diff --git a/.macroscope/check-run-agents/effect-service-conventions.md b/.macroscope/check-run-agents/effect-service-conventions.md
index b49cda95ebe..1bbd8192cb5 100644
--- a/.macroscope/check-run-agents/effect-service-conventions.md
+++ b/.macroscope/check-run-agents/effect-service-conventions.md
@@ -46,11 +46,12 @@ Review changed TypeScript and directly affected call sites for the conventions b
 ## Errors and predicates
 
 - Define service failures with `Schema.TaggedErrorClass` and structured attributes. Derive `message` from those attributes rather than storing an unstructured message as the only data.
-- `Schema.Defect()` is not a substitute for modeling an error. Flag error classes whose only meaningful field is `cause`, or whose `message` merely stringifies an opaque cause.
+- `Schema.Defect()` is not a substitute for modeling a generic error: its tag, fields, or both must identify the failure structurally, and its `message` must not merely stringify an opaque cause. A semantically precise error tag may preserve a real `cause` without inventing a redundant singleton field when no additional variable context exists; still retain any real path, resource, request, or entity context available at the wrapping site.
 - Capture stable, serializable domain context such as the operation or stage, resource/path or entity identifier, and normalized category/status. Map failures where that context is known instead of wrapping an entire multi-step pipeline in one generic error. Do not add a `detail` field that merely copies `cause.message` and then use it to construct the wrapper message.
 - When translating or wrapping a real failure, preserve the immediate underlying error itself as `cause` alongside the structural fields so the complete error chain and stack remain available. If every construction wraps a failure, `cause` should be required; make it optional only when the same error can legitimately originate without an underlying failure.
 - Derive the wrapper's `message` exclusively from its stable structural attributes, never from `cause`, `cause.message`, or a stringified defect. Do not replace the immediate error with only `error.cause`, erase a structured upstream error into a string, or manufacture an `Error` merely to populate `cause`. Pure validation/domain errors created without an underlying failure do not need a cause.
 - Do not encode the same distinction twice with both a specific error tag and a single-value `operation`, `reason`, `kind`, or `phase` literal. Choose one coherent model: use distinct error classes and omit the redundant discriminator when callers or messages treat the failures as genuinely different, or use one service-level error with a multi-value operation discriminator and a generic message derived from that operation when the failures share the same semantics.
+- Treat an error message exposed through an HTTP/RPC response, persisted state, UI, or another caller-visible boundary as behavior. Preserve those messages during a structural refactor. Existing distinct caller-visible messages are evidence that the failures should normally remain distinct error tags without redundant singleton discriminators, rather than being collapsed into a generic operation error.
 - Split semantically distinct failures into separate error classes when a `reason`, `kind`, `phase`, or similar discriminator is used to choose the user-facing message or drive caller control flow. A discriminator used only for internal diagnostics may remain a field.
 - Use `Schema.Union` of error classes when a shared schema, predicate, or helper type is useful.
 - Export direct schema predicates such as `export const isFoo = Schema.is(Foo)`. Flag a private `Schema.is` constant wrapped by a redundant function with the same signature.