Pre-execution evidence gate
Verify what the agent can do before it does it.
Turn a proposed action into explicit gates, hard stops, and a reusable decision record. No wallet connection. No invented proof.
Two required gates still need independent evidence.
Define the blast radius
Challenge every required claim
No active hard stop
The action remains in review because evidence is incomplete, not because a policy boundary failed.
Evidence is incomplete, but recoverable.
The action is reversible and human-gated. Independent destination evidence is the shortest path to a stronger decision.
Compare the evidence, not the confidence theater.
Save the current review locally, restore it later, and inspect why the decision changed.
No saved reviews yet.
A review draft records claims before consequential execution
Read the explanation
The saved review describes rebalancing eight percent of treasury stablecoins with seventy five thousand dollars exposure, human approval and a thirty minute rollback window. If the seventy five thousand represents that eight percent, the implied treasury is nine hundred thirty seven thousand five hundred dollars. The page does not state this equality, so this is a conditional arithmetic example. At one pixel per four thousand dollars, exposure measures eighteen point seven five and implied treasury two hundred thirty four point three seven five; the remaining eight hundred sixty two thousand five hundred measures two hundred fifteen point six two five. No wallet or market execution occurs. The initial saved markup says three of five claims are verified. Three divided by five is sixty percent, while the displayed evidence score is sixty two out of one hundred. At fifty pixels per count, three verified claims measure one hundred fifty, two remaining one hundred and total five two hundred fifty. The two scores need not mean the same thing, and the local application script is absent, so no weighting formula can be inferred. Authorization and destination integrity are described as needing stronger evidence. The visible text is an initial review draft, not independent proof that any financial action is safe. A hypothetical reviewer validating the two remaining claims would move a five-claim checklist from three to five. At sixty pixels per claim the before bar is one hundred eighty, additional two one hundred twenty and total five three hundred. That arithmetic does not establish the actual decision score or unlock execution. The page advertises local scenario save, review brief copy, JSON export and reset, but no application code accompanies this saved snapshot. Native fields can still be edited, while these actions lack handlers offline. Preserve the original tool and place the explainer before its first custom footer, without inventing a functioning ledger, cryptographic attestation or backend approval.