Tune on 3 files before running on 200

I need to apply this task across [N] files or records: [TASK].

Do not start with all [N]. We go in phases, and each one has a gate you have to pass.

PHASE 1, sample. Pick 3 cases and tell me why those: one typical, one unusual, and the one you think is most likely to fail. Apply the task only to those 3 and show me the before and after of each.

PHASE 2, gate. Before continuing, answer:
- What came out different from what you expected in the 3.
- What rule you had to add or fix while doing them.
- What case you saw that the sample does not represent.
If you had to fix the rule, go back to phase 1 with 3 new cases. Do not move on with an untested rule.

PHASE 3, batch. Apply to the rest, but:
- Work in batches of [BATCH SIZE] and report at the end of each batch.
- Keep a count of the ones you skipped and why. A silent skip is a hidden error.
- If a batch has more than [THRESHOLD] unusual cases, stop and tell me instead of continuing.

PHASE 4, close. Give me the final count: processed, skipped, and modified differently than expected. If the total does not match [N], do not close: find the difference.

Hard rule: never report "done" on a batch you did not count.