The North Star metric, and why its textbook form fails

What the textbook claims. One metric that best captures the core value a product delivers to customers (Sean Ellis, GrowthHackers). Why only one: when two good things conflict, the star is the tiebreaker. A good star passes three tests: it tracks real user value, the team can move it with its own work, and it leads revenue rather than following it. Around it sit input metrics (the levers) and guardrail metrics (the floors that must not break).

Lucerna candidates, judged 2026-10-28. Registered users: vanity, marketing can inflate it, weak link to reading. Paid subscriptions: a lagging result of revenue, purchasable by discounting. Weekly readers completing an effective reading session: closest, value-related, team-movable, leading. Users finishing a whole book: real value delivered, but feedback too slow for a weekly cadence, better as a long-term health metric. No Lucerna pick was made: the session ended in a critique of the concept itself, not a selection.

The critique, point by point.

1. Analogy is not reasoning. Polaris is fixed and ships steer by it, but the name's origin proves nothing about how many metrics a company should track. An analogy can explain, it cannot prove.
2. The tiebreaker premise assumes real tradeoffs. Many decisions are win-win: fixing crashes and growing the library raise retention and word of mouth together. Where nothing trades off, the tiebreaker has nothing to break.
3. One number cannot induce a full ordering. The effect of each initiative on the star is unmeasurable in advance, horizons differ (two weeks versus two months), and attribution is unclear when several changes land together. What the star actually gives is a veto (kill what visibly harms it), a tiebreak for otherwise equal options, and a shared language for arguing, not a computed ranking.
4. Legislation is not judiciary. Setting the star only enacts the law; execution is interpretation, and interpretation diverges. Every proposal can tell a story about raising the star, because several causal layers sit between an action and the metric. The real star is the accumulated precedents: the definitions, the attribution rules, the time windows. Without them the star is an empty text.
5. The foresight premise fails. Needs and problems emerge in the doing: what users say they want and what actually blocks them are often different things. Predicting every pitfall in advance is circular, because a pitfall is defined by information you only gain after doing the work.
6. A warranted pivot is indistinguishable from rationalization in advance. Both look like unsatisfying data plus a new story; the difference only appears in hindsight. What helps is not a better judgment rule but a cheaper-error structure: small-step validation instead of all-in pivots, reasons written down before the change, validation criteria set in advance, pivots confined to fixed review windows, and external anchors such as observed user behavior and willingness to pay. The dead-star signal: the definition keeps being rewritten after the fact to match whatever is already being done.
7. Goodhart's law. When a measure becomes a target, it ceases to be a good measure (Charles Goodhart, 1975). A metric is a proxy for value, not value itself, and once rewards hang on the proxy, effort flows into the gap between proxy and value. YouTube is the gap-chase in public: clicks to watch time (2012, against clickbait) to satisfaction surveys (2016, against outlandish content) to satisfaction-weighted valued watch time, each generation gamed in turn. Amazon's structural defense is managing controllable inputs instead of financial outputs (Bezos, 2009 shareholder letter). Facebook's shift from monthly to daily active users, with addictiveness rising into public scrutiny, is the failure case.
8. No verified pure cases. Company strategy cannot be randomized, so no controlled study exists; the evidence is retrospective stories with survivorship bias and reversed causality. Airbnb's nights booked is revenue restated in nights, partly violating the leading test. Facebook's daily active users are success restated as a metric. Every surviving case is a variant: a star plus an input tree (Amplitude's own framework evolved exactly there), or one direction plus several key results (OKR).

The verdict (2026-10-28). The concept is self-contradictory. It works only if fixed: fixity, pre-commitment, and drift prevention are all properties of not moving. It survives only if changed: patching Goodhart, adapting to emergence, and settling interpretation fights all require moving. Fixed, it dies of Goodhart and emergence; changed, it dies of drift and becomes a shadow of whatever is already being done. What honestly remains is weaker but real: a brake pad that vetoes visibly harmful moves, a yardstick for honest retrospectives, and a shorthand for communicating what the team is doing. None of those requires it to stay put, and the promise that it stays put is the part that cannot be delivered.

Transfer. For a solo builder the alignment argument is vacuous: there is no team to align. The largest remaining use is anti-self-deception, a fixed reference against which to check one's own storytelling. And timing matters: a fixed star costs most in the exploration stage, when emergence is strongest; it is a scaling-stage tool for preventing drift after direction is settled, not an exploration-stage tool for finding direction.

Kin. retention-curve

Agent audit checklist: when someone proposes a North Star, check every line, and report each item with evidence.

1. Is any analogy doing the work of an argument? If so, strike it and re-derive the claim without it.
2. Name a real tradeoff this star would actually break. If decisions here are mostly win-win, say so.
3. Can the star rank two live initiatives in advance, or only veto and tiebreak? Say which.
4. Are the definitions, attribution rules, and time windows written down, or is interpretation still open?
5. What would falsify this star, and what validation was set before the next change?
6. Is the definition being rewritten after the fact to match current work? If yes, the star is already dead.