Week 12: Archive or Delete?

Retention, digital decay, and what it takes to build a system that tells you the truth about which version is current

10 May 2027

Open the slides

Keeping everything is not a retention policy. It is the absence of one, and it has a specific cost: every version you keep is another candidate for being mistaken for the current one. Week 1’s filename existed because four files were kept and none of them was marked. Deleting is how a system stops lying about which version counts, and deleting the wrong thing is how it starts lying in the other direction.

Decay does not wait for a decision#

Files degrade without anyone touching them. Bit rot on failing media. A proprietary format whose only reader stops shipping. An external link that returns a different page than it did when it was cited. A cloud document that is a pointer, not a file, and resolves to nothing once the account lapses. Encryption whose key is no longer held. None of these are deletions. Each produces an artefact that is still listed in the directory, still opens or appears to, and no longer contains what it did.

The archival response is not storage. It is migration: periodically re-encoding holdings into currently readable formats, which is week 10’s conversion damage accepted deliberately, at a known cost, because the alternative is total loss later rather than partial loss now. Retention is the choice of which losses to take and when.

What a non-lying system requires#

Everything in this course has been a system telling a small untruth about which version was current. A filename asserting finality it could not enforce. A timestamp standing in for authority. A sync status standing in for agreement. A conversion reporting success while dropping content. So the requirements are the inverse of those failures:

• Currency is stated, not inferred. One version is marked current, by someone, on a ground they could name. Not the newest by default.
• History is recorded where the work happens. Not reconstructible from residue later: recorded at the time, as week 11’s archivists insisted.
• Losses are logged. A conversion, a migration, a deletion each leave a record of what was dropped, so week 10’s silent gaps become visible ones.
• Deletion is a decision with an author. Not a cleanup, not a quota eviction, not the disk filling up.

A version-control system does most of this, which is why it works. It also does not solve the problem, because the problem was never the tooling. The four files called final were produced by people who had access to version-control software and did not reach for it, because the naming workaround was cheaper at the moment of the save and the cost arrived later, distributed across everyone who had to guess.

Where this leaves you#

The course opened by claiming digital work is rarely one file: that it is a family of near-duplicates competing to become the truth. Twelve weeks later the claim holds, and the reason is now specific: nothing in the ordinary handling of files records which one won, or why. Deciding is still human work. What a well-built system can do is refuse to pretend it has already happened.