Everything before this verified a fact at the moment you asked, and then the verification evaporated into a sentence nobody could re-check. This release turns a verified fact into something you can hand to someone else.
Two new tools
evidence_bundle packages FTSO Scaling anchor feeds and finalized FDC attestations (Web2Json included) into one portable object carrying each fact's Merkle proof, voting round, protocol id and the Relay address. Every fact is folded against the on-chain root before inclusion — a fact that fails is refused rather than flagged, and a Web2Json fact without expect_source is refused too, since an unbound proof attests to whatever URL sits inside it.
evidence_verify is the receiving end. It rebuilds every leaf from its own body, folds it through its proof, and compares against Relay.merkleRoots(protocol, round) read on-chain at verification time. Nothing in the bundle is trusted, including its own hash. The verifier needs an RPC endpoint and nothing else — not this server, not the DA layer, not the agent that produced the bundle.
The hash is not the security boundary
bundle_hash is tamper-evidence, and it is documented as exactly that. Verified live on mainnet: an attacker who triples a price and recomputes the hash still fails verification, because the proof no longer folds to a root Flare's validators signed.
```
verified: false | bundle_hash_intact: true | facts_failed: 1
reason: the proof does not fold to the on-chain FTSO Scaling root
```
The hash catches accidents. Consensus catches attackers.
What a bundle does not prove
Not that the agent was honest about its reasoning, not that these were the only facts it saw, not that its conclusion follows. Only that these specific facts were true, in these rounds. That is a narrower claim than "audit trail", and it is a different claim from attesting that a server executed correctly — which still requires trusting the operator or their hardware.
New prompt prove_what_you_acted_on walks through building a bundle, verifying it, then forging it two ways and watching both get caught.
23 tools, 6 prompts, 2 resources, 103 tests.