Ask an agent what it did and it will tell you a story. The story is usually plausible. It is written by the same system that did the work, in the same confident register whether the work happened or not. If you have spent time with agents, you have caught one narrating a success that did not occur: a file that was not written, an email that was not sent, a number that came from nowhere. The industry's answer so far has been screenshots: pictures of the agent appearing to do things. A screenshot is a story with production values.
We think delegation runs on a different kind of artifact. When a contractor finishes a job, you don't accept a paragraph about how it went. You get an invoice with line items, and behind it a paper trail you could audit if you had to. Centra ends every run the same way: with a receipt.
Anatomy of a receipt
A Centra receipt is a signed record with five load-bearing parts. Inputs: what the run was given (the task, the parameters, the standing instructions in force). Outputs: what it produced, as artifacts, not summaries of artifacts. Spend: what the run cost, in compute and in money, itemized. A replay hash: a fingerprint of the run's deterministic execution, which we will come back to. And a signature over all of it, so the record cannot be quietly edited after the fact.
Notice what's absent: the agent's opinion of its own work. A receipt doesn't say the run went well. It says what went in, what came out, and what it cost. Judgment is left to the reader, who finally has the material to exercise it.
Replayability is the load-bearing property
The replay hash is the part people ask about, and it's the part everything else hangs on. Centra agents execute consequential work through CentraScript, our deterministic execution language. A task compiles to a plan, the plan is walked step by step, and the walk is deterministic: same compiled plan, same inputs, same walk, byte for byte. The replay hash fingerprints that walk.
This is what upgrades a receipt from a claim to a proof. A log says "trust that we wrote this down accurately." A replay hash says "run it again and check." If the recorded walk and the replayed walk produce the same hash, the receipt describes what actually happened not what anyone, human or model, says happened. Determinism is not a performance feature for us. It is an accountability feature that happens to run fast.
Verification without an account
A Centra receipt can be verified by anyone, with no account and no relationship to us. That's a deliberate constraint, and it shapes the design. If verification required logging into our platform, then "proof" would really mean "trust Sidiora Labs." The whole point is that you shouldn't have to. Your accountant should be able to check a receipt. So should a counterparty who has never heard of Centra and has no reason to believe us about anything.
It also keeps us honest. A verification path that works without our cooperation is a verification path we can't degrade when it becomes inconvenient. We built the door with no lock on our side on purpose.
What this unlocks
Here is the practical payoff. Today, delegating to an agent means absorbing its risk personally if it misfires, "the AI did it" is your explanation, and it's a bad one. A receipt changes the shape of that conversation. When your boss asks what the agent did with the budget, you hand over a signed record of inputs, outputs, and spend. When an auditor asks how a number was produced, you hand over a replayable trace. When a regulator asks whether the process you described is the process that ran, the replay hash answers before you do.
Delegation you can defend is a different product from delegation you have to vouch for. Screenshots ask for your trust. Receipts make trust unnecessary which, we've found, is the fastest way to earn it.