populace

strategies / composition

strategy 05 · composition

The interactions.

The support and the weights each pass their own single-operator checks in isolation. A real release runs both on the same file — imputed values get reweighted, and a distortion invisible to either operator's own validation can appear only in that combination. Composition is the study of what actually happens there.

in progress paper: design doc, pre-experiment Read the plan Repository
01 — the 30-second version

Four questions, none answered yet.

The imputation and calibration strategy pages each report results from their own dossier. Neither dossier can see what happens when both operators touch the same file — imputation is validated before any calibration runs, and calibration is validated on frames that were already imputed some fixed way. Composition asks the question that sits between them, and — as of this page — has not yet run the experiments that would answer it.

The interaction neither operator's validation can see The support operator and the weight operator each pass their own single-operator checks in isolation. A real release runs both on the same file, and a distortion invisible to either operator's own validation can appear only in that combination. The illustrated mechanism: a low-weight extreme record — a heavy tail value the imputation step drew, harmless at its design weight — has its weight multiplied up when the calibrator leans on it to hit a target, and the extreme value times the inflated weight blows up an aggregate. This realized landmine is the interaction blamed for a prior incident in the enhanced CPS pipeline; composition has not yet run the experiments to measure it. operator · support imputed file passes its own checks operator · weights reweighted file passes its own checks the composition shipped file checked by neither alone ? the interaction — invisible to either operator's single-operator validation the landmine mechanism (illustrative — not yet measured) before reweight extreme tail value × small design weight → harmless calibrator leans on it same extreme value × inflated weight (≤50×) → aggregate blows up one record's weight, reweighted design weight inflated up to 50× to hit a thin target — the bound is the guard; composition asks if it holds
Each operator passes its own checks; the shipped file is the composition of both, and a distortion can hide in that combination alone. The mechanism the first experiment targets: a heavy tail value the imputation step drew sits harmless at its small design weight, until the calibrator multiplies that weight up to hit a thin target — and the extreme value times the inflated weight blows up an aggregate. This realized landmine is the interaction blamed for a prior incident in the enhanced CPS pipeline. It is drawn here as the question composition exists to answer, not a measured result.
01planned

Realized landmines

Run actual calibrators on each imputation method's output and measure realized tail distortion — the interaction blamed for a prior incident in the enhanced CPS pipeline. Not yet run.

02planned

Repairability, quantified

How much of a weight-blind imputation's tail inflation can calibration claw back, and at what weight-concentration cost. Not yet run.

03planned

The target-vs-survey trade-off

An unscalarized Pareto frontier between administrative target attainment and population-view score, rather than a single blended number. Not yet run.

04planned

Effective-support collapse

Coverage variants restricted to the records carrying the top 99% of mass, to check whether composition quietly abandons the tail while still passing a whole-sample coverage check. Not yet run.

blocked on populace#302 (holdout-masked builds), the calibration-paper method-adapter registry, and the popdgp harness extraction — all in progress in their own repos

02 — why this is its own strategy

Neither operator dossier can look here.

The support paper scores the fill operator against survey holdouts before any calibration runs. The calibration paper plans to score the reweight operator with imputation held fixed. Both designs are correct for isolating one operator's contribution — and both are structurally unable to see what the composition of both operators does to the file that actually ships. That gap is this strategy's entire scope.

Read the experiment questions as filed.

composition-paper on GitHub Read the plan