CASE / NEW VERIFICATION

Can Table 2 be reproduced?

A Google ADK coordinator runs five scoped agents that map the claim, reproduce it, diagnose drift, repair the configuration, and challenge the conclusion.

Ready to verify

Source anchor Table 2 · FAST-SENT · test column

Reported macro F1 0.942

Tolerance ±0.002

SYNTHETIC FIXTURE orchestration benchmark

EXECUTABLE CONTROL-PLANE EVIDENCE

Fortified controls

Not evaluated

Four independently inspectable controls. Values appear only after the server returns executable evidence.

C1 WAITING

Agent registry

— versioned agent cards

Evidence: Not requested

C2 WAITING

Least privilege

— deny + quarantine decisions

Evidence: Not requested

C3 WAITING

Crash recovery

— lease-expiry recovery semantics

Evidence: Not requested

C4 WAITING

Authoritative context

— effective-time context lookup

Evidence: Not requested

RECORDED ADK EXECUTION TRACE

Five specialist events + one coordinator event.

00:00.0

SELECTED EVIDENCE

Verification contract

UNRUN

Run the case to populate exact configs, metric deltas, execution fingerprints, and agent decisions. Every status must be backed by a tool result.