The DLA Pricing Policy

In This Component

The pricing policy is a Direct Lookahead Approximation (DLA) — Powell’s most sophisticated policy class. It doesn’t just react to the current order. It simulates what the world will look like after accepting or rejecting, across $N$ stochastic futures, and selects the action that maximizes expected cumulative reward over a horizon $H$.

The DLAPolicyEngine

For each incoming B2B order, the engine executes:

1. Candidate Generation Three price tiers × three schedule options = up to 9 candidate B2BAction objects. Schedule options are min(request_day, SLA) - 4, - 2, and 0.

2. Feasibility Filter Each candidate must pass MarginObjectiveFunction.is_valid_action(): sellable inventory must cover quantity, schedule must not exceed delivery offset, and immediate margin must be positive.

3. Seeded Monte Carlo Rollouts Per feasible action, $N$ simulation paths are run. Seeds are deterministically derived from the order UUID using NumPy’s SeedSequence — the same order always produces the same rollout, making decisions fully auditable:

event_seed_words = [(event_id.int >> shift) & 0xFFFFFFFF for shift in (0, 32, 64, 96)]
seeds = tuple(
    SeedSequence([self.simulation_seed, *event_seed_words, i]).generate_state(1)[0]
    for i in range(self.num_simulations)
)

4. Path Scoring

$V(x_t)=\frac{1}{N}{\sum }_{n=1}^N{G}_n(x_t)\text{where}{G}_n={\sum }_{h=0}^H{C}_h$

Invalid future actions make the entire path worth $-\infty$.

5. Stable Argmax + Business Guard The action with the highest expected value is selected. A final safety guard converts an invalid accepted action to a reject, preventing a pathological edge case from reaching the customer.