Findings · Agent control
Over 10,000 injected events, a monitoring arm recorded 1,026 actions as prevented and all 1,026 executed; an in-series arm consuming the same events, evaluator and policy recorded 6 of 9,880.
Evidence
Two arms consumed the same events, with the same evaluator, the same policy and the same injected faults. The only difference was position: one sat on the path and could refuse, the other observed alongside.
The observing arm produced a dashboard reporting that 1,026 actions had been prevented. Every one of them had executed. This is not a degraded detection rate with a tail of misses; it is the whole population, and it follows from what a monitor is. A component that records decisions without standing in the path of the action is recording intentions, and a record of intentions reads exactly like a record of outcomes.
A second result in the same harness concerns who writes the record rather than where the control sits. When telemetry and enforcement come from one writer, the same run reports perfect enforcement, no divergences, and six fewer unsafe executions than actually occurred. A system whose account of itself has a single author cannot detect its own enforcement failures.
Sample. 10,000 events per arm, identical events, identical evaluator, identical policy and identical injected faults across both arms; the arms differ only in enforcement position.
Boundary
A property of one synthetic, deterministic testbed under injected faults, not a measurement of any deployed safeguards product and not a rate that transfers outside this harness. The agent environment is deliberately trivial so that what is measured is the pipeline rather than model behaviour.
Independent reproduction. None known.
Reproduction
./demo.sh && make experiments- Toolchain
- Python, Docker Compose, Redis; runs offline on one machine
- Expected output
- The monitor arm reports 1,026 of 10,000 as prevented-but-executed and 0 of 818 intervention recall; the in-series arm reports 6 of 9,880 and 697 of 818
- Claim stated at
- repowazdogz-droid/safeguards-control-plane · README.md, section The result, in one table
- Verified at
- repowazdogz-droid/safeguards-control-plane@08e52549 (2026-08-12)
This pass. Re-verified this pass by source read at commit 08e52549; the measured run and unit tests are in the repository and were not re-executed in this pass.