The full, reproducible evidence behind the integration page. A real google-adk 2.8.0 agent, run locally and in-process with the Mission SDK — no Google Cloud — measured on multi-hop mesh authority, provenance-gated synthesis, and a governed action. Every comparison is classified — this is conformance, not a scorecard.
| Experiment | Native role | ReDevOps addition | Result | Classification |
|---|---|---|---|---|
| GO-A Transitive authority | hop runs its own capabilities | chain-wide intersection | 2/2 escalations denied · 0 wrong | REDEVOPS_DELTA |
| GO-A Provenance gate | coordinator trusts returns | attributable-only synthesis | 2/2 unprovenanced excluded | REDEVOPS_DELTA |
| GO-A In-scope preserved | — | authorized work admitted | 2/2 admitted | PARITY |
| GO-B Governed action | mesh produces a finding | authority · approval · replay | 0 unauthorized · 0 duplicates | REDEVOPS_DELTA |
| GO-C Real ADK mesh | ADK agent answers | provenance on every finding | 4/4 answered · 4/4 attributable | PARITY |
| GO-D No re-amplification | hop declares broad grants | chain intersection | dropped grant stays dropped | REDEVOPS_DELTA |
| GO-E Telemetry | ADK agent spans | Mission-root nesting | preserved (nested) | REDEVOPS_DELTA |
Real ADK agents form a delegation chain: a coordinator (grants research:read · web:fetch · kb:write) delegates to a web researcher that drops kb:write, which delegates to a scraper that re-declares kb:write. Effective authority is the intersection of the whole chain, so the scraper cannot regain the dropped grant. When it attempts to write the KB or delete, ReDevOps denies it — 2/2 transitive escalations, 0 wrong, while admitting 2/2 in-scope actions. Separately, a finding enters the coordinator's synthesis only when its provenance chain is attributable to authorized agents: a finding relayed by an unknown agent or with no provenance is excluded — 2/2 unprovenanced findings kept out (a raw mesh trusts both).
A delegated finding's shared-KB write runs only through a Mission. Missing grant ⇒ 0 unauthorized. Grant but no approval ⇒ parked, 0 writes. Approved ⇒ fires exactly once. Replay ⇒ 0 duplicate writes.
A real google-adk agent (via LiteLlm) answers four delegated research sub-questions end-to-end — 4/4 answered — and ReDevOps attaches a provenance chain to each so all 4/4 findings are attributable. ADK owns the agent loop; ReDevOps makes the mesh's output traceable without changing what it produces.
Each run writes a reproducibility record — framework version, model ids, per-finding classification, and a content-addressed result digest.
// reproducible · local, no Google Cloud · github.com/redevops-io/redevops-integrations/integrations/google · integration page · architecture article