ARCHITECTURE SPEC: CONTENT GENERATION: NONE LANGUAGE MODELS: NONE NET DISPATCH: 0 BYTES CPU BUDGET: BOUNDED

Actorics is not generative AI in any form. It does not generate art, audio, text, levels or dialogue. The runtime selects an action from a set your designers authored — the same thing a behavior tree does, except the opponent stops falling for the same tactic. No language models. No network calls. No tokens.

On-device behavioral runtime for game characters.

Opponents, bosses and squadmates that stop falling for repetitive player tactics during combat. Evaluated locally on the CPU within your frame budget. Zero network calls. Zero cloud token bills. Fully constrained by Designer Envelopes.

[INSPECT UNREAL / UNITY INTEGRATION] [REQUEST A PILOT]
1
The player repeats one moveRolls behind the boss and strikes.
2
The boss spots the patternThree repeats of the same move are enough.
3
The boss stops falling for itIt reads the roll, steps back and blocks. Same speed, same damage.
Just watch — it plays by itself and restarts every ~25 seconds.Or try it: click the arena, ←↑↓→ move, ShiftShift roll, Space strike.
ATTRACT · PLAYER REPEATS ROLL_BEHIND [0/3] T+00:00.000
Click the arena to take control
WASD move · Space strike · Shift×2 roll · P pause
SCRUB — / —

This sandbox is a JavaScript illustration of the mechanic, not the Actorics core. Timing values come from the Envelope controls below; no performance number is measured here.

AGENT TELEMETRYBOSS_01
AGENT_STATE:
COMBAT_ENGAGED
EVALUATION_CORE:
ON-DEVICE CPU
TICK_RATE:
60 Hz (FIXED STEP)
POLICY:
LEARNING
DETECTED_EXPLOIT:
NONE
ENVELOPE_STATUS:
CONSTRAINED_OK
NET_DISPATCH:
0 BYTES (LOCAL)
FRAME_OVERHEAD:
— . — — µs [AWAITING HARNESS]
DECISION LOG0 EVENTS
SEC.02 / PROBLEM

Gameplay combat breaks down in the frame budget.

Hand-authored trees get solved by players. Cloud agents don't fit in a frame. The behavioral layer sits between the two, and it is still written by hand.

INDUSTRY STANDARDSTATIC

Static behavior trees

  • The player finds a dominant tactic within minutes and repeats it for the rest of the encounter.
  • Every new enemy type means hours of manual tree tuning by a programmer.
GENERATIVE CLOUD AGENTSEXTERNAL

Remote inference per character

  • External inference with round-trip latency that combat timing can't absorb.
  • A running bill per session, per player.
  • Late decisions break animation blending and telegraphs.
  • Live generation requires a storefront AI disclosure.
ACTORICSLOCAL

Behavioral runtime

  • A compact state evaluator on the player's CPU. No GPU, no server.
  • Punishes repetition of a single tactic instead of scaling damage or health.
  • Runs inside hard boundaries set by your designers.
  • Target: zero allocations per tick, zero external dependencies.
SEC.03 / ARCHITECTURE

Dynamic adaptation inside rigid designer boundaries.

One tick, five stages. The controller proposes; the Envelope decides what is allowed; commitment keeps animation intact. Select a stage.

Designer Envelope

These values drive the sandbox above and the struct on the right.

Recovery window after an attack. Adaptation can never shorten it.

The designer sets the delay; the floor is fixed at 150 ms.

How much better a new action must score before the current one is dropped.

ActoricsEnvelopeConfig.h

    

  







  

    SEC.04 / WORKLOADS
    

      
Four workloads. Zero content generation.

      
The same runtime, pointed at four jobs studios already pay for in hand-tuning, QA hours or empty lobbies. Pilots start with the boss.

    

  

  

    

      
SHIPS IN GAMECLIENT

      
Adaptive Boss Encounters

      

        
GENRES
Action-RPG, soulslike, character action

        
WHAT CHANGES
Tempo, counter choice and positioning. Damage and health stay where the designer set them.

      

    

    

      
NEVER SHIPSPRE-RELEASE

      
Headless Exploit Finder

      

        
WHAT IT DOES
Runs your encounter against scripted player habits offline and flags any tactic a player could repeat forever.

        
HOW IT RUNS
Headless builds on your own machines. Nothing ships, no cloud training cluster.

      

    

    

      
LIVE SERVICEDEDICATED SERVER

      
Human-like Lobby Fill Bots

      

        
GENRES
Battle royale, tactical shooters, extraction

        
HOW IT RUNS
On the dedicated server with standard actor replication. No cloud bill per match.

      

    

    

      
ASYNC MULTIPLAYERPROFILE

      
Adaptive Ghost Profiles

      

        
GENRES
Racing, fighting, speedrun

        
WHAT IT IS
A player's style as a portable profile. Race a live opponent that drives like your friend, not a frame-by-frame replay.

      

    

  







  

    SEC.05 / TESTBENCH
    

      
Testbench protocol.

      
Nothing below is estimated. The first table is how each character will be measured in the engine: a value appears when the testbench produces it, until then the slot stays empty. The second is a measured run of the core on abstract subtasks, not characters.

    

  

  

    
PER-CHARACTER ENGINE TESTBENCHNOT YET MEASURED

    

      
CHANNELSTATUSMETHOD

      
        
CPU time per agent tick
          
— . — — µsNOT YET MEASURED

          Engine profiler, mean and peak per frame

        
Scaling to 50+ active agents
          
— . — — msNOT YET MEASURED

          1 / 10 / 50 agents, same scene

        
Memory per character
          
— — KBNOT YET MEASURED

          Module plus learned state, measured size

        
Learned state save size
          
— — BNOT YET MEASURED

          Serialized blob, versioned format

        
Network traffic
          
BY CONSTRUCTION [0 BYTES]

          By construction: the runtime opens no sockets

        
Allocations inside tick
          
NOT YET MEASURED

          Allocation tracing under the engine profiler

        
Cross-platform determinism
          
NOT YET MEASURED

          Record inputs, replay 100×, compare action hash

      
    

  

  

    
CORE SUBTASKS, NOT CHARACTERS · 48 HETEROGENEOUS MATH SUBTASKSMEASURED

    

      
CHANNELRESULTCONDITIONS

      
        
Training time, all 48 subtasks15–20 sStandard CPU, no GPU · 0.2 s per epoch

        
Reload of learned state0.1 msFrom the saved state file, no retraining

        
Accuracy on unseen data99.20–99.26%Balanced accuracy · 10,000 held-out samples · holds whether trained from scratch or reloaded from saved state

        
Learned state file≤ 1 MBAll 48 subtasks in one file

        
Forgetting0%Learning new subtasks does not degrade earlier ones

        
Build targetsx86-64 · ARM64Rust + ISPC

      
    

    
Abstract math subtasks, not game characters. It shows the core learns, keeps what it learned and reloads it from file. Per-character numbers come only from the engine testbench above.

  


  
We don't publish synthetic benchmarks. Every value will be published raw, without smoothing, together with the testbench project.







  

    SEC.06 / INTEGRATION
    

      
Native tooling integration. No standalone runtimes.

      
Built to sit next to your Behavior Tree or StateTree, not to replace them. Every decision is written into the debugger your team already opens. Pause the fight, scrub back two seconds, read why the character did what it did.

    

  

  

    
VISUAL LOGGER · BOSS_01FORMAT ILLUSTRATION

    

      

        
        
        
        
      

      

        

          
          
ANTICIPATIONCOMMITMENT → RECOVERY

        

        


10:30:22.140 [PERCEPTION]   PLAYER SPAM DETECTED: ROLL_LEFT (COUNT: 4)


10:30:22.142 [DESIRED]      FIRE ACTION: SWEEP_LOW_COUNTER


10:30:22.143 [ENVELOPE]     PASSED: REACTION_TIME 180ms >= FLOOR 150ms


10:30:22.144 [COMMITMENT]   LOCKED: POINT_OF_NO_RETURN REACHED

        

      

    

    
Illustration of the log format.

  

  
Plugin design. In development; the first evaluation builds go to pilot studios.

  

    
UNREAL ENGINE 5Native C++ module. StateTree tasks. Trace streams into Visual Logger.

    
UNITYC# bridge to the native runtime. Scene View gizmos and Profiler markers.

    
CORERust + ISPC. x86-64 and ARM64.

    
DEPENDENCIESNo external virtual machines. No Python stack.

  







  

    SEC.07 / LICENSING
    

      
Licensed like engine middleware.

      
The runtime MVP is ready and pilot slots are open. Every engagement starts with a pilot on one encounter, measured on your own profiler. A per-title license follows only if the numbers hold. PC and mobile first.

    

  

  

    

      
STEP 1PILOTS OPEN

      
Design Partner Program

      
Fixed fee per pilot

      
    
        
  • For studios in pre-production or on a vertical slice
    • 
        
  • Integration with the author of the technology
    • 
        
  • One encounter calibrated, blind playtest report delivered
    • 
      

    

    

      
STEP 2COMMERCIAL

      
Production Engine Integration

      
Fixed fee per title, or optional 1% royalty

      
    
        
  • Unreal Engine 5 and Unity plugins
    • 
        
  • Designer Envelope editor
    • 
        
  • Visual Logger and Profiler hooks
    • 
      

    

  


  

    
PILOT SPECIFICATION REQUESTpilots@actorics.com

    

      

      

        

      

        

      

    

    
> Opens your email client with the request filled in. No client? Write to pilots@actorics.com.