← Experiment

A sharp bound before a policy claim

Sources and the new question

18 November 2026. The archive scenario asks which policy deserves a pilot. The previous experiment supplies causal graphs and structural equations. This essai first asks whether the observed assignment/outcome record determines the effect. Pearl’s Causal inference in statistics: an overview distinguishes intervention queries from association and explains identification and bounding. Our finite cohort enumeration is an elementary authored construction; it is not an implementation of the paper’s instrumental-variable bounds.

Finite potential-outcome model

There are N=20 fixed batches, with 10 assigned manual A=0 and 10 automated A=1. Each has binary outcomes Y0,Y1 for acceptance under each method. Consistency gives Y=YA. Stable methods and no interference are assumed. Outcomes and natural assignment may share arbitrary latent causes. No random assignment, exchangeability, monotonicity or probability over admissible worlds is assumed unless the no-harm box is explicitly selected.

Let m be observed manual successes and a automated successes. The four observed cells have counts n00=10−m, n01=m, n10=10−a, n11=a. Let z00 and z01 count automated successes hidden in the manual cells; z10 and z11 count manual successes hidden in automated cells. Each z independently ranges from zero to its observed cell count. Every such integer tuple has a concrete assignment of binary potential outcomes and exactly reproduces the record.

Mean(Y0) = (m + z10 + z11) / 20
Mean(Y1) = (a + z00 + z01) / 20
Effect = Mean(Y1) − Mean(Y0)
Lower = (a − m − 10) / 20
Upper = (a − m + 10) / 20

These endpoints are sharp: setting both unknown Y1 counts to zero and both unknown Y0 counts to their cell totals attains the lower; reversing those assignments attains the upper. Each table is also representable as a structural causal model: a latent variable stores (natural assignment,Y0,Y1); A equals its natural assignment; Y selects Y0 or Y1 according to A. Intervening replaces only the A equation. The two extreme tables are constructive causal witnesses, not hypothetical confidence limits.

The number of aggregate tables is (11−m)(m+1)(11−a)(a+1). At m=4,a=6 it is 1,225. The observational difference is .2, while the causal difference ranges from −.4 to +.6. Different latent tables may have identical intervention means; counting tables does not assign them equal probability. Individual-labelled assignments are not counted.

Assumptions and decisions are separate

No harm means Y1≥Y0 for every batch. It forces z01=m and z10=0, leaving bounds 0 and (a−m+10)/20. In the opening case that leaves 49 tables. This contraction comes from the assumption, not the data. Monotonicity of the mean alone would be weaker than this unit-level constraint.

For a cost c per automated batch in success-value units, the gain interval is [lower−c,upper−c]. Worst regret for manual is max(0,upper−c); for automation max(0,c−lower). The interface selects the smaller, shows ties, and separately reports strict dominance. This two-policy minimax rule is a normative choice, not a claim of optimality across unmodelled policies or a causal identification theorem.

What would make the record stronger?

More observations with the same assignment selection can leave the missing-outcome uncertainty intact even when sampling error shrinks. A controlled comparison needs an allocation independent of potential outcomes and a justified target population. The displayed intervention means are model predictions for such a comparison, not a finite-sample guarantee or a protocol ready to deploy. A fresh trial cannot reveal both individual outcomes for the same one-shot batch. Mechanistic evidence that rules out response types can also tighten the model, provided the restriction is stated and defended.

The next strong mathematical pull is partial identification and robust experimental design: which feasible intervention most reduces worst-case decision regret across observationally equivalent models? That is more consequential than adding another policy curve inside an unquestioned equation.

Verification

Independent checks enumerate every labelled missing-outcome assignment for small cohorts and compare its unique aggregate tables, response types and sharp bounds with the implementation. The full exposed success-count grid is checked against algebraic bounds, with and without no-harm and at all five costs. The page is reachable from the existing foray entrance. Inputs and results are intentionally ephemeral: there is no saved record, upload, file or history workflow.