Batch mcts-m5. Model 0.2.0, graph 0.2.0, parameters 0.1.1. 10 simulations, seeds 1 to 10; probes on 3. 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
South Korea played 20 rounds from 2026 under the tree search (mcts) method, with stochastic events, 10 times. Its capability index moved from 16.00 to a median of 17.11 (+1.11), with 90 per cent of simulations between 16.62 and 18.24 [experiment_objectives].
Starting position
In 2026 the nation ranks 6 on the index against the frozen field, holds 2 of nine domains at Tier 1 and has a tier rank of 7. Its base income is 1,014 credits a year, calibrated so that its 2026 position is affordable with headroom for investment.
What was decided
A simulation spent a median of 2,816 credits on actions. The five groups that took most were Technical & Vocational Skills (build, 87 credits, 3.2 per cent); Lifelong Learning & Reskilling (build, 84 credits, 3.0 per cent); Diplomatic Statecraft & Global Networks (partner, 82 credits, 3.0 per cent); Fiscal Capacity & Public Finance (partner, 78 credits, 2.8 per cent); Public Integrity & Anti-Corruption (build, 78 credits, 2.8 per cent) [experiment_decisions_by_group]. By mode: build 48 per cent, buy 27 per cent, partner 25 per cent [experiment_decisions_by_group]. Concentration of spending (HHI) had a median of 0.024, and the most repeated action appeared in 15 per cent of rounds.
The leading domain by period: rounds 1 to 5, Human Capital; rounds 6 to 10, Trade & Investment; rounds 11 to 15, Governance & Integrity; rounds 16 to 20, Critical Technology [experiment_spend_by_period].
The method's chosen action was within 5 per cent of its best alternative in 95 per cent of actions and ahead by more than 25 per cent in 0 per cent [experiment_decision_margins].
What the world did
A median simulation saw 3 shocks and 1 upside events. The most damaging per firing were S03 (6 firings in 5 simulations, −31.7 points each); S09 (1 firings in 1 simulations, −30.7 points each); S04 (4 firings in 3 simulations, −21.5 points each) [experiment_shocks]. Upside events fired 14 times across all simulations [experiment_shocks].
The groups that lost most to shocks, per simulation: Fiscal Capacity & Public Finance −1.73; Trade Facilitation & Logistics −1.61; Economic Leverage & Dependence −1.49; Digital & Services Trade −1.46 [experiment_shock_damage_by_group].
The edges that carried most change through the graph, per simulation: Fiscal Capacity & Public Finance to Social Infrastructure (+0.32 points); Fiscal Capacity & Public Finance to Force Generation & Mobilisation (+0.25 points); Economic Output & Productivity to Fiscal Capacity & Public Finance (+0.29 points) [experiment_propagation_by_edge].
Where it ended
At 2046 the median rank is 2 (5th to 95th percentile 1 to 3), the median count of Tier 1 domains 6 (5 to 7) and the median tier rank 2; the resilient score, the mean score sum over simulations, is 1,258 [experiment_objectives].
Tier 1 probability changed most in: financial strength (0 to 90 per cent of simulations); governance and integrity (0 to 100 per cent of simulations); information and influence (0 to 100 per cent of simulations); strategic infrastructure (0 to 30 per cent of simulations); trade and investment (0 to 100 per cent of simulations) [experiment_tiers_by_round].
Why
Without shocks the index ends +0.599 higher on average, so the shocks cost about 0.599 index points [experiment_probe_attribution].
Holding instead of following this method changes the index by −2.342: 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 −0.018 and −0.017, for example 2045 buy Space Operations (33 credits) [experiment_probe_attribution]. The largest actions by credits come late, when income is highest, and are rarely the most valuable.
Removing the largest shock of a simulation raises the index by +0.281 on average [experiment_probe_attribution].
The replay probe reproduces the simulations exactly (index change 0.000000), so the attributions are clean [experiment_probe_attribution].
Surprises
tier_jump (warn): largest gain in one domain within five years: 5 tiers (10 simulations above 2); tiers are reported, not an objective.
saturation_high (warn): 60% of simulations end with a group at 20 that did not start there (median 2 groups).
The simulation furthest from the median (seed 4) ends +1.14 from it, with 1 shocks and 3 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 search
Monte Carlo tree search 0.1.0 with candidate generators 0.1.0 (goal gap; top 4 groups each), 500 iterations a decision, UCT c = 1.4 scaled to the reward spread, widening alpha = 0.5, greedy playouts to the horizon, objective index; playouts drew their futures from the none (certainty-equivalent playouts: the planner ignores shocks) scenario.
The committed decision most often chosen, round by round: round 1 hold (100 per cent of simulations, visit share 96 per cent); round 2 hold (60 per cent of simulations, visit share 75 per cent); round 3 greedy (60 per cent of simulations, visit share 47 per cent); round 4 greedy (70 per cent of simulations, visit share 69 per cent); round 5 greedy (50 per cent of simulations, visit share 44 per cent) [mcts_decision_sequence]. The search hesitated (top two root candidates within 10 per cent) in 9 per cent of decisions, considered a mean of 13 candidates, grew trees of a mean 414 nodes to depth 5.3, and took 0.8 seconds a decision [mcts_decision_sequence].
Where the committed actions came from, by proposing generator (committed simulation-rounds): goal gap 13.4; greedy 12.9; hold 1.9 [mcts_root_candidates].
Against the greedy experiment with the same seeds (m5-kor-greedy), the search's index at the horizon is −0.085 on average (10th to 90th percentile −0.243 to +0.097), better in 30 per cent of seeds; on the index −0.085, the score sum −5.1 and Tier 1 domains −0.20 [mcts_regret].
Across seeds the same root decision won in a mean 72 per cent of simulations (first five rounds 68 per cent) [mcts_stability].
The principal variation was followed 100 per cent at depth 1, 71 per cent at depth 2, 68 per cent at depth 3, 72 per cent at depth 4 [mcts_pv_vs_path].