Model-cascade success concealed reads of feedback unavailable at decision time

A model router decides whether to answer cheaply, escalate to a stronger model or reuse a cache. If its decision stage can see feedback that should only arrive after that decision, an evaluation can report unrealistically strong answers while preserving ordinary-looking latency and billing. CascadeLedger treats this as an information-ordering violation instead of assuming a correct final answer proves an honest process.

The study binds stage calls to response provenance and compares counterfactual prefixes, withholding information not yet available at each decision point. In controlled injected faults, the researchers report near-perfect detection for the defined hidden-read categories, while also showing that some contaminated conclusions change when recomputed without unavailable feedback. These measurements depend on the study's own trusted logging and fault design; they cannot establish a universal detector for opaque commercial inference APIs.

The investigator must record when each piece of feedback was created, which stage was permitted to read it, and which cache or endpoint actually supplied it. Matching hashes alone would not expose a valid artifact introduced too early. The boundary is causal availability at decision time, not just integrity of bytes in a final report.