COUNTERPHASE / LOCAL SCIENTIFIC ADVERSARY

What did the model learn from evidence it never selected?

Two constrained learning policies. One complete original opportunity ledger. Unselected future outcomes are changed one at a time. A changed decision is a reproducible counterexample.

Cohort SYNTHETIC_BUDGET_20 · 20 T0 opportunities · Offline

Adaptive policy

NO COUNTEREXAMPLE WITHIN TESTED CHALLENGES

34 challenges run · 0 not tested

Fixed policy

NO COUNTEREXAMPLE WITHIN TESTED CHALLENGES

34 challenges run · 0 not tested

Evidence boundary

A passing challenge does not prove the policy never accessed additional labels. Supplied timestamps and selection flags are not independently authenticated. A failing challenge demonstrates dependence on the altered inputs under this local runner, not the causal cause of real-world losses.

Source SHA-256: e7f296f1f8dd9ce9f12cfec515582d50a3d05aab73ae0b97b535b88556142dc4
Policy SHA-256: 8f3cf40cfca5813983b9e9cbaffb9817f3249587db8e3164549c4c7285d7b070