Counterfactual programme steering · Essai 2 · 18 November 2026
What the same project record cannot decide
Automation looks better in the record. Enumerate everything that record leaves possible before choosing a policy.
Twenty archive batches
Ten batches were assigned manual processing and ten automation. A success means a usable catalogue entry accepted within the same six-week window. Each batch has a fixed potential outcome under either method; only its assigned outcome was observed. Batches do not affect each other, and methods and acceptance criteria have consistent meanings.
This is a complete fictional finite cohort, not a random sample. Assignment may depend on difficulty that we did not record. We ask what would happen to these same twenty batches if all had received one method. There is no assumed exchangeability and no sampling confidence interval.
Hold the record; vary what it leaves unknown
Controls recalculate immediately. No records are saved or transmitted; reload restores defaults. The no-harm checkbox adds an assumption, not new evidence.
| Outcome under an intervention | Least favourable world | Most favourable world |
|---|
Inspect the response types that construct both worlds
These are hidden potential-outcome types across all twenty batches. They are not estimates of real archive types. The four unknown cell counts below also show how each world reproduces the assignment record.
| Type or unknown count | Least favourable | Most favourable |
|---|
A decision can still be explicit
Maximum regret compares a policy with the better of these two policies in each admissible world. Minimising it is a declared decision rule. It neither discovers the true world nor establishes that its selected policy has a positive causal effect. Implementation costs, values, equity and other policies would change the decision problem.
What evidence could separate the worlds?
A controlled allocation that breaks the link between batch difficulty and method can address the missing comparison. The table shows the intervention means that an ideal randomized comparison on the same target population would estimate in each extreme world. A finite pilot would be noisy; it would need uncertainty analysis, stable treatment definitions, no interference and a justified bridge to the target batches. Merely collecting more batches under the same selection rule need not narrow these identification bounds.
The no-harm assumption is especially strong in archive work: automation can introduce an error into a batch that a person would have completed correctly. Switch it on to see exactly how much of the conclusion that assumption supplies.
What moved forward
Essai 1 rehearsed causal paths and compared policies inside supplied equations. This essai asks whether the record determines the effect at all. It constructs every compatible aggregate potential-outcome table, obtains sharp lower and upper bounds, and gives actual worlds attaining both endpoints. Uncertainty about the model now constrains the decision instead of disappearing inside the simulator.
We do not infer an individual batch’s missing outcome, or claim that these are all causal explanations of a real project. These are all response-table completions within the stated finite binary model.
Derivation, source boundaries and verification · Executable enumeration