PHASE 3.18 — LAYMAN RESULTS EXPLANATION
TRUE-DIVERGENCE FAMILY AUDIT AND HARD 9 INTERVENTION

WHAT DID 3.18 DO FIRST?
Phase 3.18 used the corrected Phase 3.17 definition of unresolved policy damage. Only hands where Selective Aggression actually chose a different first action from Textbook were allowed into the candidate audit.

This prevents ordinary blackjack losses under an already-correct action from being mistaken for a policy defect.

CORRECTED FAMILY COMPARISON
Residual HARD 5–8 vs dealer 2–6:
- Primary: 4,898 divergent hands, mean net -0.2246, aggregate net -1,100.
- Holdout: 4,765 hands, mean net -0.2191, aggregate net -1,044.

HARD 9 vs dealer 9–Ace:
- Primary: 2,594 divergent hands, mean net -0.3978, aggregate net -1,032.
- Holdout: 2,532 hands, mean net -0.4186, aggregate net -1,060.

All remaining SOFT divergences taken together were actually net positive:
- Primary +538 aggregate units.
- Holdout +206.
So it would have been wrong to label all remaining soft hands as one damage family.

A narrower SOFT 18–19 vs 9–Ace subgroup remained negative, but its aggregate damage was smaller than the two hard-hand candidates.

WHY HARD 9 WAS SELECTED
HARD 5–8 had slightly greater total aggregate loss, but HARD 9 vs 9–Ace showed much stronger damage per occurrence and an exceptionally coherent mechanism:
- Textbook says HIT throughout the family.
- Maximum Aggression says DOUBLE.
- Immediate bust risk from taking one card is 0%.
- Primary and Holdout agree closely.

This therefore gives us a clean independent test rather than absorbing the larger residual low-total family all at once.

WHAT CHANGED
HARD 9 versus dealer 9–Ace now HIT instead of automatically DOUBLE.
All previous validated rules remain frozen.

RESULT
Primary average ending bankroll:
97.992 -> 98.167, improvement +0.175.

Holdout:
97.931 -> 98.109, improvement +0.178.

Positive sessions:
Primary 37.60% -> 38.36% (+0.76 percentage points).
Holdout 36.96% -> 37.78% (+0.82 percentage points).

The gain is now small, but the replication is exceptionally close: +0.175 vs +0.178.

WHAT DOES THIS TELL US?
This is increasingly strong evidence of diminishing marginal returns.

The early programme removed broad, highly damaging forms of Maximum Aggression and produced large improvements. By 3.18, a well-replicated intervention affecting more than 2,500 hands per cohort produces only about +0.18 units at programme level.

That does not make the intervention invalid. It tells us that the remaining policy defects are contributing progressively less to the overall 30-hand bankroll distribution.

DISTANCE FROM TEXTBOOK
Textbook average ending bankroll:
- Primary 99.956.
- Holdout 99.972.

Selective Aggression after 3.18:
- Primary 98.167.
- Holdout 98.109.

The remaining gap is approximately 1.79 units Primary and 1.86 units Holdout.

PEAK BANKROLL AND MULTIPLIERS
Maximum Selective bankroll:
- Primary 128.0 at Hand 26.
- Holdout 125.5 at Hand 29.

No session finishes at 2x, 3x, 5x or 10x.

The multiplier result remains expected under the frozen 1-unit wagering architecture and does not count against the quality of the decision policy.

WHAT REMAINS?
The corrected true-divergence ranking is now dominated by relatively narrow groups:
- HARD 5/6 vs weak dealers;
- residual HARD 7/8 weak-dealer states;
- SOFT 18 vs 9/10;
- SOFT 19 vs 10.

These are much narrower than the broad errors removed earlier.

NEXT QUESTION — PHASE 3.19
Aggregate the residual HARD 5–8 weak-dealer family again after 3.18 and compare it directly with the narrower SOFT 18–19 strong-dealer family.

Because 3.18 produced only about +0.18 units, Phase 3.19 should also begin formal closure-threshold thinking: how much replicated improvement is enough to justify adding another permanent rule to the policy?

Do not force closure solely because the gain is small, but do not add complexity without measurable replicated benefit.

PUBLICATION NOTE
This Layman Results report contains the complete explanatory interpretation for Phase 3.18.
