Feedback improves learning only when material observations and challenges can propagate into review of dependent models, positions, strategies, and actions.

An observation is not corrective merely because it is stored, retrieved, or described as surprising. It improves collective learning only when its materiality can be assessed, the claims and commitments that depend on it can be found, accountable reviewers reconsider those dependents, and any accepted correction can change later reasoning or action. Propagation should follow explicit semantic dependence and blast radius rather than textual similarity or a rule that every change travels globally. The claim assumes dependencies are represented well enough to route review, evaluators can distinguish material from cosmetic change, and consequences are observable often enough to test later improvement. It does not authorize automatic belief or identity mutation. Architecturally, it motivates root and parent lineage, exact-version dependencies, impact review tasks, preserved rationales, and separate proposal, approval, compilation, and activation stages.