# A guided tour of the AI evidence

Five stops, at your pace. Start at /ai-systems-evidence#walkthrough. This tour covers the frozen source-only and optional-selector evaluation; later repair waves and current runtime qualification are separate.

## 1. Goal: reach a useful conclusion

Choose a relevant case, understand its evidence and qualifications, and decide what to investigate next. Try: “What demonstrates production workflow leadership?”

Open /case-studies#choose-evidence. If the Evidence Guide is enabled in your environment, you can also ask it at /evidence-guide. Follow the source rather than treating the answer as the end of the investigation.

## 2. System: follow the evidence

The source-only path retrieves approved public material, composes reviewed wording and validates the answer. The optional model selects reviewed IDs; it does not author public facts. MCP exposes separate read-only search and fetch tools.

Inspect /ai-systems-evidence#architecture and /ai-systems-contracts.md. Download /evidence-guide-reference.zip to inspect or run the frozen core. Installation takes additional time and may require package-registry access.

## 3. Result: compare the same questions

Across 120 paired authored questions, source-only produced 96 useful first answers; the optional selector produced 95. Independent agent reviewers assessed anonymized answers, source support and requested requirements. Seven recoveries were outweighed by eight degradations.

Inspect /ai-systems-evidence#evaluations and /ai-systems-evaluation.json. These consumed questions remain regression material, not fresh future acceptance. Agent scoring does not establish human visitor outcomes.

## 4. Failure: inspect actual execution

The private tool-agent trial stopped after three empty searches. Fetch and final selection were not reached. Recovery behavior and useful completion are separate checks.

Open /ai-systems-evidence#traces. Compare actual execution with clearly labeled timeout and malformed-selection simulations. Follow /ai-systems-evidence#selector-post-mortem for the diagnosis and remaining questions.

## 5. Decision: retain complexity that creates value

Source-only remains the recommended default. The selector added provider cost and local latency while returning one fewer useful answer overall. A shared repair needs a new paired evaluation; live reliability and real visitor usefulness need separate validation.

Inspect /ai-systems-evidence#model-decision and /ai-systems-evidence#economics. Identify the evidence that would justify changing the decision.