Revision to Jules's Bottle-0 capsule in 3k (task 8811d91d-5012-4f1e-a3a8-709565fa9d91, source revision bottle-0). Prior exposure: I read the full source mirrored in the task body, so this is a self-check, not a blind reader test.

Problem with 3k: it states "word count: exactly 60 whitespace-separated tokens", but splitting the quoted capsule on whitespace gives 61. I counted each run of spaces as one separator; commas, periods and hyphens stay inside the token (ledger-07.csv = 1).

The overage is small and fixable before the identifiers line:

• keep: restart needs Mira's approval, not yet given.
• drop: the word yet

Revised capsule (60 whitespace tokens):

"Aster import paused; restart needs Mira's approval, not given. Active file ledger-07.csv, 120 rows, excluding header; ledger-06.csv obsolete. Duplicate invoice ID may be installments; do not delete. Amounts: decimal EUR strings, not cents; unknown rejected, not zeroed. Account IDs are strings; never convert to integers. Last run was dry, not committed. Next: ask Mira for installment policy; keep source unchanged."

Six answers still reconstructable from the capsule:

• input: ledger-07.csv; 120 data rows excluding header.
• duplicate: No; may be installments; do not delete.
• money: decimal EUR strings, not cents; unknown stays rejected, not zeroed.
• identifiers: No; never convert to integers.
• commit: dry run only; paused; restart not approved.
• next: ask Mira for installment policy; keep source unchanged; Mira can approve but has not.

Tradeoffs and limits: "not given" is blunter than "not yet given"; nothing tells the reader unknown is a literal string rather than a placeholder — a reader must infer that from the money answer. The original baseline capsule in the task body is 64 tokens by the same count, so it is also over budget. This is a hand count, not a test run by another participant.