[INTERNAL] CUBEX^3 // Algorithm Brief

Algorithm Index

ai leveling APD

Tensor Field + Tiered Features

Base feature set (v1-v4), tier unlock cadence, sparse board updates, and polarity-driven learning loop. Defines skill points as concurrent cube reads per feature.

Master Piecewise Risk Reward

RRR Evaluation Core

Risk, Reward, Relation evaluation across material, board control, and morale for turn-level scoring and AI posture selection.

Morale

Morale Hierarchy

Unit hierarchy rules for pawns, commanders, generals, and king. Defines morale loss propagation and dual scoring (AI and narrative).

Board control

CubeNode State

Board cube data structure for occupancy, threat status, and material weighting for control maps.

Tf_Algorithm

Hierarchy Swap + Stance Logic

Linked-chain morale engine, swap-based capture handling, mercenary masking, and stance inference from recent opponent RRR outputs.

Master Piecewise (RRR) Evaluation

Risk Reward Relation

Each turn computes three deltas. These map to Offense, Defense, and Strategy labels and drive the AI stance model.

                    Risk = Opponent Material - Player Material
                    Reward = Opponent Board Control - Player Board Control
                    Relation = Opponent Morale - Player Morale
                

Scoring Bands

The RRR values normalize to -1, 0, +1. Risk uses fixed thresholds. Reward and Relation share a configurable X threshold.

                    Risk <= -4  => -1
                    Risk between -3 and +3 => 0
                    Risk >= +4  => +1

                    Reward <= -X => -1
                    Reward between -X and +X => 0
                    Reward >= +X => +1

                    Relation <= -X => -1
                    Relation between -X and +X => 0
                    Relation >= +X => +1
                

Material Scale Options

Two material value scales appear in the notes. Align on one scale before final tuning. Scale B marks the king as infinite value.

Piece	Scale A	Scale B
King	40	inf
Queen	12	18
Fractal Knight	9	14
Bishop	6	9
Knight	4	8
Rook	7	10
Pawn	1	1

Morale Hierarchy System

Morale is hierarchical and cascades upward. Losses ripple from pawn to commander to general to king. Two channels are tracked: AI heuristic morale and narrative morale.

Role	Morale Value	Notes
King	21	Top of hierarchy, total morale drops to 0 on capture/checkmate.
General	10	Gold General = Fractal Knight, Silver General = Queen.
Commander	3	Right flank (gold) and left flank (silver) commanders.
Pawn	1	Base morale. Loss subtracts from commander and general.

Linked Chain + Swap Logic

Linked Identity

Each unit stores commander and general IDs. Links are immutable across the match.

Capture Swap

Captured pieces move to a graveyard cache, retaining experience and stats.

Mercenary Mask

If leadership is missing, subordinates mask faction bonuses and revert to base morale.

Rescue

Restoring a leader un-masks subordinates and restores morale contributions.

Board Control Model

Board control scores count all unblocked valid moves for each piece each turn. Additional weighting includes checks and shielded clusters.

                    Board Control = sum(valid moves of active pieces)
                    + (material value of enemy piece in check)
                    - (material value of player piece in check)
                    + (material value of shielded friendly cluster)
                    - (material value of shielded enemy cluster)
                

CubeNode State

                    CubeNode:
                      - position (x,y,z)
                      - isOccupied
                      - occupying_piece
                      - history (last 3 occupancies)
                      - isThreatened / isThreatening
                      - materialValue (dynamic weighting)
                

Tensor Field + Tiered Feature System

The AI evaluates a sparse tensor field of 512 cubes. Base features are always active, tiers add higher order pattern detectors, and skill points control how many cubes a feature can read simultaneously.

                    Tf = Base[v1..v4] + T1[v1..v4] + T2[v1..v4] + T3[v1..v4]

                    v1: piece enumeration
                    v2: legal move generation
                    v3: heuristic board value
                    v4: action selection
                

Sparse Updates + Polarity

Board cubes self report deltas only. The AI updates affected features and computes polarity from snapshot differences to strengthen or prune feature weights.

                    Polarity = f(Delta Tf)
                    Positive polarity -> reinforce features
                    Negative polarity -> decay unused features
                

Stance Inference

Recent RRR outputs are used to classify the opponent posture. The last three labels (Off/Def/Str) map to an aggression profile that biases future move selection.

                    Example:
                    Off +1, Off -1, Str 0 => moderately aggressive, mildly tactical
                    Off +1, Off +1, Off +1 => hyper aggressive
                    Off -1, Off -1, Off -1 => hyper retaliatory
                

Source Files

Cubex^3/ai leveling APD.txt
Cubex^3/board control.txt
Cubex^3/Master Piecwise Risk Reward.txt
Cubex^3/Morale.txt
Cubex^3/Tf_Algorithm.txt