Mechanical Gates over Model Consensus
Evidence-Governed Harness Engineering for Agentic Software Development
Vibe coding creates applications; harness engineering makes them governable.
For regulated teams that need evidence, auditability, and rollback.
AI can build software faster than we can govern it
Code opacity
AI-generated native code is often a black box — unreadable to the business operators and IT reviewers expected to audit and maintain it.
AI-verifier risk
When the AI that verifies shares the producer's context or bias, "all tests green" can verify nothing — a correlated shared-context false pass.
Regulated audit gap
In banking and government, internal audit and control cannot govern what they cannot read — turning AI-built apps into unmanaged shadow IT.
Govern with evidence, not consensus
Mechanical gates over model consensus
At release gates, privilege executable evidence — mutation testing, property-based testing, spec invariants, real execution — over "several models agreed."
Shared-context false pass hypothesis
A named and situated failure mode: the AI verifying the AI can pass tests while the specification is never truly checked. Not solved by adding more models.
Skill capitalization
Each delivered project becomes a reusable, testable, governable Skill with a named-state lifecycle — so quality compounds instead of decaying.
Regulated teams govern what they can see
In banking and public-sector settings under internal-audit and internal-control mandates, business units increasingly build their own applications by vibe coding — but the native code is opaque to the very functions expected to review it. Re-expressing the repeatable, judgment-bearing steps as inspectable workflow graphs lets governance attach at the workflow layer, and — the property auditors valued most — a non-engineer auditor can visually follow a transaction directly on the graph.
Preliminary field observation — not a validated comparisonFrom opaque code to an audit trail
Creation stays fast; the repeatable, judgment-bearing steps become controllable and auditable.
Working Paper
Conceptual framework + preliminary field observation
Controlled studies (EXP-001 / a comparative case study) are planned, not yet reported.
This is not peer-reviewed and contains no validated empirical results yet. Produced with a multi-engine AI workflow assisting drafting, critique, and verification; the named human author set the research questions, reviewed the evidence, and accepts responsibility — AI systems are not authors. A DOI certifies permanent archiving, not peer review.
Cite this work
Lu, Y.-H. (2026). Mechanical Gates over Model Consensus:
Evidence-Governed Harness Engineering for Agentic Software
Development (v3.4, Working Paper). Zenodo. DOI: 10.5281/zenodo.21390753.
ORCID: 0009-0006-5373-0586. License: CC BY 4.0.