INTERACTIVE DEMO

Follow an AI action.
Inspect the evidence.

See how a consequential action moves from agent context and policy to human review, a downstream system response, and independently verifiable evidence.

Real SHA-256 + Ed25519 checks Runs locally in your browser Independent Python verifier
LOOPGRID / EVIDENCE LABLOCAL SIMULATION
One action.
Six connected records.
01Agent contextIdentity · authority
02Policy & reviewVersion · approval
03Tool actionRequest · response
04Evidence proofHashes · signatures
CAPTURE → SEAL → INVESTIGATE → PROVE
Clean-room synthetic demo. Fictional identities and data, scripted agent decisions, simulated reviewers, and an in-memory business system. No real calls, payments, patient records, loan decisions or external business APIs. What is real and what is simulated →
01 / CHOOSE A WORKFLOW

Different industries.
One evidence model.

Choose a scenario, inspect the policy, and decide whether the synthetic action should proceed. Every workflow uses the same reference cryptographic recorder.

VOICE AI · FINANCIAL ACTION

A customer disputes a duplicate charge.

A scripted voice agent requests a refund.

◎ LOCAL / SYNTHETIC
EVIDENCE RUN
Loading sampleLoading
0 events
Workspace Recorder synthetic.v1

Loading the signed reference history.

SIGNED EVENT HISTORY SELECT TO INSPECT
EVENT INSPECTOR PROTECTED PAYLOAD

No event selected

{}
Inspect hashes & signature
THE IMPORTANT DISTINCTION

A verified record and a successful business action are separate claims. Reject the request or simulate a downstream failure to see the difference.

02 / INDEPENDENT VERIFICATION

Don't trust the badge.
Check the bytes.

The browser recomputes commitments and verifies actual Ed25519 signatures. The trusted signer fingerprint is supplied separately from the evidence being inspected.

CRYPTOGRAPHIC EVIDENCE VERDICTNOT CHECKED

Loading the reference evidence.

EVENTS CHECKED
BUSINESS OUTCOMEA verified record is not a business-success verdict.
EXPECTED SIGNER / TRUST INPUTPublished synthetic reference key

The pin is an independent trust input. For this demo, it is supplied with the reference fixture or created by your own browser. Do not treat a key supplied inside an untrusted bundle as trusted merely because it verifies.

Evidence signer
Signed checkpoint / final chain hash
Reconstructed chain
Verification signer
THE TAMPER LAB05 TESTS

Break the copy.
Keep the original.

Each experiment modifies a copy of the evidence. The original signed bundle and trusted fingerprint remain available.

Original evidence. No modification is currently applied.
03 / PORTABLE EVIDENCE

Take the evidence
out of the UI.

Export the synthetic bundle and verify it with the included independent Python script. No LoopGrid service, API key or external business system is needed for that verification.

Import is limited to the synthetic reference format. Obtain the expected signer fingerprint separately; the uploaded bundle does not establish its own trust.

INDEPENDENT CHECKPYTHON
▤   manifest.json▤   events.jsonl▤   checkpoint.json▤   public-key.pem▤   bundle.json
$ python verify_synthetic.py \  evidence.zip \  --pin-file trusted-pin.txt
VERIFIED / INVALIDExit 0 / 1 · No network calls

Requires Python 3.10+ and the established cryptography package. Install the dependency before running offline. This verifier supports only the documented synthetic reference format.

04 / TECHNICAL TRANSPARENCY

What the demo proves.
What it doesn't.

01

Real cryptographic checks

Actual SHA-256 commitments, Ed25519 signatures, chain reconstruction, signed checkpoint and external signer pinning. Exported browser runs can be checked by the separate Python implementation.

02

Synthetic business execution

The agent is scripted, the reviewer is simulated, and the downstream adapter changes only browser memory. No real LLM, voice provider, bank, EHR, payment processor or identity system is connected.

03

Integrity is not completeness

Verification does not establish that every real-world event was captured, that upstream statements were true, that a human actually approved, or that an action was correct or compliant. Those require appropriate trusted capture boundaries and deployment controls.

DESIGN PARTNER PROGRAM

Have one consequential
workflow worth testing?

We're looking for technical teams to validate whether portable, independently verifiable evidence adds value beyond their existing logs. Start with one synthetic workflow; discuss real integration requirements only if the problem is relevant.

Share a workflow to test →Explore the Design Partner program ↗