Status: AUTHORING LAW / STATIC DOC / RUNTIME PROOF NOT IMPLIED Scope: AI director, creature cognition, navigation, flocking, stimulus memory, encounter pacing, pathfinding, and AI proof gates.
- First-20 moment: first shallow exit and hazard beat where creature pressure, sonar uncertainty, and Director pacing make the route feel alive without unavoidable failure.
- Route blocker removed: prevents AI work from drifting into random patrols or late-game-only brains before the opening swim, resource, tool, and return-path decisions are legible.
- Proof class: STATIC_DOC hook only; acceptance still requires encounter capture, Director/cognition telemetry, profiler/GC evidence, and first-route fairness proof.
AI exists to create pressure, doubt, pursuit, avoidance, ecology, and evidence. It does not exist to run expensive brains everywhere. HECTON-8 rejects random patrol monsters, omniscient enemies, fake intelligence hidden behind jump scares, and AI that spends frame time without changing player decisions.
The player should infer AI through sound, light response, route pressure, disturbed ecology, sonar uncertainty, and remembered stimulus, not through creatures sprinting directly at the camera because a trigger fired.
AI has explicit lanes:
- Director pacing: cold cadence, usually 1 Hz or lower, owns stress, tokens, phase, and spawn permission.
- Cognition: batched Burst jobs over active entities, owns drives, memory, state choice, and intent.
- Navigation: path requests, threat-grid snapshots, flow fields, funnel smoothing, and path validity.
- Flocking/swarm: GPU or data-oriented spatial hash, owns group motion only when swarm behavior matters.
- Presentation: animation, audio, VFX, and UI consume AI intent; they do not mutate cognition truth.
No AI system may own a private Update loop for gameplay truth. No AI system may search scene objects in hot paths.
The AI Director controls pressure curve, not cheap surprise. It uses tokens, phase timing, route context, player stress, oxygen/power/clarity state, depth, noise, light, and recent encounter memory.
Required phases:
- buildup;
- peak;
- decay;
- relax.
The Director must never spawn threats just because the game is quiet. Quiet is a resource. The correct question is what physical pressure the player is already under and whether another stimulus improves or flattens that pressure.
Creature cognition must use bounded drives and state flags:
- hunger;
- fear;
- fatigue;
- threat;
- territory;
- curiosity;
- injury;
- stimulus memory.
Stimulus types must be explicit: acoustic, visual, chemical, electrical, pressure, blood, light, tool heat, hull vibration, and sonar ping. Memory is a ring buffer, not an unbounded list.
Creatures must have readable intent:
- ignore;
- investigate;
- stalk;
- flee;
- feed;
- defend;
- lure;
- attack;
- retreat.
If a creature has only idle, chase, and attack, it is not a HECTON-8 creature.
Navigation must respect the medium. Underwater AI is 3D, occluded, current-influenced, and sonar-limited.
Required:
- non-negative path costs;
- threat-grid or voxel snapshot reads instead of live physics raycasts in jobs;
- A* only where needed;
- weighted A* for distance LOD;
- flow-field fallback for distant entities;
- funnel/string-pulling smoothing for corridor paths;
- route invalidation when doors, floods, voxel edits, or pressure states change.
Forbidden:
- Bezier smoothing that cuts through geometry;
- live
Physics.Raycastinside path jobs; - direct path recalculation every frame;
- AI that ignores doors, flood, light, sound, or pressure state.
Swarms are atmosphere only until they affect visibility, route, threat, or resource decisions. Background swarm motion belongs on GPU or low-cadence data jobs.
Rules:
- fixed-capacity spatial hash;
- overflow discard or load-shed, never stall;
- behavior state packed into bitfields;
- group response to light, sound, flow, and predator proximity;
- no per-entity GameObject logic for swarm members;
- no collision truth for decorative swarm particles.
Evidence class: STATIC_SOURCE only. Compile, Unity import, Play Mode, profiler, GC, encounter capture, and player-build proof remain PENDING VERIFICATION.
| Runtime | Owner / boundary | Static route | GlobalQualityWeight consequence | Missing proof |
|---|---|---|---|---|
Assets/_Project/Scripts/HectonBoidController.cs |
Hecton8.AI.GPU, DataVault owner SystemID.AIEcology; GPU swarm/flocking presentation owner. It does not own creature cognition, collision truth, save identity, or Director authority. |
Implements ITickable, IUpdatable, ILateFrameTickable, and hot-swap listener; registers late-frame environment route. Uses ping-pong boid GraphicsBuffers, spatial grid buffers, visible index/indirect args buffers, Graphics.RenderMeshIndirect, GPU frustum culling, and owner-local material state. Owns a 300-frame DataVault black-box ring at BufferID 71979 with dump target Docs/AgentLogs/Dump_1301_Boids.bin. Consumes SignalBus<AcousticPingSignal> frame snapshots and caches IFoveatedSimulationDirector/flow read models through GlobalRegistry hot-swap. |
Reads MathLodRuntimeConfig.GlobalQualityWeight or HomeostasisBrain.GlobalQualityWeight; derives social LOD weight and foveated VAT time scale. Foveated tier may freeze/peripheral-scale presentation, but creature truth, player fairness cues, and acoustic signal meaning must not change by quality tier. |
No GPU profiler, GCMonitor, RenderDoc/Frame Debugger, swarm capture, signal overflow proof, black-box dump artifact, or gameplay fairness proof was provided by this static audit. |
AI may be hostile, not dishonest.
Required fairness cues:
- acoustic cue before close threat;
- partial silhouette or sonar read before lethal contact where possible;
- readable reason for escalation;
- safe explanation after failure through evidence or black-box trail;
- route pressure that can be answered by player tools.
The Director must not chain unavoidable failures. If oxygen, power, clarity, route, and enemy pressure all collapse together, the system must create a readable recovery path or explicitly enter catastrophic mode.
GlobalQualityWeight scales AI cadence, active presentation density, optional memory richness, flock sample density, and Director staging breadth. It never changes gameplay truth ownership, perception fairness, save identity, command authority, or the meaning of a creature state.
Compact uses fewer active cognition entities, wider cold tick cadence, flow-field steering, lower swarm density, and stronger audio/UI telegraphs. Middle uses full local cognition with conservative Director tokens. High adds richer memory and ecology. Ultra adds denser encounter staging and more secondary swarms, not omniscient or unbounded brains.
AI work must provide:
- Director phase, token budget, and cadence;
- active cognition count and maximum admitted count;
- path request count, path failure count, and load-shed behavior;
- stimulus memory schema and capacity;
- black-box fields routed through
telemetry.md; - Compact and High encounter capture when player-facing behavior changed;
- profiler/GC proof for runtime AI changes;
- fairness cues: what the player heard, saw, or inferred before danger escalated.
If these artifacts are absent, the AI claim remains PENDING VERIFICATION.
Reject:
- random patrol routes as final behavior;
- omniscient player tracking;
- AI spawns with no token budget;
- pathfinding with live scene searches;
- per-creature managed allocations;
- swarms built from individual GameObjects;
- AI reports without active counts, tick cadence, token budget, path cost, and load-shed behavior;
- horror that comes only from a scripted scream.
AI is accepted only when it is bounded, readable, data-owned, fair under pressure, scalable across hardware, and capable of creating dread through systems instead of cheap triggers.