PHASE 3.02 — RESULTS AND PLAIN-ENGLISH EXPLANATION

WHY THIS EXPERIMENT EXISTS
The new Phase 3.xx programme is trying to stabilise the previously defined Maximum Aggression blackjack player. The long-term aim is to discover whether aggression can be applied selectively, using only information visible at the decision point, so that the aggressive player's probability of finishing a 30-hand session with a higher bankroll can be improved.

Before changing the policy, Phase 3.02 resolves an important rule question: when the first two cards are any combination of 10, J, Q or K, does Maximum Aggression split them before considering a double?

WHAT THE CODE CONFIRMED
The simulator stores 10, J, Q and K as the same numerical value: 10. Therefore combinations such as K+J, Q+10 and J+K arrive at the split test as 10+10. Maximum Aggression checks SPLIT before DOUBLE, so all such initial two-card combinations are split-eligible and are routed to SPLIT first.

This is consistent with the historical Maximum Aggression implementation. Phase 3.02 therefore clarifies the existing baseline; it does not rewrite the previous experiments.

WHAT REMAINED FROZEN
- Six-deck deterministic shoe.
- 30 hands per session.
- 100-unit starting bankroll.
- 1-unit base wager.
- Dealer stands on every 17, including soft 17.
- Natural blackjack pays 3:2.
- One split only; recursive re-splitting is not modelled.
- Doubling after a split remains permitted.
- No surrender action exists in the historical engine.
- One deterministic shuffle occurs at the start of each session.
- No reshuffle occurs during the 30-hand session.
- Primary and holdout each contain 5,000 paired sessions.
- Textbook player remains the control.
- Maximum Aggression remains the extreme baseline.
- No card counting, previous-card-history feature, remaining-rank inventory, hidden-card knowledge or future-card knowledge is available to the policy.

BASELINE REPLICATION
Primary average ending bankroll: Textbook 99.956; Maximum Aggression 78.978.
Primary positive sessions: Textbook 48.14%; Maximum Aggression 2.48%.
Holdout average ending bankroll: Textbook 99.972; Maximum Aggression 78.895.
Holdout positive sessions: Textbook 47.42%; Maximum Aggression 2.42%.

PLAIN-ENGLISH INTERPRETATION
Maximum Aggression is still performing very poorly compared with textbook play, which is exactly why it is useful as the extreme baseline for the new optimisation programme. Phase 3.02 does not attempt to improve it. Instead, it makes sure that the baseline policy is being interpreted correctly before any optimisation begins.

The new state-level evidence also shows why labels matter. A HARD 20 DOUBLE can occur after a split because only one split is permitted and a split child can subsequently receive another ten-valued card. It must not be confused with an initial K+J, Q+10 or similar two-card state, because those initial states are routed to SPLIT first.

WHAT THIS DOES NOT PROVE
Phase 3.02 does not prove that any particular policy change will improve Maximum Aggression. It also does not introduce shoe-history information or optimise against the known future card sequence. Its purpose is baseline integrity and correct interpretation.

WHY PHASE 3.03 FOLLOWS
With the ten-value split precedence now explicit and the finite-shoe baseline replicated in both primary and holdout cohorts, Phase 3.03 can make the first controlled policy-stabilisation change. Only one replicated visible-state policy family should be changed so that any bankroll effect can be attributed to that change.
