01 / PURPOSE
Perspective, not omniscience
An away-team agent should not receive the complete state of the simulation. It should see from its own camera, notice only some details, remain uncertain about others, and carry those experiences into later decisions.
Ambiguity is not a defect. It creates curiosity, disagreement, investigation, and individual experience.
The perception system reports visible evidence. The agent combines that evidence with its role, goals, instruments, and memories. A security officer and an engineer can therefore reach different hypotheses from the same scene without changing what the camera actually recorded.
02 / DATA FLOW
Observe only when something matters
- 01
Capture a viewpoint
EVONC renders an image from one agent's camera and attaches only the instrument readings that agent is permitted to know.
- 02
Gate the frame locally
Fast image hashes and change detection reject duplicates. A deliberate look, a mission event, or meaningful visual change opens the gate.
- 03
Perceive with a replaceable worker
A separate Python process runs a compact local vision model. The game engine knows only the file-and-JSON contract, not the model vendor.
- 04
Record evidence and uncertainty
The result contains observations, approximate locations, confidence, limitations, and unknowns. It does not declare guesses to be world truth.
- 05
Let the agent decide
Role and memory turn evidence into a hypothesis, speech, or action: inspect, move closer, retreat, ask another agent, or remember the encounter.
03 / INDEPENDENCE
Compute is a budget, not a dependency
Local by default
Capture, change detection, storage, and the initial vision model operate without a network service.
Event-driven
The system analyzes selected moments instead of continuously sending every rendered frame through a model.
One shared worker
A bounded priority queue lets one loaded model serve multiple agents while stale background requests can be discarded.
Replaceable backend
Local, test, specialist, and optional remote backends produce the same observation schema.
Compact agent context
Agents reason over small observation records and consult source images only when visual reinspection is worthwhile.
Graceful absence
EVONC continues to operate if the perception worker is stopped, unavailable, or intentionally disabled.
04 / OBSERVATION MODEL
Keep seeing separate from believing
Observed A gray humanoid stands near a cyan luminous object.
Uncertain The object's purpose and the humanoid's intent are unknown.
Engineering hypothesis The object may be an active device worth inspecting.
Security hypothesis The figure may be guarding a hazard or boundary.
Only the first two statements belong to shared visual evidence. Hypotheses remain with the agent that formed them, unless that agent communicates them to the team.
05 / DELIVERY PLAN
Build the smallest useful loop
Stage A
Offline proof
Image in; observation JSON and readable report out.Stage B
Live observer
Agent camera capture, change gate, one local worker, visible status.Stage C
Embodied action
Look, turn, approach, inspect, retreat, report, and request another view.Stage D