HARBORSHIELD / v0.2 / PROTOCOL 2026-10-04
From a recommendation
to reproducible evidence.
These are synthetic experiments, not historical shipping observations. Independent samples check decisions under an assumed model; they do not validate insurance pricing or real-world performance.
20 containers · 70% route cap · 30-day mean-transit limit · 1,000 training worlds · 10,000 held-out worlds · five training seeds.
Normal operations
| Policy | Direct / Busan / Klang | Test mean | Test CVaR95 |
|---|---|---|---|
| Joint CVaR | [6, 14, 0] | $153,987 | $218,369 |
| Cost only | [6, 14, 0] | $153,987 | $218,369 |
| Balanced feasible | [7, 7, 6] | $161,345 | $228,547 |
| Additive tail proxy | [6, 14, 0] | $153,987 | $218,369 |
Monsoon weather
| Policy | Direct / Busan / Klang | Test mean | Test CVaR95 |
|---|---|---|---|
| Joint CVaR | [6, 14, 0] | $223,572 | $313,742 |
| Cost only | [6, 14, 0] | $223,572 | $313,742 |
| Balanced feasible | [7, 7, 6] | $240,822 | $336,379 |
| Additive tail proxy | [6, 14, 0] | $223,572 | $313,742 |
Port congestion
| Policy | Direct / Busan / Klang | Test mean | Test CVaR95 |
|---|---|---|---|
| Joint CVaR | [6, 14, 0] | $254,268 | $360,812 |
| Cost only | [6, 14, 0] | $254,268 | $360,812 |
| Balanced feasible | [7, 7, 6] | $272,107 | $383,175 |
| Additive tail proxy | [6, 14, 0] | $254,268 | $360,812 |
Strait disruption
| Policy | Direct / Busan / Klang | Test mean | Test CVaR95 |
|---|---|---|---|
| Joint CVaR | [10, 10, 0] | $298,116 | $426,050 |
| Cost only | [6, 14, 0] | $296,801 | $426,536 |
| Balanced feasible | [7, 7, 6] | $310,744 | $440,676 |
| Additive tail proxy | [6, 14, 0] | $296,801 | $426,536 |
Shared-shock sensitivity
Normal regime: held-out batch CVaR95. Policies are re-selected at each latent dependence assumption. The horizontal bars use a zero cost origin.
What the results do—and do not—show
Baseline policies can coincide. If cost-only and joint-risk allocations are the same, the measured improvement is zero. This model does not guarantee that a more complex method will outperform a simpler policy.
The full report includes paired bootstrap intervals, training-seed stability, sample-size diagnostics, equations, input hashes and exact unrounded values. The MPA arrivals series is displayed separately as public-data context.