Lecture 08: Verification Gates & Evidence

"Looks Good To Me" is Not Evidence

In unharnessed agent sessions, the agent evaluates its own output through language generation:

"I have reviewed the code and confirmed that the canonical tag is properly configured."

This is subjective hallucination. The agent cannot "see" PHP fatal errors or regex matching bugs through conversational reflection.

---

Deterministic Verification Gates

The Genisoft Harness requires verifiable execution evidence produced by real runtime processes:

               ┌───────────────────────────┐
               │    Modified PHP/JSON      │
               └─────────────┬─────────────┘
                             │
            ┌────────────────┴────────────────┐
            ▼                                 ▼
┌───────────────────────┐         ┌───────────────────────┐
│   verify-seo-safety   │         │schema-graph-validator │
│   - Insecure scheme   │         │ - Missing @type       │
│   - Unescaped meta    │         │ - Dangling @id        │
│   - Sitewide noindex  │         │ - Missing rich fields │
└───────────┬───────────┘         └───────────┬───────────┘
            │                                 │
            └────────────────┬────────────────┘
                             │
                             ▼
              [ EXIT CODE 0: VERIFIED PASS ]
                             │
                             ▼
              Evidence appended to progress.md

If the tool exits with non-zero status, the agent is blocked from committing and must self-correct.