Judge guide
See the whole flow in about three minutes
Write a policy in plain English, publish it, run a customer through it with a synthetic document, and see what the platform proposes.
Synthetic specimen files
Fictional people and generated documents, made for this demo. Download a case folder and upload its id.png as the photo ID, bill.png as proof of address and selfie.png as the selfie.
| Case folder | What it shows | Expected decision |
|---|---|---|
p1-clean | Everything in order | Auto-approve auto_approve |
p1-expired | The ID card has expired | Request more information request_info |
p2-name-mismatch | The bill is in another person's name | Manual review manual_review |
p2-blurry | Blurry photos, stopped by the quality check | Request more information request_info |
Expected decisions assume the specimens' evaluation date, 2026-11-01. With Meta, dates are checked against today, so an older bill can also be flagged as too old.
Step 1: Unlock the demo 15 s
Enter the demo access code in the amber bar at the top of any page and press Unlock. The team shares the code with judges; without it, anything that uses AI or saves a policy is refused.
Step 2: Compile an example policy 45 s
In the Policy Studio, click the example With an unsupported rule, give it a name and category, then press Compile. Each setting in the preview shows the sentence it came from, and the sanctions sentence is flagged as unsupported instead of being guessed.
Open the Policy StudioStep 3: Publish it 20 s
Press Save draft, then Publish. The policy appears in the list as a new published version, with a link to its customer journey.
See all policiesStep 4: Run the customer journey with a specimen 60 s
Open the housing-loan journey, agree to the consent text, and upload the files from the
Start the journeyp1-expiredspecimen folder (see Synthetic specimen files above):id.pngas the photo ID,bill.pngas proof of address andselfie.pngas the selfie. Each upload is quality-checked straight away; tryp2-blurry/id.pngto see a retake hint.Step 5: See the result 20 s
The customer sees a calm outcome and what to fix (here, the expired ID), never a risk score. Repeat with
p1-cleanfor approval orp2-name-mismatchfor manual review. The developer tool shows the full pipeline activity for the same files, and Audit shows the events the run appended to the hash chain.
What this version does
The panel above says which AI provider this server runs. Every AI result carries a badge saying which provider produced it.
Meta Calls Meta's model when connected
- Policy compiler: labels each sentence with a rule kind or flags it; code checks every label against the sentence
- Customer step wording, written once when a version is published
- Document classification and field extraction, with a box showing where each value was read
- Comparing values across documents, the reviewer summary and the customer message
Code Always deterministic
- Every compiled rule is validated against a strict schema before it is saved
- Policy rules: expiry, minimum age, proof-of-address age, required documents
- Image quality gate: file type, size, resolution and a sharpness check
- Risk scoring and the decision rule: integrity issues go to a person, nothing is auto-rejected
- Audit hash chain, demo access code, rate limiting and the daily AI-call cap
Stand-in Synthetic or simplified
- Without a Meta key (mock mode): specimens return their ground truth, other files get placeholder data, and the compiler and wording use fixed patterns and templates
- ID registry: a small synthetic list, not a government service
- Selfie: quality check only, then discarded; no liveness detection or face matching
Not in this version Planned, not built yet
- Storing cases: documents and extracted details are not kept; only the case id, policy version and decision are recorded
- Encrypting case data with Vault (the per-customer key setup exists but holds no customer data)
- A reviewer console with evidence boxes
- Deleting a customer and the retention purge