A constrained slate needs a feasible exposure allocation, not only a weighted score

Let each eligible item have a declared utility estimate and let each displayed position have an exposure weight. Position-weighted utility combines them under an explicit model; it is not automatically a causal estimate of satisfaction. Exposure records how visible items are expected to be, which can differ from a raw item count. A group appearing once in the last position may receive less exposure than another group appearing once at the top. State whether the weights are measured, modeled, or simply chosen for a teaching calculation.

A randomized ranking can be described by the probability that each item occupies each position. Valid assignment probabilities must correspond to an admissible policy: positions cannot contain multiple items and an item cannot occupy two slots in one slate. Expected utility and exposure then average over that policy's randomness. A deterministic request constraint requires every realized slate to satisfy it; an expectation constraint allows individual slates to differ. Confusing the two can make a statistically feasible allocation violate the promise made for a particular request.

Hard eligibility can exclude items before any objective is optimized. Minimum representation requirements can also fail if the candidate union contains no eligible member of a required group. For disjoint groups that partition the eligible items, the total available exposure in a fully filled slate is the sum of position weights. Requested minimum expected exposures cannot sum beyond that total. This is only a necessary check: item availability, per-slate rules, position restrictions, and other constraints can still make an apparently reasonable set of minima infeasible.

A weighted objective usually expresses tradeoffs among feasible alternatives; it does not guarantee a hard constraint merely because its coefficient is large. Conversely, imposing a constraint can reduce the chosen utility proxy while serving another declared objective. Report that tradeoff rather than presenting every metric as free to improve simultaneously. When no feasible slate exists, apply a documented priority or fallback rule and record the violation context. Silently removing difficult requests from aggregate reports hides exactly the cases needed to evaluate the design.

Specify whether constraints apply within one request, a session, or a population window, and preserve the denominator and policy version. Inspect user and item cohorts because an aggregate exposure target can conceal a subset receiving no useful recommendations. Validate utility predictions and outcomes with an appropriate experiment; a mathematical optimum for a click proxy may reward behavior that does not improve long-term user benefit. Include candidate-source limitations and latency in the serving evaluation so a constrained reranker is judged on the slate it can actually construct.

The probability lab can illustrate the random count in one explicitly chosen mixture. Reset, set both p=0.2 and n=10, predict mean two and variance 1.6, then Play or Step. Reset and set both p=0.2 and n=20 for mean four and variance 3.2. Its exact binomial bars assume independent fixed-probability decisions and playback highlights outcomes rather than sampling. It has no item assignment, exposure weights, eligibility, constraint solver, or user utility. A long-run expected count in that toy model does not prove every realized slate satisfies a hard request-level requirement.