Contest simulation & service timing

Simulate N competitors over discrete rounds and compare service-timing options under common random numbers, with simultaneous intervals and an explicit abstain.

contest_strategy answers one shaped question: when should this agent stop for service? It simulates the whole field of competitors forward over the remaining rounds, evaluates several timing options against the same simulated futures, and returns the contrast between them with honest intervals.

The shape recurs wherever agents compete over discrete rounds while a resource decays and a service interruption resets it — delivery fleets deciding when to recharge, harvesters when to unload, crews when to rotate, any process where stopping costs time now and buys performance later.

from algenta import MojoRuntime

runtime = MojoRuntime(mode="local")

params = runtime.execute("contest_strategy", "default_contest_params")["result"]
n = 12
board = runtime.execute(
    "contest_strategy", "option_board",
    [[90.0 + 0.12 * i for i in range(n)],   # per-agent rate
     [0.010] * n,                           # per-agent decay slope
     [2.5 * i for i in range(n)],           # elapsed so far
     [8.0] * n,                             # resource age
     [1.0] * n,                             # active
     3, 2, 30,                              # focal agent, target rival, rounds remaining
     params, 20000, 400, 42, 11],
)["result"]

board["v"][7]   # recommended option
board["v"][8]   # verdict: 0 STAY, 1 RECOMMENDED, 2 INCONCLUSIVE

The modules

Module	What it does
contest_strategy	Simulates the whole field of competitors forward under common random numbers to compare service-timing options, returning simultaneous confidence intervals and an explicit abstain verdict.

Functions

Function	What it does
option_board(...)	Four service-timing options evaluated against shared futures. Returns per-option success probability, paired deltas, simultaneous intervals, a recommendation and a verdict.
outcome_distribution(...)	Full-field finishing-position distribution per agent — win, top-three and points-equivalent probabilities, plus expected position.
default_contest_params()	The parameter vector, with documented defaults.
wear_penalty(decay, age, cliff_age, cliff_slope)	Resource penalty: linear decay plus the cliff past a threshold. One definition, shared by both simulators.
option_service_round(option)	Maps an option index to the round it services on.
contest_worker_count() / contest_strategy_version()	Parallel width and the contract identifier.

Common random numbers, and why they are not optional

Comparing options by running each against its own random futures buries the contrast in sampling noise. Every option here is evaluated against the same simulated world: the same incidents, the same service-time draws, the same per-agent form. Only the decision differs.

The gain is measurable rather than assumed — the board reports its achieved variance ratio, typically around 0.37, which is a 2.7× effective replication gain over independent sampling.

Randomness is counter-based, so a draw is addressed by (replication, round, agent, process) rather than consumed in sequence. That matters after options diverge: once one branch services and another does not, sequential streams hand them different numbers and the pairing silently disappears. Exogenous draws stay paired; endogenous events still differ, because the state differs.

The intervals are simultaneous, and the board can decline

The board selects the best of several options, so marginal intervals overstate confidence. It resamples whole replication rows jointly and uses a max-statistic, giving family-wise coverage.

It then applies an explicit abstention rule: recommend a service option only when its lower simultaneous bound clears the practical threshold ε. Otherwise the verdict is INCONCLUSIVE — a statistically real 0.05-point edge does not justify an operational action, and a board that always answers is not applying the rule.

How many replications you need

More is not automatically better, and the useful number is empirical. Across 45 real decision states:

replications	interval width	matches the N=50,000 answer
1,000	4.37 pp	87%
10,000	1.36 pp	93%
20,000	0.97 pp	96%
50,000	0.61 pp	100%

Interval width tracks 1/√N to within 1–2% at every level, which is the check that these are genuinely sampling intervals. Below about 20,000 the recommendation itself starts moving with N — at that point the board is partly reporting its random seed rather than its state.

Verification

Directional sanity is asserted rather than assumed: with too few rounds remaining to repay the service cost, servicing must never be recommended. Success probabilities are checked to sum to exactly 1.0, which is what catches a lost update in the parallel reduction. Degenerate inputs — zero or negative replications — return INCONCLUSIVE rather than NaN or a crash.

Next: model the decaying resource with asset degradation state, or score the resulting probabilities with forecast scoring.