Prompt-policy comparator
79
Risky proposal-leakage rows.
Reported in the prompt-policy comparator path. This is a proposal endpoint, not an executed-side-effect measure.
SPQR / runtime evaluation
Proofs, not promises.
Read the limits too.
The reported Aegis result from Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution.
Reported sandbox evaluation
What was measured.
Governed sandbox path
0 / 2,100
Risky side-effect completions.
Zero observed in the evaluated governed rows. The same path also recorded zero governed mock-tool applications.
Attempted governed decisions
1,832 / 1,832
Trusted provenance retained.
All attempted governed decisions preserved runtime-resolved provenance in this evaluation.
The result is bounded. The next test counts useful work.
6,300 total rows across 42 tasks, five run families, three conditions and ten repeats per family. Different endpoints must not be presented as a percentage reduction. These results do not establish useful allowed-task completion, universal AI safety or production bypass resistance.
Read the methods and limitations · Review the next evaluation protocol
Study context
Keep the denominator and the endpoint.
The runtime-governance manuscript evaluates five run families, 42 tasks, three conditions and ten repeats per family. That totals 6,300 rows, including 2,100 governed rows. Public source: the runtime-governance preprint linked below.
Comparator endpoint
79 risky prompt-policy comparator-path leakage rows. Leakage refers to risky proposals in the comparator, not the same endpoint as a governed side-effect completion.
Governed endpoints
Zero governed risky side-effect completions and zero governed mock-tool applications. Trusted runtime-resolved provenance was retained for all 1,832 attempted governed decisions.
The public source is a preprint. This page does not assert journal acceptance or independent replication.
What this result does not establish
Zero observed events is not zero risk.
Not universal safety
The result concerns the evaluated sandbox corpus and governed path. It does not prove the model is intrinsically safe.
Not useful allowed execution
Zero governed mock-tool applications does not show legitimate work completing. A stronger comparison must measure utility and false denials.
Not deployment assurance
Direct credentials, bypass paths, tool vulnerabilities, malicious policy changes and recovery behaviour require separate review and testing.
No claim is made that this evaluation would have prevented a particular public incident. Proposal leakage and executed effects must not be collapsed into a “100% reduction” chart.
Read examples as examples
A diagram is not a runtime receipt.
The website’s fictional tenant, policy identifiers, verdicts, sample timings and receipt graphics illustrate the control pattern. The interactive preview is scripted. They are separate from the reported sandbox evaluation above.
One clear next step
A useful challenge asks:
what would change the result?
Keep the existing safeguards and define what legitimate work, prohibited effects and acceptable cost mean before the run.
One workflow. Defined scope. A decision about fit.