes a flattering or discouraging zero.
• Archive, WIP, and semantically conflicted pages receive no definitive marker.

The implementation and its fail-closed tests are public in the Task Library rollout. The shared rule that future agents inherit is in the merged definitive-article standard.

The live numbers belong to the generated hub footers and the Local Service Spotlight Task Library, where they can be rebuilt when a new meta-article is published. They are intentionally not copied into this narrative as a second source of truth.

The larger lesson: a badge is useful only when it reveals the graph behind it. If the count cannot name the worked examples, link back to them, and survive a rebuild from the source corpus, it is decoration.