sema-evalsIndependent evaluations

A2A Drift — 20260722T080212387Z-order-20260714

Validates the A2A semantic-extension middleware and two-agent demo: Agent Card extension advertisement, acceptance-contract message parts, controlled cross-agent registry drift, silent execution under baseline, voluntary detection, enforced halt, the no-drift false-halt guard, condition pairing, and bundle/summary reproduction. Scripted-agent outcomes are a construction, not evidence about language models, and not conformance evidence against a real A2A SDK (ADR 0012).

Condition Trials Drift trials Detected Silent executions Correct halts False halts Task successes
baseline 8 4 0/4 4/4 0/4 0/8 4/8
advertised-voluntary 8 4 4/4 0/4 0/4 0/8 4/8
advertised-enforced 8 4 4/4 0/4 4/4 0/8 8/8

Drift-scoped columns use drift-injected trials as the denominator; the rest use all trials in the condition. Every count is recomputed from trials.public.jsonl at build time.