Intervene

W12 · 5 h 20 min

Optimization under competing queries

Essential question

Which gain survives a portfolio of intents?

What you should be able to do

  1. Define competing objectives and hard constraints.
  2. Interpret a Pareto frontier.
  3. Design a bounded ablation and preregister a choice rule.
PrerequisitesW06 strata and weights.W08 estimand.W09 single-factor intervention.
Builds on

W06 strata and weights. · W08 estimand. · W09 single-factor intervention.

Investigates

Which gain survives a portfolio of intents?

Feeds

Intervention design review.

Arrive with a prepared artifact

Reading route
Core PAPER-03/PAPER-31; Extend PAPER-04/PAPER-27/PAPER-30.
Viewing route
Open the complete text-first lecture packageSix-minute Pareto and ablation example. The current equivalent is notes, slide script, worked case, and no-video transcript; no recording is claimed.
Readiness check
List two objectives, one budget, and three non-negotiable constraints.
Bring
Preregistered objective and choice-rule card.

Concepts, assumptions, and boundary

One intervention serves multiple intents

Content changes can help one query class while harming another. Optimization therefore needs a declared portfolio, objective vector, query weights, content or cost budget, and hard factual, legal, safety, and accessibility constraints. A weighted sum is a policy choice, not a natural truth. Pareto analysis makes trade-offs visible by showing solutions that cannot improve one objective without worsening another.

Ablation tests the mechanism claim

An ablation removes or varies one component while preserving the rest of the design. The analysis should report per-stratum effects, dominated solutions, constraint violations, and uncertainty. Selecting weights after viewing results turns the objective into another researcher degree of freedom. The choice rule, tie handling, and failure threshold are therefore frozen in advance, and a null or conflicting result remains a valid outcome.

Inspect the mechanism or evidence structure

Synthetic illustrationConstrained Pareto choice—schematic only
Constrained Pareto choice—schematic onlySeven deliberately schematic candidate positions occupy axes labeled Objective A and Objective B. Five illustrative points form a frontier, one is dominated, and one crossed point is excluded before choice by a hard constraint. Positions do not represent a dataset or estimate.Objective A · schematicObjective B · schematicConstraint-excludedexcluded before choiceillustrative set · not to scale
Figure design. A not-to-scale two-objective diagram distinguishes an illustrative nondominated frontier, dominated candidates, and a constraint-excluded candidate. It contains no observed values, sample size, or uncertainty estimate. Schematic candidate positions are not observations or estimates. Weights express policy; facts and safety remain non-negotiable.
Long description

Seven deliberately schematic candidate positions occupy axes labeled Objective A and Objective B. Five illustrative points form a frontier, one is dominated, and one crossed point is excluded before choice by a hard constraint. Positions do not represent a dataset or estimate.

One action, one feedback state

Action

Move objective weights and inspect which solutions become dominated or violate constraints.

Feedback

The state reports the implied policy choice and never permits factual integrity or safety to become a soft weight.

Accessible alternative

A table enumerates all weight settings and selected solutions.

Open the intervention-diff explorer

Produce a reviewable intermediate file

Task
Compare bounded interventions across a weighted query portfolio.
Inputs
Frozen per-query outcomes, objective template, constraint matrix, plotting script.
Timebox
75 minutes
Intermediate file
ablation-matrix.csv and pareto.svg
Stop condition
Exclude any solution that changes facts, fails safety/accessibility, or exceeds the authorized scope.

Why each controlled source is here

Complete, retrieve, and revise

Checkpoint
Intervention design review.
Reflection
Explain who or which query class loses under the selected solution.
Revision
Add per-stratum uncertainty and a null-result decision path.
Low-compute route
Use the supplied outcome matrix; only table calculations and a lightweight plot are required.

Five retrieval questions

01What makes a solution Pareto dominated?

Another feasible solution is at least as good on every objective and strictly better on at least one.

02Why freeze objective weights?

Post-hoc weights can select a preferred result after observing outcomes and obscure who bears losses.

03Which constraints should not become soft rewards?

Factual integrity, authorization, safety, privacy, legal obligations, and essential accessibility.

04What is the evidence boundary for W12?

Optimization results are objective-, budget-, model-, and query-distribution-specific; they do not establish a universally best content strategy.

05What must you submit or revise after this week?

Intervention design review. Add per-stratum uncertainty and a null-result decision path.

Continue in the practice package