Pareto exploration

The editorial cost of clicks is the first claim this report makes. If the platform optimizes only for the articles a user is most likely to click, the newsroom gets a clean metric and a narrower product. ADR-0004 rejects that trade as the center of the tutorial: editorial accountability should be legible in the platform, not hidden inside a ranker or postponed to a manual review meeting. ADR-0009 turns that thesis into an evaluation shape by comparing constraint configurations across click metrics and editorial metrics at the same time.

The chart below is the compact version of the argument. Every point is a configuration evaluated against a publisher dataset. The x-axis is NDCG@10, which is the familiar click-prediction question: did the list put likely clicks near the top? The y-axis is intra-list diversity, which asks a different editorial question: did the list spread attention across topics instead of collapsing into the narrowest possible click pattern? Bubble size carries coverage so a point can be attractive for more than one reason. The point of the Pareto frontier is not to find a morally pure corner. It is to make the frontier visible enough that a newsroom can choose deliberately.

Moving along the frontier costs something. In this fixture, the click-first configuration has the lowest diversity weight and therefore the least pressure to widen the recommendation list. The balanced configuration spends some click efficiency to gain diversity, coverage, and fresher editorial shape. The diversity-forward configuration spends more. That phrasing matters: the cost is not a bug in the evaluation harness. It is the measurement the newsroom needs before it can decide whether the extra diversity is worth the loss in a click metric.

The decision also cannot be delegated to the chart. A Pareto point can say that one configuration dominates another, but it cannot decide how much topical variety a front page owes readers, or how much click loss is acceptable during a major news cycle. Those are editorial choices. The platform's responsibility is to make the choices concrete enough that editors and analysts argue about the same tradeoff rather than about competing anecdotes.

The table underneath keeps the wider metric family visible. NDCG@10 sits next to diversity, coverage, recency, sentiment distribution divergence, and sensitive-topic exposure. ADR-0015 is the reason those dimensions are comparable: every configuration is a named tuple of soft weights and hard-rule settings. Without that tuple, an analyst would be comparing charts that happen to share labels. With it, the analyst can say "this point changed because the diversity weight moved" or "this point changed while the hard guardrail stayed constant."

The useful reading is not "maximize diversity" any more than it is "maximize clicks." The useful reading is that a configuration makes a policy visible. If the newsroom wants a stronger diversity posture, the chart shows the click metric it is likely to give up and the coverage it is likely to gain. If the newsroom wants a click-first posture for a particular product surface, the same chart shows which editorial measures will narrow. A platform earns trust when those choices are inspectable before they become defaults.

That is why this page stays close to the data rather than turning into a scorecard. A scorecard would hide the frontier by declaring a winner. The frontier is the story: editorial policy is a movement through tradeoffs, and the analyst surface makes the movement legible enough to debate.

The next useful question is not whether the frontier exists. It is which frontier a newsroom wants for each product surface. A breaking-news feed, a personalized homepage, and a weekend reading module may all choose different points. The same analytical contract lets those choices be compared without pretending they should collapse into one universal optimum.