Sensitive-topic exposure trend

This is the most editorially loaded report in the analyst surface. It argues for the constraint, not against it. Sensitive-topic exposure is not a metric that says sensitive journalism is bad, or that hard stories should be hidden from readers. It is a metric that asks whether the recommendation list is amplifying sensitive material beyond the limits editors intentionally set. That distinction is the reason the sensitive-topic guard is a hard rule in ADR-0010 rather than just another slider blended into a score.

The line chart below follows sensitive-topic exposure across the same configuration sweep used in the Pareto report. The x-axis is diversity weight, which means the chart is not pretending sensitivity is only a sensitivity problem. Editorial constraints interact. A list that becomes more diverse can also change which sensitive articles appear. A list with better click metrics can still violate an editorial boundary if the guardrail is missing or set too loosely. The analyst task is to inspect the interaction without turning it into a simplistic "less is always better" story.

The careful reading is that exposure should be bounded, not erased. A newsroom may need to recommend reporting on conflict, crime, death, courts, or other public-interest topics. The platform should not silently suppress that work. But the platform also should not let a relevance model fill a list with sensitive material simply because those stories attract attention. ADR-0015 defines the configuration shape that lets those two statements coexist: soft-weight tuning for diversity, recency, and sentiment; hard enforcement for the sensitive-topic cap.

That is why this page keeps sensitivity_guard_enabled and max_sensitive attached to the metric. Those settings are editorial context. If exposure is low while the guard is enabled and the cap is zero, the report says the hard rule is doing its job. If exposure remains low when the cap moves to one, the report says the platform allowed a limited amount of sensitive material without letting the list become dominated by it. If exposure rises sharply under a future configuration, the next question should be whether that rise came from a deliberate policy change, a source-mix change, or a bug in enforcement.

The summary table turns that reading into an audit question. For each publisher and guard setting, it shows the average, minimum, and maximum exposure observed in the sweep. The numbers are not a substitute for editorial judgement. They are the substrate for editorial judgement: a way to ask whether the policy is being enforced consistently enough that editors can reason about it.

This is also where the report needs restraint. The chart can show exposure moving down as a cap tightens, but it cannot know whether the underlying stories are exploitative, necessary, repetitive, urgent, or central to the public record. The page therefore treats the metric as a warning system, not as a moral verdict.

This report should therefore be read with two cautions. First, a lower exposure number is not automatically a better newsroom outcome. It may indicate a healthy cap, or it may indicate that important journalism is being kept out of recommendation lists too aggressively. Second, a higher exposure number is not automatically a failure. It may be correct during a major public-interest news cycle, provided the limit was chosen deliberately and recorded in the configuration history.

The guardrail exists because automated recommendation systems can over-amplify signals that editors would handle with more care. The analyst surface gives the organisation a place to inspect that risk in plain data. It does not replace editorial judgement with a number. It gives editorial judgement a stable table, a trend line, and an audit trail to argue from.