LuckyJ digital mahjong

Digital Points for Playing Like LuckyJ

What this book is

What LuckyJ teaches.

LuckyJ is the self-play-trained AI at the center of this book. Its live Tokujou games show a digital riichi style built around practical table judgment layered on top of tile efficiency. The recurring habits are grouped into things you can actually use mid-hand: value, shape, safe tiles, calls, pressure, and placement.

NAGA is the AI review lens. Its model heads, including Nishiki, Hibakari, and Kagashi, show strong candidate lines and expose the moments where LuckyJ chooses another plan. Those disagreements are review prompts, not automatic improvements over ordinary efficient play. LuckyJ still matches Nishiki's top discard about 73% of the time.

LuckyJ

A self-play-trained policy observed making repeated choices under score pressure, opponent pressure, and awkward hand shape.

NAGA

A strong riichi AI review system used here as a comparison partner and calibration tool.

Teacher

The split is a question: what the model values, what LuckyJ may be buying, and whether a human can verify the condition before copying it.

Start with point 01, then use the full playbook for point-by-point examples with the hand, table context, and review prompt shown together.

Start with point 01 Read the evidence audit

Why it teaches

The lesson is the decision loop.

  1. See the table. Start from score, row, dora, calls, riichi, and the tiles that become dangerous later.
  2. Compare the plans. NAGA gives the strong model line; LuckyJ's choice shows what table condition changed the price.
  3. Copy the condition. Learn when value, shape, safety, or placement should take priority.
qruz,157

Source scope

Tokujou was the test, not the training data.

LuckyJ learned by self-play. This book analyzes a mixed Tenhou review corpus—1,079 Tokujou, 127 Upper, and 49 General games—with no Houou or Mahjong Soul sample. The full playbook now uses Tokujou-only teaching examples and labels the room beside each source.

Copy the portable mechanics first. Recalibrate calling, riichi, keiten, safety inventory, and placement thresholds for your room and rank-point system.

For Saint 3 four-player South, a nominal fourth at even score is about -255 points while first is about +125 in Jade or +135 in Throne. That makes fourth-place control valuable, but it supports target-specific exits—not indiscriminate safety hoarding against opponents who give you fewer free turns. Read the calibration and limits.

Safe tiles and timing

The correction is narrower defense.

The follow-up review sharpened a normal riichi concept: genbutsu and suji matter most when they answer a real threat. A safe tile by itself is too vague. The stronger LuckyJ habit is timed, target-specific, and connected to hand value.

Safe tiles

Safe tiles expire.

A genbutsu tile loses value when it no longer defends the live danger or when keeping it steals the hand's route.

Read point 14
Shape

Late outside cuts can preserve route.

Under pressure, cutting the edge tile can keep an inner winning path while holding immediate danger down.

Read point 15
Defense

Defense is target-specific.

Kept genbutsu and suji matter most when they answer a live riichi or open threat in the middle or late row.

Read point 13
Honors

Honors need role labels.

Self value, opponent yaku-condition tile, dead honor, and safe tile are different honor jobs.

Read point 16
Placement

Leader safety is a qualifier.

With a lead, reduced ambition is only good when the discard also avoids extra danger and preserves a clean next turn.

Read point 01
Defense

Danger labels need live context.

Safe-looking labels are only priors; update them with visible blockers, opponent routes, and suit constraints.

Read synthesis
Shape

Route value beats surface shape.

Wait shape is one input. Value, yaku certainty, ron likelihood, and fallback defense can change the best route.

Read synthesis
Commitment

Inventory comes before commitment.

Riichi freezes the hand, so price the leftover tiles before declaring; after calls or pushes, keep updating their jobs.

Read synthesis
Revision to point 14: Keep safe tiles when they defend the next real threat, then spend them when that job expires. The evidence audit adds one more check: before calling a tile safe, test the label against live table context.

Method notes

How to read the examples.

The examples come from LuckyJ games and AI review output, but the page is written for play. Use the numbers as a way to find interesting hands, then judge each spot through normal riichi ideas: value, shape, safe tiles, opponent yaku conditions, and placement.