babel-relay

Babel Relay measures what happens when the meaning of a shared term silently drifts while work passes through a relay of agents — and which mechanism, if any, surfaces the drift before it becomes a wrong outcome.

How to read the results

Each trial sends a small contract definition through a three-hop relay (specification → implementation → audit). In drift scenarios, one field of the definition is silently mutated at a relay boundary; the audit agent must then decide whether the implementation matches the definition it can see. The primary endpoints are silent divergence (drift that no agent ever surfaced) and task success, scored by a frozen deterministic parser — no LLM judge.

The five conditions decompose where protection could come from, so that a content effect is never attributed to hashing, and a detection effect is never attributed to enforcement.

baseline

No semantic material beyond the task itself. Measures the floor.

equal-prose

The full definitions travel inline as prose. Isolates the value of content alone.

opaque-resolver

The same definitions behind a content-free lookup ID. Controls for compact references without content addressing.

addressed-voluntary

Content-addressed references; agents may check them but nothing compels action. Isolates detection.

addressed-enforced

Content-addressed references plus a runtime that refuses to proceed while reference and definition disagree. Isolates enforcement.

Runs are labelled by evidence tier — deterministic harness validation, exploratory model pilot, or preregistered experiment — and nothing stronger than its label should be inferred from any table.