Finding the orphan initiatives in a strategy

In an X-Matrix, links only exist between adjacent quadrants: vision to annual objectives, objectives to initiatives, initiatives to KPIs, vision to KPIs. That constraint is what turns a vague sense of misalignment into arithmetic.

An element is an orphan when it appears in no correlation of non-zero strength. Not "weakly linked" — not linked at all. On that definition, finding orphans is a set operation over the correlation graph, and it takes milliseconds.

The related and more revealing case is the broken chain: an element that is linked, but whose thread does not reach the end. A vision correlated to an annual objective that no initiative delivers is not an orphan — and it is arguably worse, because it looks connected on the wall.

The completeness score turns the orphan ratio into a sub-score, weighted at 20% of the total, using deliberately blunt bands: half or more of your elements orphaned scores 20 out of 100; a third or more scores 40; a sixth or more scores 60; any orphan at all caps you at 80; only a fully wired matrix scores 100.

The bands are coarse on purpose. The difference between 14% and 16% orphans is noise; the difference between "a few loose ends" and "half the strategy is decoration" is the whole diagnosis.

Read the number as a symptom, not a grade. A high orphan ratio early in a planning cycle is normal — you are still wiring. The same number in month nine means the matrix was filled in once and never used to steer.

• Orphans weigh 20% of the completeness score; complete chains weigh 25%.
• Together with correlation density (20%), connectedness accounts for 65% of the score — volume of elements accounts for 15%.
• The full specification is public and the engine is MIT-licensed, so you can audit exactly how your strategy is judged.

Two patterns come up in almost every real matrix, and both are more interesting than a plain orphan.

The unmeasured objective: an annual objective wired to initiatives but to no indicator. The work is happening and nobody can say whether it is working. In the score this shows up in the indicator balance, and it is the single most common finding we see.

The lagging-only objective: an objective tracked exclusively by result indicators — revenue, NPS, churn — with no leading indicator on the process that is supposed to produce them. It is not an orphan, it is worse: you will discover the failure at the exact moment it is too late to react. The score flags it by name and asks for a process indicator.

And one honest caveat: the score says nothing about whether an objective is ambitious, whether a KPI is measurable, or whether an initiative is funded. Those are judgement calls, and inventing a number for them would be the false precision the score exists to avoid.

The reason to make this mechanical is not elegance, it is repetition. Detection that depends on a human noticing happens when someone is preparing a board pack; detection that is a function call happens every week, in the conversation, whether or not anyone wanted to hear it.

On hoshin.app the check is an MCP tool, so your own assistant runs it: `get_matrix_completeness` returns the score, the sub-scores and a list of suggestions naming the elements at fault, in a form your agent can act on rather than a paragraph it has to interpret. Ask it to wire what it found, and it can — `set_correlation` and `upsert_element` let it fix the structure it just criticised.

A test we run in CI asserts that applying the score's own suggestions raises the score. That sounds obvious; it is not. A diagnostic that cannot be acted on is decoration, and we would rather find out in the build than in a review.