A deterministic decision engine with traceable execution and regime arbitration. Zero dependencies. Bit-identical replays.
$ git clone https://github.com/adam-scott-thomas/maelstrom-runtime.git
$ cd maelstrom-runtime && pip install -e .
$ python examples/demo_runner.py examples/minimal_spec.json
Six regimes compete for control. Stress deforms legality. Bypasses collapse deliberation. Every cycle is hashed and auditable.
Scroll to see how it works ↓
The Core Loop
Each decision passes through structured conflict:
Evaluate Clarity, risk assessment
Generate Optionality, alternative paths
Select Defensibility under regime constraints
Execute Commitment and action
Reflect Structural comparison and regret
Conflict is not failure.
It is the mechanism through which volition emerges.
Regime Arbitration
At any moment, one regime dominates because its penalty gradient is rising fastest.
r*(t) = argmaxr (ΔPr / Δt)
The system does not "choose" a regime. The environment forces the shift.
Transitions are hysteretic:
Entry into crisis is fast.
Recovery is slow.
Arbitration is continuous.
Dominance is emergent.
Legality Deformation
Stress is not a scalar penalty.
It deforms the legality graph — the space of procedurally valid transitions.
A'ij = Aij − αij · S(t)
Admissibility shrinks under stress
W'ij = Wij + βij · S(t)
Transition cost increases
High stress does not merely reprioritize options.
It removes entire pathways from consideration.
Deliberation operates within what remains admissible.
At low stress, all five transitions are admissible. The full deliberative loop operates.
Bypass Dynamics
Under extreme stress or latency pressure, full evaluation may exceed the available budget.
The system collapses into a lawful shortcut — a bypass.
Bypasses are not bugs.
They are regime-appropriate compressions of the loop.
Bypasses trade completeness for viability.
Counterfactual Regret
Each cycle archives:
Regret(t) = max(counterfactual_value) − executed_value
Positive regret indicates a superior forgone alternative.
Regret is not emotion.
It is structural comparison.
Accumulated regret informs doctrine — persistent modifications to future behavior formed from repeated constraint.
The system learns without replaying catastrophe.
Perfect Determinism
Maelstrom is fully deterministic.
Given the same scenario specification and seed, the runtime produces bit-identical traces.
Validated over:
Every cycle is hashed.
Every regime shift, veto, bypass event, and regret score is logged.
Merkle-verified trace integrity.
Full post-hoc auditability.
Why It Exists
Autonomous systems increasingly operate in environments where:
Maelstrom provides:
It is a foundation for systems where legitimacy, survivability, and recoverability matter as much as output quality.