What is running right now, read from each service by the indexer every few seconds, and what broke before. The vaults need none of this to keep their floors: the program runs on the chain, and anyone can crank it.
The policy: a post-mortem within 72 hours of resolution, every time, with the time, the impact, the cause and the fix. Each entry is a file in the public repository (docs/incidents), so the log can't be edited quietly.