CR Pipeline Authoring
AI · Collective Reasoning
CR Pipeline
Authoring
Collective Reasoning 뉴럴 파이프라인 계약을 설계하고
.crpipe 패키지로 Unity에 임포트합니다.
SCROLL
Pipeline Spec
Top-level properties that define the overall architecture of a CR pipeline.
Display Name
string
Human-readable identifier for the pipeline asset inside Unity.
Candidate Count
int ≥ 1
Number of candidate actions evaluated per pawn per decision step. Scales the PairEncode batch multiplier.
Latent Dimension
int ≥ 1
Width of the shared latent vector passed between passes. Controls model capacity.
Feature Schemas
Each schema is a newline-separated list of named channels. The channel count directly determines network input widths.
Actor Feature Schema
→ PawnEncode input
hp · energy · stress · danger
action0…3 · faction0…3
socialPressure · aggression · fear · focus
isGroup · groupSize · groupSpread
velocityX · velocityZ
action0…3 · faction0…3
socialPressure · aggression · fear · focus
isGroup · groupSize · groupSpread
velocityX · velocityZ
Relation Feature Schema
→ PairEncode input (supplemental)
Spatial / social relationships
between actor and candidate target
between actor and candidate target
Knowledge Feature Schema
→ State input (supplemental)
World-state awareness channels
fed into the State pass
fed into the State pass
Action Catalog
→ Action output labels
One action name per line.
Output dimension = action count.
Output dimension = action count.
Neural Passes
Contracts are derived automatically from the spec. Each pass defines its name, batch multiplier, input channels, and output channels.
01
PawnEncode
in: actorFeatures → out: latentDim
Encodes per-pawn state into a latent representation. Batch ×1.
02
PairEncode
in: latent×2 + relationFeatures → out: latentDim + 1
Encodes pairwise relationships between actor and each candidate. Batch ×candidateCount.
03
State
in: latent×3 + knowledgeFeatures → out: latentDim
Aggregates pairwise encodings with world-state knowledge into a global context vector. Batch ×1.
04
Action
in: latentDim → out: latentDim
Produces the final action logits from the context vector. Output mapped to Action Catalog. Batch ×1.
Export
Export
.crpipeClick Export .crpipe in the toolbar to serialise the current pipeline spec and derived pass contracts to a binary package. Import the file into Unity to instantiate a CrPipeline asset.
1 Author pipeline spec and feature schemas in the web tool
2 Review derived pass contracts in the Pipeline Spec panel
3 Click Export .crpipe → file saved via browser File System Access API
4 In Unity, import the
.crpipe asset and assign to a CollectiveReasoning component