PHASE 3.02 — HISTORICAL RULE ALIGNMENT AUDIT

PURPOSE
Before Phase 3.03 changes any policy rule, the preserved historical code was re-audited so that the Phase 3.xx programme remains aligned with the actual blackjack engine used previously.

CODE SOURCES CHECKED
- Phase 1.29 — archived Maximum Aggression policy.
- Phase 2.01 — deterministic finite six-deck hand-level validation.
- Phase 2.02 — deterministic continuous six-deck behavioural divergence validation.
- Phase 3.01 / 3.02 — optimisation-entry instrumentation and audit phases.

VERIFIED HISTORICAL RULES
1. Six physical decks are represented by a 312-card shoe in the finite-shoe phases.
2. Phase 2.02 deterministically shuffles the shoe once at the start of each 30-hand session.
3. No in-session reshuffle occurs in Phase 2.02 / Phase 3.01 / Phase 3.02.
4. Dealer stands on every total of 17, including soft 17.
5. Natural blackjack pays 3:2.
6. Only one split is modelled; recursive re-splitting is not modelled.
7. Doubling after a split is permitted.
8. No blackjack surrender action is implemented.
9. Maximum Aggression checks split eligibility before its aggressive first-decision DOUBLE.
10. Because 10/J/Q/K are stored as value 10, any initial two-card combination of those ten-valued ranks is split-eligible under Maximum Aggression.
11. No card counting, remaining-shoe composition, previous-card-history feature, hidden-card information or future-card knowledge is supplied to the policy.

DATASET SIZE
- Primary: 5,000 paired sessions × 30 rounds.
- Independent holdout: 5,000 paired sessions × 30 rounds.
These are session cohorts, not '60,000 shoes' or '30,000 shoes'.

CORRECTIONS TO THE EARLY DRAFT PHASE 3.02 PLATE
The following labels were presentation errors and are NOT properties of the historical engine:
- 'LS / Late surrender allowed' — REMOVE.
- 'RSA / Re-splitting Aces allowed' — REMOVE; recursive re-splitting was not modelled.
- 'reshuffled when ≤18 cards remain' — REMOVE; no in-session reshuffle occurred.
- 'Primary 60,000 shoes / Holdout 30,000 shoes' — REMOVE; each cohort contains 5,000 paired 30-hand sessions.

SCIENTIFIC CONSEQUENCE
No historical experiment result is invalidated by these corrections. The errors occurred only in the newly drafted explanatory plate. The Phase 3.02 Java code and reproduced Phase 2.02 bankroll results remain aligned with the preserved historical implementation.
