Engineering case study · Kiran Yenaganti
A targeted repair to HOMININI’s report verifier.
Prototype · AI-assisted engineering · Developer-reported local results
HOMININI explores how automated workflows can produce inspectable reports. During development, its verifier accepted appended statements including “all tests passed” and “no security vulnerabilities” without supporting evidence.
A report containing source references could still overstate what the workflow had established. The problem was in the verification path.
I repaired the verifier and added targeted negative tests for those unsupported additions. The reported regression results show that the tested additions are now rejected.
The repair supports a narrow conclusion: the verification path rejects the specific unsupported additions covered by those tests. It does not establish that every false claim will be detected.
One isolated run processed the public psf/requests repository through four workflow steps and generated a private report in 3.469 seconds. It made no model calls and no changes to the external repository. This is one observed run, not a general performance benchmark.
Six Python files from a pinned repository commit were inspected, their content hashes checked and source references included. The downloaded code was not executed.
A formatting limit that removed later findings was reproduced and corrected. A six-file regression test checks that every fixture’s source-linked finding survives report generation.
In an isolated application-level test, the submitting user could see the generated artifact and an unrelated user could not. Deployed multi-user security has not been established.
An earlier repair explained a recorded execution limit without restarting work. Its developer-reported 56-test suite covered the gateway and diagnosis helper. Those tests are separate from the report-verifier repair described above.
Runtime Accountability for Persistent LLM Multi Agent Systems is a working paper by Kiran Yenaganti proposing requirements for human authority, evidence, recovery and independent verification. It is not peer reviewed and does not claim empirical validation. A public reading link is not available on this site yet.
That working paper is distinct from the runtime-governance manuscript shared on ResearchGate.
Read the runtime-governance research overview →These isolated demonstrations do not establish production readiness, deployed multi-user security, independent-process verification or fully autonomous engineering. The timing above is a single observation, not a comparative benchmark.