Make the review specific.
Bind declared tools, policy and scenario inputs to the result, so a reviewer knows exactly what was evaluated.
Evidence for agent decisions
Test the policy.
Inspect the evidence.
Run a selected agent policy against fixed synthetic scenarios. Deterministic code grades each decision. Your reviewer gets tamper-evident evidence they can re-check in their own browser.
AI-generated illustration · fictional people and setting
Synthetic demo · runs in your browser. Not production SaaS or compliance certification. Evaluate a selected policy in your browser using synthetic scenarios. Origin does not contact or execute your named agent. Browser evidence is untrusted by default.
01 / The reference check
From an assumption
to something checkable.
A clear sequence for a complicated review. Try the five stages, then run the synthetic policy evaluation yourself.
What inputs were declared?
Bind Declared inputs are bound
Declared model, tool set, selected policy, and scenario battery are pinned into a single input digest. This binds the local policy evaluation; it does not prove that the named agent was inspected or run.
input · declared model + tools + policy + battery DIGEST BOUND
Which decisions were tested?
Challenge A deterministic battery runs
The same scenarios every time: actions that must be allowed, actions that must be denied, and actions that must escalate to a named human.
battery · allow · deny · escalate DETERMINISTIC
Where did the policy disagree with the oracle?
Grade The oracle grades it
False accepts, false rejects, and catastrophic failures produce scored outcomes under the fixed oracle. A catastrophic over-grant caps the result at L0 no matter how well everything else scored.
oracle · FAR · FRR · catastrophic cap EVALUATION RESULT
What does the artifact actually attest?
Attest Browser demo evidence is issued
The browser signs a synthetic Action/Run envelope over the local policy evaluation. It records not attempted: no named agent was contacted or executed, and no provider effect was confirmed.
envelope · bound to declared input · synthetic SIGNED
Can a reviewer detect a change?
Reverify Offline UNTRUSTED, until tampered
A reviewer re-verifies browser evidence offline and gets UNTRUSTED: its session signer is unpinned and it does not represent execution. Change one bound field and the verifier returns VOID.
verify · intact browser evidence UNTRUSTED
verify · one bound field changed VOID
A deterministic oracle grades the selected policy. The demo does not run the named agent or grant permission to deploy it.
Run the reference check02 / Why Origin
A powerful model.
A different question.
Your model provider helps make an agent capable. Your reviewer needs to know what a specific policy allows, where it fails, and what the evidence actually covers.
Where Origin fitsMake the limits visible.
Show allowed, denied and escalated cases. Keep synthetic results distinct from customer evidence and deployment authority.
Make it independently checkable.
Export an artifact that a reviewer can inspect and re-verify. A changed bound field invalidates its integrity.
03 / The evidence
Trust grows from
what you can verify.
Inspect the record, its source and its limits. Each evidence level says what has been earned—and what is still ahead.
Available · sandbox
Follow a machine-emitted trace.
TR-A002 · 12 events, a real SHA-256 hash chain and an offline re-verification command.
Available · authoredUnderstand the evidence format.
TR-A001 · an authored specimen showing the record a reviewer could receive.
Not earned yet
A real customer's workflow.
TR-A003 · external design-partner evidence has not been collected. No customer validation is claimed.
For the Origin Attestation format, VALID means intact and signed by the embedded key — not signer identity, execution, issuer trust, safety or approval.
Verify an attestation04 / Work with us
Start with one
decision that matters.
We're looking for design partners running agents with production access. Bring one approval decision; we'll map the evidence your reviewer would need to decide. Founder-led, four steps, and a pilot only if there's a useful fit.
A focused first conversation
Bring the workflow.
And the approval question.
- 01 Map the action, tools and potential consequences.
- 02 Agree what the reviewer needs to see.
- 03 Identify the tests and evidence gaps.
- 04 Scope a pilot only if there is a useful fit.
The public demo is a prototype, not production SaaS or compliance certification. Customer-specific execution and runtime enforcement require separately agreed work.
05 / Built around the reviewer
Evidence, with
the boundaries included.
Capability, permission, execution and verification are different questions. A useful record keeps them separate.
Deterministic evaluation
A defined oracle grades the policy against a synthetic battery. Never an LLM grading another LLM.
Visible provenance
Authored, synthetic, machine-emitted and customer-owned evidence stay distinct. Unpinned browser signers stay UNTRUSTED.
Your authority to decide
The artifact supports review. It does not certify an agent, confirm an external effect or approve a deployment.
What Origin is not.
Not yet validated by a customer
No design partner, no pilot, no paying user — today. Everything shown here is authored, synthetic, or machine-emitted in a sandbox; none of it is customer-owned. If a real workflow from a real reviewer is earned, it will appear on /proof, labelled as plainly as everything else.
Not production SaaS
A working prototype on synthetic sandbox evidence. Customer-specific execution requires separately agreed work.
Not a compliance certification
Origin's checks produce evidence for your reviewer; your gate decides. The public demo's evidence is session-signed and unpinned, so it is not Origin-issued. Nothing here certifies anything.
Not “safe” or “correct”
A green result means the check reproduces under this verifier: integrity and determinism, never “safe” or “correct”. That is the whole claim.
Not runtime enforcement
The runtime gate, controlled proxy and kill switch are proposed pilot architecture, not shipped.
Who builds Origin.
Bo-Huei Lin, founder.
Origin Labs / Bounded simulations
The next decision
may move something.
The agent reference check is the product prototype. Labs explores the same evidence questions through synthetic fleet operations and simulated robot plans: what was tested, under which conditions, and what should happen next?
Simulation evidence is a step toward understanding a system. It is not real-world safety validation or robot certification.
Explore Origin LabsLet's make the next step clear
Bring us one approval decision.
If your team can't ship an agent because no one can show what its policy allows, see how a synthetic reference check works — or book a founder-led evidence review.