Ladmir 0.5.0: tactics that change with level
The 0.5.0 build gives Mira, the temporary Echo Hound and ordinary enemies a small neural network for choosing how to move, support and select targets. Helpers use your character level; enemies use it within a cap set by the road.
· Truckfish Games
What changes as you level
The goal was to improve helper and enemy tactics as the player gains levels. Helper competence follows levels 1 through 24. The first six roads cap enemy tactical skill at levels 3, 6, 9, 12, 15 and 18 respectively; each frontier road caps it at 24.
In a controlled Windows player check, Mira chose a nearby grunt at level 1 and a wounded support target in the paired level-12 and level-24 situations. A level-24 ranged enemy backed away to make space. On the first road, a level-24 player still faced an enemy with effective tactical level 3.
Level affects which legal choice the model favors. Existing character and mission formulas still determine health and damage, and helper damage, healing amounts and attack intervals keep their authored values. There is no separate neural upgrade currency or extra control to manage.
A small model with fixed weights
The network has 24 inputs, a hidden layer of 32 ReLU units and eight outputs. Its 1,064 fitted parameters occupy 4,256 bytes as float32 values. Tools/train-neural-ai.py fitted the model offline using seeded synthetic tactical examples, then exported the weights into NeuralWeights.cs.
GameNeuralAI.cs reads the visible combat situation, including distance, health, sight, attack readiness, nearby allies, warnings and open paths. It ranks legal actions such as approach, back off, strafe, attack, support and regroup. The fitted weights stay fixed during play. The model does not learn from your play, read your camera scan or contact an online service.
Getting around cover
Review found a movement problem in the new controller: a blocked straight-line path could remove a movement choice before the existing avoidance code could try a detour. Mira or the Hound could become stranded behind a short wall or pillar while regrouping. Ordinary enemies could hold behind cover because Approach never reached ChaseEnemy.
NeuralRoute now offers short, validated waypoints around cover and briefly retains a side to follow. Enemy Approach can reach the existing ChaseEnemy avoidance path. Helpers still check collision, visible hazards, world bounds and distance from the player before taking a step. The fitted scores choose among the admitted actions.
The native Unity suite passed 39 controller regression cases covering helper regrouping, enemy pursuit around walls and pillars, hazard detours and decision-memory reset. Those cases exercise the actual movement controllers.
The combat rules still apply
Missing or invalid model content selects bounded authored fallback choices, so combat can continue. The Windows player checks confirmed that fallback respected Mira's live cooldown and produced a projectile hit after she was ready. Neural decisions remain transient; the update introduces no neural data into saves.
What the checks establish
On the accepted source, the Unity EditMode suite passed 237 of 237 tests, including 64 neural cases. The built Windows development player passed 186 of 186 assertions, including 62 neural checks. These are automated checks with prepared runtime situations, not a human campaign playtest.
The results cover selected tactical situations. Complete comparisons for every helper and enemy role at every level, full-campaign balance, human challenge and sustained physical-phone performance remain unverified. A Windows development build supplied the scripted measurements; the Truckfish download is Android only.
The exact Android 0.5.0 package passed offline emulator checks for the Truckfish startup screen, touch Continue, automatic dismissal, menu-to-play input and the More Truckfish games link. Gameplay checks used a prepared test character. The owner-confirmed physical-phone camera test from September 13 belongs to an earlier release and does not establish a new phone test for 0.5.0.