Read Code Moves Faster Than Understanding
Attention is the amount of human scrutiny required before an artifact should be trusted, approved, merged, published, or acted on.
It is not the severity of any single finding. It is not the verdict. A high-attention artifact can still be LGTM after careful review. A low-attention artifact can still request changes if a clear issue is found.
Attention should be based on blast radius, reversibility, observability, verification difficulty, evidence quality, and whether the artifact changes future engineering understanding.
When evidence is insufficient to show that a change is bounded, reversible, or non-semantic, choose the higher attention level that matches the possible blast radius. Do not downgrade solely because CI passes, the diff is small, the author claims no behavior change, the artifact is docs-only, or the change was generated.
minimal: A quick correctness scan is enough. Use for non-semantic or mechanically verifiable changes: typos, formatting, dead links, renames, generated-file refreshes, test-message cleanup, or wording changes that preserve meaning.
routine: A routine semantic review is enough. Use when the artifact changes behavior but is bounded, familiar, easy to understand end to end, supported by direct evidence, and cheap to reverse. Examples include a localized bug fix with a regression test, a small feature implemented through an established pattern, or a simple validation change confined to one component. It should not alter contracts, shared invariants, sensitive flows, or rollout mechanics.
focused: Focused human review is required. Use when establishing trust requires reasoning about intent, edge cases, domain rules, state, cross-component interactions, internal contracts, operational effects, or claims not directly demonstrated by tests. The blast radius may still be limited, but the reviewer cannot validate the artifact through a local scan alone.
owner: Careful review by a responsible owner or domain specialist is required. Use when incorrectness could be widespread, difficult to detect, difficult to reverse, or materially consequential: security, privacy, money, data integrity, public contracts, schemas, migrations, backfills, concurrency, idempotency, state machines, rollout safety, or large agentic changes with weak evidence. This also applies to source-of-truth documents whose semantic drift could misdirect future decisions.