DETERMINISTIC BLACKJACK — MAXIMUM AGGRESSION POLICY OPTIMISATION
RUNNING RESEARCH GUIDE

Programme start: Phase 3.01
Current documented phase: Phase 3.04

======================================================================
1. WHY THIS NEW PROGRAMME EXISTS
======================================================================
The completed blackjack research established a broad evidence base comparing textbook and non-textbook behaviour, including a deterministic finite six-deck shoe and an intentionally extreme Maximum Aggression player.

The Phase 3.xx programme asks a narrower new question:

Can Maximum Aggression be stabilised into a selective, evidence-driven policy that improves its probability of finishing a fixed 30-hand session with a higher bankroll, without using card counting, hidden information or future-card knowledge?

The immediate target is not to maximise a 2x, 3x, 5x or 10x bankroll. First, the programme must determine whether indiscriminate aggression can be made materially more stable. Larger bankroll targets belong to later work only after the policy mechanism is understood and replicated.

======================================================================
2. EVIDENCE BASIS AND HISTORICAL BOUNDARY
======================================================================
PRIMARY DESIGN AUTHORITY
The optimisation programme is driven primarily by the deterministic finite six-deck evidence from Phase 2.xx and by new Phase 3.xx primary/holdout datasets.

SECONDARY HISTORICAL EVIDENCE
Earlier non-depleting/infinite-generator behaviour-profile experiments remain useful for historical comparison, hypothesis generation and corroboration. They do not directly define a new optimisation rule unless the finite-shoe evidence independently supports the same conclusion.

PROGRESSIVE EVIDENCE RULE
As Phase 3.xx develops, newly generated controlled optimisation datasets become the strongest evidence for subsequent policy changes.

======================================================================
3. FROZEN STARTING ARCHITECTURE
======================================================================
The programme initially preserves the established finite-shoe architecture so that old and new evidence remain comparable:

- Six decks / 312-card shoe.
- Deterministic session shuffle.
- 30 hands per session.
- 100-unit starting bankroll.
- 1-unit base wager.
- Existing dealer rules and blackjack payout.
- Existing split limit and double rules.
- No in-session reshuffle at the optimisation-entry stage.
- Textbook strategy remains the control/reference.
- Frozen Maximum Aggression remains the extreme baseline.

Thirty hands is retained initially as the established experimental horizon. Alternative session lengths may later be examined as a separate sensitivity study; they should not be changed at the same time as the first policy rules.

A realistic interval/penetration reshuffle is also reserved for a later robustness stage. It should not be introduced simultaneously with initial policy optimisation because doing so would make attribution harder.

======================================================================
4. INFORMATION BOUNDARY FOR THE NEW POLICY
======================================================================
The optimiser may eventually use information legitimately visible at the current decision point, including:

- player total and hand composition;
- hard/soft status;
- pair status;
- dealer upcard;
- legal actions;
- previously validated experimental risk/outcome trends associated with those visible states.

The initial programme does NOT give the policy a running count, previous-card-history feature, remaining-rank inventory, hidden-card information or knowledge of future cards.

Historical experimental trends may inform a rule before a new run begins. That is different from adapting a decision during the run because particular ranks have already left the current shoe.

======================================================================
5. REFERENCE POLICIES
======================================================================
TEXTBOOK
The established basic-strategy implementation remains the control/reference policy.

MAXIMUM AGGRESSION
Maximum Aggression is retained as the deliberately extreme boundary condition. Its action precedence is important:

1. If the initial two-card state is split-eligible, SPLIT first.
2. Otherwise, DOUBLE on the first legal decision under the frozen aggressive rule.
3. Continue according to the established Maximum Aggression implementation.

Because the simulator stores 10, J, Q and K as numerical value 10, all initial two-card combinations of these ten-valued ranks are treated as 10+10 for split eligibility. Thus K+J, K+Q, J+Q, 10+K and equivalent combinations are routed to SPLIT before DOUBLE.

======================================================================
6. PHASE 3.01 — FROZEN BASELINE AND OPTIMISATION ENTRY POINT
======================================================================
PURPOSE
Phase 3.01 entered the optimisation programme without changing Maximum Aggression. Its job was to preserve the finite six-deck baseline while adding state-level instrumentation capable of identifying where aggressive decisions cause repeated bankroll damage.

ARCHIVAL STATUS
The original Phase 3.01 experiment package has now been preserved as the official archival package and supplemented with a retrospective layman results explanation. This documentation addition does not change the Phase 3.01 Java code, datasets, output or experimental results.

WHAT WAS ADDED
The evidence record was expanded to include visible starting state, hard/soft status, dealer upcard, split-child status, textbook first action for the same visible state, aggressive action, stake, cards consumed, third-card behaviour, hand outcome and bankroll contribution.

WHAT WAS NOT CHANGED
The Maximum Aggression policy itself was not optimised. The six-deck architecture, 30-hand horizon, bankroll start and textbook control remained frozen.

WHY IT MATTERED
The instrumentation made it possible to move away from vague labels such as 'more aggressive' and instead identify adequately sampled visible-state families whose losses replicate in both primary and holdout evidence.

IMPORTANT FOLLOW-UP
Review of the initial state labels raised a question about hard-20 doubles and whether ten-valued combinations such as K+J were being split first. This required an explicit historical/code audit before optimisation could safely begin.

DOCUMENTED DESIGN EVOLUTION
The original Phase 3.01 next-design note anticipated that Phase 3.02 would make the first one-rule intervention. The hard-20 / ten-value split question emerged before that intervention was accepted, so Phase 3.02 was correctly reassigned as a verification gateway. The first genuine stabilisation intervention therefore moves to Phase 3.03.

======================================================================
7. PHASE 3.02 — TEN-VALUE SPLIT PRECEDENCE AUDIT
======================================================================
PURPOSE
Phase 3.02 resolved the ten-value split question before making any policy change.

AUDIT RESULT
PASS. The simulator represents 10, J, Q and K as integer value 10. Maximum Aggression evaluates split eligibility before the aggressive double rule. Therefore all initial two-card ten-value combinations are split-eligible and are routed to SPLIT first.

Examples:
10+J -> stored 10+10 -> SPLIT
K+Q  -> stored 10+10 -> SPLIT
J+K  -> stored 10+10 -> SPLIT
Q+Q  -> stored 10+10 -> SPLIT

This behaviour is consistent with the historical baseline. Phase 3.02 clarifies the interpretation rather than changing previous evidence.

BASELINE REPLICATION
PRIMARY
Average ending bankroll — Textbook: 99.956; Maximum Aggression: 78.978.
Positive sessions — Textbook: 48.14%; Maximum Aggression: 2.48%.
Paired winner — Textbook: 95.72%; Maximum Aggression: 3.88%; tie: 0.40%.
Mean cards consumed — Textbook: 165.674; Maximum Aggression: 177.840.

HOLDOUT
Average ending bankroll — Textbook: 99.972; Maximum Aggression: 78.895.
Positive sessions — Textbook: 47.42%; Maximum Aggression: 2.42%.
Paired winner — Textbook: 95.86%; Maximum Aggression: 3.86%; tie: 0.28%.
Mean cards consumed — Textbook: 165.647; Maximum Aggression: 177.749.

STATE-LEVEL CONTEXT
Primary resolved aggressive hands: 171,838; mean net per resolved hand -0.6117; bust rate 35.76%.
Holdout resolved aggressive hands: 171,789; mean net per resolved hand -0.6143; bust rate 35.72%.

The state ranking contains HARD 20 DOUBLE states marked [split-child]. These are not unsplit initial K+J-type hands. Under the one-split limit, a child created by a split may receive another ten-valued card and subsequently reach the aggressive DOUBLE rule.

LAYMAN INTERPRETATION
Phase 3.02 is a quality-control experiment. Before trying to make Maximum Aggression better, it confirms exactly what the aggressive player is already doing. The very poor bankroll result remains replicated in independent primary and holdout cohorts, while the split-precedence ambiguity is removed.

WHAT PHASE 3.02 DOES NOT CLAIM
It does not identify a proven improved policy. It does not use card counting or live shoe-composition information. It does not change the historical Maximum Aggression definition.

NEXT DESIGN — PHASE 3.03
Phase 3.03 should make the first genuine stabilisation change. Only ONE replicated visible-state policy family should be altered, chosen from adequately sampled finite-shoe evidence that shows the same damaging direction in both primary and holdout cohorts.

The purpose is attribution: if bankroll behaviour changes, we should be able to say which policy intervention caused it.

======================================================================
8. DOCUMENTATION AND PUBLICATION WORKFLOW
======================================================================
Each Phase 3.xx experiment should preserve four levels of evidence:

1. Java code — exact implementation.
2. Raw/detailed results and CSV evidence — audit trail.
3. Individual experiment results TXT — concise numerical results plus plain-English explanation.
4. This cumulative Research Guide — the connected scientific narrative across all experiments.

Publication plates are selective. A plate should be created when an experiment establishes an important baseline, milestone, conceptual transition or finding that benefits from visual explanation. Plates should explain why the programme has reached that point, which evidence is being used, what remains frozen, what is being tested and what the evidence means.


======================================================================
9. PHASE 3.03 — FIRST CONTROLLED STABILISATION INTERVENTION
======================================================================
PURPOSE
Phase 3.03 makes the first genuine policy change in the optimisation programme. It changes one replicated visible-state family only: a split-child HARD 20 now STANDS at its first decision instead of taking Maximum Aggression's automatic DOUBLE.

EVIDENCE FOR SELECTION
Phase 3.02 Primary and Holdout rankings independently placed split-child HARD 20 states at the top of the adequately sampled damage table. Across dealer upcards, the automatic DOUBLE produced approximately -1.65 to -1.77 mean units per resolved hand with roughly 91-93% bust rates; Textbook would STAND.

WHAT REMAINED FROZEN
The six-deck finite shoe, deterministic session-start shuffle, 30-hand horizon, 100-unit start, 1-unit wager, dealer/payout/split architecture, no in-session reshuffle, Textbook control and information boundary all remain unchanged. Initial ten-value split precedence remains unchanged. Every other Maximum Aggression first-decision state still follows the frozen aggressive rule.

PRIMARY RESULTS
Average ending bankroll — Textbook: 99.956; Selective Aggression: 82.767.
Positive sessions — Textbook: 48.14%; Selective Aggression: 5.48%.
Paired winner — Textbook: 92.04%; Selective Aggression: 7.54%; tie: 0.42%.
Mean cards consumed — Textbook: 165.674; Selective Aggression: 176.838.

HOLDOUT RESULTS
Average ending bankroll — Textbook: 99.972; Selective Aggression: 82.748.
Positive sessions — Textbook: 47.42%; Selective Aggression: 5.10%.
Paired winner — Textbook: 92.80%; Selective Aggression: 6.64%; tie: 0.56%.
Mean cards consumed — Textbook: 165.647; Selective Aggression: 176.801.

COMPARISON WITH FROZEN MAXIMUM AGGRESSION
Phase 3.02 Maximum Aggression averaged 78.978 Primary / 78.895 Holdout and produced 2.48% / 2.42% positive sessions. Phase 3.03 therefore recovers about 3.8 average bankroll units in both cohorts and raises positive-session frequency to 5.48% / 5.10%. The direction and approximate magnitude replicate independently.

MULTIPLIER MONITORING
From Phase 3.03 onward, ending bankroll thresholds >=200 (2x), >=300 (3x), >=500 (5x) and >=1000 (10x) are recorded as observational outcomes. All were 0.00% in both Phase 3.03 cohorts. These thresholds do not influence the policy at this stage.

LAYMAN INTERPRETATION
The first surgical correction worked in the intended direction. Maximum Aggression remains a poor player overall, but simply refusing to double a split-child hard 20 materially reduces damage. This is the first evidence that the extreme baseline can be stabilised incrementally using visible-state evidence rather than card counting or live shoe history.

LIMITATION
This experiment validates only one intervention. It does not establish the final selective-aggression policy and does not justify changing several neighbouring states together.

NEXT DESIGN — PHASE 3.04
Carry forward the Phase 3.03 HARD 20 split-child STAND rule. The remaining replicated ranking is now dominated by HARD 19 states. Phase 3.04 should select and alter only one next adequately sampled HARD 19 family, preserving one-rule-at-a-time attribution and all frozen architecture.

======================================================================
10. PHASE 3.04 — SECOND CONTROLLED STABILISATION INTERVENTION
======================================================================
PURPOSE
Phase 3.04 carries forward the Phase 3.03 split-child HARD 20 STAND rule and adds one further visible-state policy family only: HARD 19 now STANDS at its first decision instead of taking Maximum Aggression's automatic DOUBLE. The new HARD 19 rule applies to ordinary and split-child hands.

WHY HARD 19
After the Phase 3.03 intervention, HARD 19 states dominated the remaining replicated damage ranking across Primary and Holdout evidence. The family was strongly negative across dealer upcards, with approximately 80% bust rates under the aggressive DOUBLE, while Textbook uniformly STANDS on HARD 19.

WHAT REMAINED FROZEN
The six-deck finite shoe, deterministic session-start shuffle, 30-hand horizon, 100-unit start, 1-unit base wager, dealer/payout/split architecture, no surrender, no in-session reshuffle, Textbook control and information boundary remain unchanged. The Phase 3.03 split-child HARD 20 rule remains active. Every other Maximum Aggression state remains unchanged.

PRIMARY RESULTS
Average ending bankroll — Textbook: 99.956; Selective Aggression: 86.287.
Positive sessions — Textbook: 48.14%; Selective Aggression: 9.40%.
Paired winner — Textbook: 87.24%; Selective Aggression: 11.82%; tie: 0.94%.
Mean cards consumed — Textbook: 165.674; Selective Aggression: 176.252.

HOLDOUT RESULTS
Average ending bankroll — Textbook: 99.972; Selective Aggression: 86.023.
Positive sessions — Textbook: 47.42%; Selective Aggression: 8.76%.
Paired winner — Textbook: 88.72%; Selective Aggression: 10.48%; tie: 0.80%.
Mean cards consumed — Textbook: 165.647; Selective Aggression: 176.155.

COMPARISON WITH PHASE 3.03
Phase 3.03 Selective Aggression averaged 82.767 Primary / 82.748 Holdout with 5.48% / 5.10% positive sessions. Phase 3.04 therefore adds +3.52 / +3.28 average bankroll units and +3.92 / +3.66 percentage points of positive-session frequency. The direction again replicates independently.

CUMULATIVE POLICY LAYERS
1. Phase 3.03 — after an initial split, if a child hand subsequently forms HARD 20, STAND rather than DOUBLE.
2. Phase 3.04 — if a first-decision hand is HARD 19, STAND rather than DOUBLE, whether ordinary or split-child.

MULTIPLIER MONITORING
Ending-bankroll thresholds >=200 (2x), >=300 (3x), >=500 (5x) and >=1000 (10x) remain observational only. All were 0.00% in both Phase 3.04 cohorts.

LAYMAN INTERPRETATION
A second precisely targeted correction again improves the Maximum Aggression bankroll without replacing the whole policy with Textbook strategy. The player remains substantially worse than Textbook, but the evidence now shows a replicated progression: removing two specific forms of indiscriminate doubling materially reduces damage.

NEXT DESIGN — PHASE 3.05
Carry forward both validated rules. HARD 18 now dominates the remaining replicated damage table. Phase 3.05 should investigate one coherent HARD 18 family only, with all architecture and information controls frozen.

======================================================================
END OF GUIDE — CURRENT THROUGH PHASE 3.04
======================================================================
