Pilot plan
How this prototype becomes a real trial: one corridor, ninety days, three modes of increasing effect, with measured gates and written reasons to stop. Thresholds are proposals to agree with customs and the authorities, not findings.
Scope
- One entry point and one corridor, for example a sea or air flow with a mix of food, electronics and regulated goods that touches at least three authorities.
- Customs and each participating authority keep every legal decision. Madoun supplies a shared file, a parallel plan and a lane suggestion.
- Data stays in the UAE, on infrastructure the government chooses. Madoun reads from ATLP and MAMAR through adapters and does not replace either.
Gates
| Gate | When | Effect on real decisions | Passes when |
|---|---|---|---|
| Gate 0: ready | Before day 1 | None | Baseline release times measured with the WCO Time Release Study method; data-sharing agreement signed; security review and AI impact assessment signed off. |
| Gate 1: shadow | Days 1 to 30 | None. Runs beside real decisions. | At least 95% of declarations map without manual repair; API answers within 500 ms for 95% of calls; no high-severity security finding; lane agreement with officers measured and reported. |
| Gate 2: advisory | Days 31 to 60 | Officers see the parallel plan and owners. Nothing is automatic. | Every waiting step has a named owner; officers use the shared evidence file in most reviews; override rate stays under 40% and every override carries a reason. |
| Gate 3: limited effect | Days 61 to 90 | Pre-arrival release for green shipments of authorised operators only. | Median release time falls against the baseline with the whole 95% interval below zero; audits of green releases show a violation rate within the agreed tolerance; no group bears a clearly higher share of checks that find nothing. |
Stop criteria
- The violation rate found by post-release audits of green shipments rises above the agreed tolerance.
- Any high-severity security or privacy incident.
- Sustained override rate above 40%, which means officers do not trust the lanes.
- Data quality too poor to assess fairly, for example more than 5% of declarations unmappable.
- A legal or policy objection from any participating authority. Their veto holds.
What changes on day one
| Who | What they see and do |
|---|---|
| Customs officer | Opens one file per shipment with lane, reasons, owners and clocks. Overrides a lane with a reason. Resolves exceptions. |
| Regulator officer | Sees only the evidence shared with their authority, reuses a receipt instead of asking again, and records their own approval in parallel. |
| Trader or agent | Sees the status of their own shipment and what is missing, earlier. Does not see other parties' receipts. |
| Auditor | Reads the full trail and exports a clearance record. Cannot change anything. |
| Inspection operator | Receives inspection tasks from any performer: officer, scanner, drone or robot. A person confirms any serious finding a machine reports. |
Where each part stands
| Part | Readiness | What is missing |
|---|---|---|
| Rule engine, event log, review plan | Level 4: validated on synthetic data in a lab | Validation on real, labelled history. |
| Evidence receipts and signatures | Level 3 to 4 | Real key custody in each authority, external anchoring of the ledger head. |
| Integration API | Level 3: mock | Authentication, connection to ATLP and MAMAR, load testing. |
| Learning loop and model check | Level 3 | Real outcomes, audit sample large enough to trust, drift monitoring. |
| Robotic inspection seam | Level 2 to 3: contract and simulation only | A real robot, a safety case, terminal access, officer confirmation rules. |
| Adapters to ATLP and MAMAR | Level 1 to 2: designed, not built | Access to interface specifications and a test environment. |
How many audits would the pilot need?
To trust the violation rate in the green lane, the pilot must audit enough green shipments. Enter the rate you expect and how close you need to be.
About 1,118 audited green shipments are needed to know the rate within ±1 points with 95% confidence.
At 1 to 3 violations per 100 and a margin of one point this takes several hundred to over a thousand audits, which is why audits start in shadow mode, long before any release changes. If it is too many for one corridor, widen the margin or the corridor.
What the pilot needs from partners
- Customs: baseline timestamps, read access to declarations, an officer lead and a decision on lane policy.
- Each authority: one named reviewer, its evidence-sharing rules, and the right to stop.
- Technology partners: a test environment, key management, and a robot cell for the research track.
- Legal and privacy: lawful basis, retention and audit rules, and sign-off on the AI impact assessment.
Fuller detail is in the repository: pilot plan, research programme, AI impact assessment, governance, threat model, model card.