Batch cmaes-dev. Model 0.2.0, graph 0.2.0, parameters 0.1.1. 2 simulations, seeds 1 to 2; probes on 1. Every figure is a distribution over simulations of a structural model with seed parameter values; it shows what the model implies under these settings, not what will happen. Other nations are frozen at their 2026 ratings. No reconstructed historical data is used in this experiment.
Headline
Estonia played 4 rounds from 2026 under the evolution strategy (cma-es) method, with stochastic events, 2 times. Its capability index moved from 11.49 to a median of 11.65 (+0.16), with 90 per cent of simulations between 11.52 and 11.78 [experiment_objectives].
Starting position
In 2026 the nation ranks 33 on the index against the frozen field, holds 0 of nine domains at Tier 1 and has a tier rank of 27. Its base income is 390 credits a year, calibrated so that its 2026 position is affordable with headroom for investment.
What was decided
A simulation spent a median of 232 credits on actions. The five groups that took most were Space Technologies (partner, 35 credits, 7.5 per cent); Development Cooperation & Global Assistance (buy, 35 credits, 7.5 per cent); Diplomatic Statecraft & Global Networks (partner, 29 credits, 12.7 per cent); Alliances & Security Cooperation (partner, 29 credits, 12.6 per cent); Development Cooperation & Global Assistance (partner, 24 credits, 5.1 per cent) [experiment_decisions_by_group]. By mode: partner 69 per cent, buy 27 per cent, build 3 per cent [experiment_decisions_by_group]. Concentration of spending (HHI) had a median of 0.140, and the most repeated action appeared in 75 per cent of rounds.
The leading domain by period: rounds 1 to 5, Information & Influence [experiment_spend_by_period].
What the world did
A median simulation saw 0 shocks and 0 upside events. The most damaging per firing were S06 (1 firings in 1 simulations, −21.5 points each) [experiment_shocks].
The groups that lost most to shocks, per simulation: Trade Facilitation & Logistics −0.41; Diplomatic Statecraft & Global Networks −0.37; Monetary & Currency Capability −0.35; Fiscal Capacity & Public Finance −0.34 [experiment_shock_damage_by_group].
The edges that carried most change through the graph, per simulation: Semiconductors & Advanced Computing to Artificial Intelligence (+0.61 points); Semiconductors & Advanced Computing to Economic Leverage & Dependence (+0.25 points); Space Technologies to Sensing, Timing & Navigation (+0.09 points) [experiment_propagation_by_edge].
Where it ended
At 2030 the median rank is 32 (5th to 95th percentile 30 to 33), the median count of Tier 1 domains 0 (0 to 0) and the median tier rank 28; the resilient score, the mean score sum over simulations, is 827 [experiment_objectives].
Why
Without shocks the index ends 0.000 higher on average, so the shocks cost about 0.000 index points [experiment_probe_attribution].
Holding instead of following this method changes the index by −0.302: that is the method's worth against doing nothing [experiment_probe_attribution].
Removing any single one of the five largest actions changes the index by between −0.017 and −0.009, for example 2030 partner Diplomatic Statecraft & Global Networks (12 credits) [experiment_probe_attribution]. The largest actions by credits come late, when income is highest, and are rarely the most valuable.
The replay probe reproduces the simulations exactly (index change 0.000000), so the attributions are clean [experiment_probe_attribution].
Surprises
same_action (warn): the most repeated action (group and mode) appears in a median 75% of rounds.
fitness_improved (warn): 38 per cent of 8 optimisations improved their hold-out fitness or stopped cleanly on stagnation; mean gain +0.0031 in reward units.
fitness_noise (warn): in the final generation the noise over shock seeds exceeded the best-minus-mean gap in 88 per cent of optimisations (mean noise 0.1884 against a mean gap 0.0230); population 4, 2 shock seeds.
The simulation furthest from the median (seed 1) ends +0.14 from it, with 0 shocks and 0 upside events [experiment_outlier_simulations].
Comparison
This is the baseline (hold) experiment, or there is none for this nation in the batch.
No backcast comparison: this experiment does not start from a historical position [experiment_against_history].
The optimisation
CMA-ES 0.1.0 (the cma package 4.5.0) with the rule encoding (26 numbers over the full horizon), receding, population 4, up to 3 generations a round, warm-started, stagnation stop at 30, 2 shock seeds a generation (common random numbers, redrawn each generation) and 2 hold-out seeds; objective index; plans scored under the stochastic scenario; started from the greedy plan; a single action capped at the large chunk.
8 optimisations ran a mean of 3 generations (8 stopped on budget) in a mean of 0 seconds each; the hold-out index of the best plan rose by a mean of +0.0031 from the starting plan [experiment_cmaes_convergence_by_round].
After shocks, the credits actually played were 100 per cent of those the first plan intended, and 0 per cent of realised actions differed from the plan [experiment_cmaes_planned_vs_realised_by_round].
Against the greedy experiment with the same seeds (dev-est-greedy), the plan's index at the horizon is +0.032 on average (10th to 90th percentile +0.012 to +0.053), better in 100 per cent of seeds [experiment_mcts_regret_by_round].
Against the tree search on the same nation, horizon, objective and seeds (dev-est-mcts-index), the plan's index at the horizon is +0.052 on average, better in 100 per cent of seeds [experiment_cmaes_vs_mcts_by_round].
Allocation similarity by round (cosine, credits by group and mode): with greedy 0.29, with the tree search 0.24 (tree search with greedy 0.86) [experiment_cmaes_similarity_by_round].
In the final generations the noise over shock seeds was 0.1884 against a best-minus-mean gap of 0.0230; the noise was the larger in 88 per cent of optimisations [experiment_cmaes_noise].