AI-extracted understanding is a hypothesis, not documentation

When AI extracts understanding from existing code, the output looks like documentation. It is not. It is a hypothesis.

Forward context creation has a clear source of truth: the specification. If the code does not match the spec, the code is wrong. Validation is straightforward.

Backward context creation inverts this. The code is the source of truth. AI analyses it, produces a structured interpretation. But if that interpretation does not match what the code actually does in production, the interpretation is wrong. Not the code.

This is a different discipline. Validation means checking whether an interpretation matches observed reality. Confidence is bounded by what the tooling can see and what domain experts can confirm. Re-validation is structurally required, because the estate keeps changing underneath the model.

Most organisations have no process for this. They have documentation practices, design reviews, testing standards. All built for the forward direction. When they apply those same processes to backward context creation, they produce artefacts that look complete but are not attested against reality.

The result is confidence theatre. Governance that looks evidence-based while the evidence is incomplete in structurally predictable ways.