Item 73: What is the baseline for a demonstrated improvement

The improvement theorem's verdict turns on a baseline the round supplies rather than derives, and the dependence is not cosmetic. cases.cm17_baseline_changes_the_verdict runs one trace against two admissible baselines:

baseline = the unmodified conduct    evidence 48.0  demonstrated   contested 180
baseline = an already-good policy    evidence  0.0  not demonstrated  escaped 180

Same process, same losses, same withdrawal, opposite verdicts. Both baselines satisfy the two conditions the round can state — predictable, and committed before the loss is revealed — so nothing available distinguishes them.

Five candidates were prosecuted and none is privileged: the unmodified procedure's output, the recorded pre-repair recommendation, a consumer-supplied reference, an explicit shadow execution, and an abstract comparison certificate. They are one type in the reference model, which is the round's answer to which is right: the theorem does not need to know, but a consumer does.

Deliverable shape: either a condition that selects the baseline from the process's own structure, or the finding that the baseline is irreducibly consumer-supplied together with what a consumer must state when it supplies one. Acceptance check: cm17 rerun under the proposal, with the two verdicts either reconciled or the divergence shown to be the correct answer. Context: projects/normativity/legitimacy/rounds/2026-08-27-legitimate-improvement/LEGITIMATE_IMPROVEMENT.md §8; src/evidence.py. Consumed by: item 72, whose exposure premise is stated relative to a baseline.