9. What existing methods establish · Convergence Programming

Each of these is useful. Each is individually insufficient, and for the same underlying reason: each leaves some distinction outside the observation history uncertified.

Remedy	What it leaves open
Prompt engineering	Uses only information that entered the observation history. It cannot recover what never did.
Polished demos	Answer usability. Hidden critical scenarios remain untested.
Self-reflection	Without independent observation it collapses into correlated self-report.
Model-as-judge	Improves explanations. Shares the blind spots of the thing it judges.
Generated tests	Preserve known anchors, and make false convergence more convincing when generated from the same underspecified prompt.
Retrieval	Adds documents rather than executing the decisive scenario.
Human approval	Decisive when the inspector can read the implementation, weak otherwise.
Formal verification	Certifies against a specification and inherits every omitted requirement.
Runtime monitoring	Catches what has already happened, which for irreversible action is too late.

None of this says stop doing them. It says that each one certifies a distinction only when that distinction has entered the trajectory, and the harness is the part that makes sure it does.