ARABATripwire · Watch Mode
Catch the silent wrong action before a customer does.
An agent's worst failure is the one that returns ok. A refund that fires at ten times its cap. A record silently overwritten. A call to the wrong endpoint that succeeds. The run looks green, and the customer finds the bug. Tripwire is the read-only scan that catches it first.
The scan
Say what must be true. Watch it flag the rest.
Point Tripwire at your agent and declare the rules that must always hold. It runs the agent against your inputs and a library of adversarial ones, then flags every action that completed but broke a rule. It changes nothing, so it sits beside a live system today.
The same report shape the package prints. A live, interactive scan runs at deadlatch.dev.
The stack
Three primitives. One control loop.
enforce
Purse
Decides what an agent may do and refuses the rest in real time. The action never fires off-policy.
See Purse →prove
blackbox
Hash-chains what the agent actually did, so the record survives anyone trying to rewrite it after the fact.
See blackbox →watch
Tripwire
Runs the agent against your rules and adversarial inputs, and flags the action that completed but broke one.
You are hereEnforce stops the wrong action. Prove records what happened. Watch catches what still slipped through. Together they are Deadlatch.
Under the hood
Why a green run is not enough.
It checks the result, not the exit code
An action that returns ok can still be wrong. Tripwire tests the outcome against the rules you declared, so the ok-but-wrong result is caught instead of shipped.
It brings its own adversaries
Beyond your inputs, it runs a library of adversarial ones designed to trip an agent. The failure you did not think to test for surfaces before a user finds it.
Read-only by design
It watches and reports. It changes nothing and needs no write access, so you put it beside a live agent today without touching how the agent runs.
Zero dependencies
A small library that takes a run and a set of expectations and returns a report. It reads a blackbox trace directly, or scans a run of its own.
Run it yourself