Consequence-aware change evidence
Review a change through observed behavior plus reversibility of its real-world consequences.
These examples and illustrative results are independently authored teaching materials, not measured model results.
Use case
A bulk-mail fix removes duplicate recipients. The teaching dry-run has three addresses, two identical. Code can be rolled back, while sent messages generally cannot be recalled. Review evidence must address both behaviors and consequences.
Mechanism
Describe trigger and before/after behavior, using a no-send dry-run to verify recipient set and count. Identify affected users/entry points, duplication/omission risks and irreversible effects. State rollout gates, scope limits and stop conditions with test/configuration evidence rather than an author rating alone. Distinguish reverting code, stopping future sends and handling messages already sent.
Bad example
Bulk-send tests are green; any problem can be fixed by reverting one commit, so no other consequence needs review.
Good example
Review address deduplication: three teaching inputs include one duplicate, changing the plan from 3 sends to 2 unique recipients. Retain set comparison and tests, explain that code rollback cannot undo delivery, and define rollout scope and stop gates. Actual sending requires existing authorization.
Why the change matters
Change size does not establish consequence size. Before/after evidence supports improved behavior, while consequence analysis identifies recoverable effects. Together they inform review depth and rollout controls.
Observable expectation
The teaching plan changes 3 sends to 2 without losing either unique address. Add case/alias inputs and apply the product’s address rules instead of arbitrarily merging them. The record distinguishes zero external dry-run sends from checks still needed for real delivery.
Limits
A dry-run cannot establish production delivery or receipt. Address normalization, concurrent queues and retries add risks, and the author’s account remains reviewable. Recall support varies and must not be assumed; assess external effects under actual capabilities.