Does better decision-making actually help?

Measured on paired synthetic cargo staging episodes — not a claim of airline ROI. Costs are unitless. We disclose denominator, SLA, lead time, tail risk and cases where RELAY is worse.

ReportLoading…

Cost reduction — same arrivals, same resources

Each scenario has 500 paired episodes; 24 injection ticks and up to 110 drainage ticks.

Loading evidence…

These two policies are compared under the NEW shared, censored-aware v11 evaluation method. The older 21-tick percentages live in the original-v10 audit, and must not be conflated with this cohort-complete methodology.

What improved? What did not?

Scenario	Cost reduced	95% paired CI	On-time rate Δ	Staging lead time ↓	Tail cost ↓	Runs worse

On-time Δ is percentage points, not percent cost; lead time is restricted mean staging time, NOT flight recovery or human planning time. CVaR tail cost is across episode totals.

Recompute the evidence

Episodes Seed

Pinned report is generated by the bundled Python simulator; fresh runs use an isolated lab and do not interrupt your live three-Twin clock.

Scientific integrity & deployment scope

Model I — Placement

Bounded 2D floor candidates plus separate geometric 3D preflight. No aircraft load/restraint certificate.

Model II — PIVOT

Exact finite one-step Bayesian verification identity. No full sequential belief-MDP or global counterfactual reoptimization.

Model III — Network

Shared-capacity HiGHS MILP on enumerated routes. Cluster mutation remains a research fixture, not live optimizer.

Live data & cross-Twin checks

Dataset QA

Loading…

Cross-Twin lineage

Loading…

Production limit

Local-only, no airline credentials, no authorized actual capacity, no live tracking, no real customer tenancy. Do not expose on public Internet.

All detailed benchmark inputs and historical-v10 denominator comparison are available as JSON in /v11/api/kpis and /v11/api/v10-original-kpis.

Deep evidence: frequency → convergence → risk → ROI

Graphs are measured or derived from the bundled reproducible synthetic episode dataset. This is NOT an airline cost study.

Poisson input frequency compared to generated orders — 1. Synthetic order-arrival frequency (Poisson assumption; 12,000 10-minute ticks).

Paired cost reduction convergence with bootstrap confidence bands — 2. Statistical convergence: paired-sample estimator ± bootstrap 95% bounds. Not optimizer convergence.

Episode cost saving probability density — 3. Episode-level risk: some runs are worse under RELAY.

Synthetic worst decile cost improvement — 4. Tail risk (CVaR90 of full episode costs), separate from Model III scenario CVaR.

Illustrative scenario probabilities weighted cost savings — 5. Scenario probability → expected penalty costs. Scenario weights are assumed, not fitted.

Hypothetical ROI and breakeven sensitivity — 6. ROI sensitivity only: arbitrary HKD/unit and deployment volume; not measured ROI.

Probability of RELAY policy regression by scenario — 7. Percentage of runs where RELAY is worse; not all optimizations dominate baseline.

Synthetic cost breakdown by component — 8. Cost components: processing, waiting, lateness, unresolved penalty.

Temporary outage recovery backlog slower under RELAY — 9. Negative evidence: RELAY was 15.1% slower clearing 90% of disruption backlog.

MILP solve time against candidate column count — 10. Real finite-column MILP solve latency; separate from statistical convergence and 100× cinema speed.

Truth contract: The model's penalty unit has no real HKD exchange rate. Recovery lead time means ground staging, NOT airline network recovery or solver speed. No real Cathay data or observed business savings have been supplied.