Technical Architecture

Four layers of
emergent intelligence

Periphery does not impose structure. It observes data across five embedding spaces, clusters emergent patterns with HDBSCAN, and validates every structure through adversarial coherence scoring. What reaches the analyst has survived the Critic.

01

Ingest Mesh

The collection layer. RSS feeds, REST APIs, file uploads, and custom scrapers funnel through a politeness-aware ingestion system. Rate limiting is per-source. Deduplication happens at the content hash level before any processing begins.

Politeness-aware crawling with per-domain rate limits
Content-hash deduplication at ingestion boundary
Source reliability scoring based on historical accuracy
Automatic language detection and normalization
02

Crystallizer

The core analytical engine. Documents are projected into five embedding spaces simultaneously. HDBSCAN clustering identifies emergent structure without requiring predefined categories. Clusters form, merge, split, and dissolve as new data arrives.

Five-space HDBSCAN clustering (semantic, entity, relational, temporal, geospatial)
Dynamic cluster lifecycle: formation, merging, splitting, dissolution
No predefined ontology required — schema emerges from observation
Continuous re-evaluation as new documents arrive
03

Continuous Critic

The adversarial validation layer. Every emergent structure is challenged. The Critic scores coherence, identifies contradictions, and assigns legibility tiers. Nothing reaches the analyst without passing adversarial review.

Adversarial coherence scoring for every cluster
Contradiction detection across sources
Five-tier legibility gradient: Solid → Defined → Emerging → Haze → Whisper
Epistemological honesty — the system knows what it doesn't know
04

Query Interface

The analyst's surface. Natural language queries are decomposed into analytical operations. Claude synthesizes responses with full confidence rendering — every claim is traceable to source documents with explicit uncertainty markers.

Natural language analytical queries
Claude-powered synthesis with source attribution
Confidence rendering on every assertion
Exportable intelligence products with full provenance

Embedding Architecture

Five spaces, one structure

Every document is projected into five orthogonal embedding spaces simultaneously. Clustering operates across the combined space, capturing patterns invisible to any single dimension.

SEMANTIC

Dense vector representations capturing meaning and context across documents.

ENTITY

Named entity co-occurrence and resolution across multilingual sources.

RELATIONAL

Graph-derived embeddings encoding relationships between actors, organizations, and events.

TEMPORAL

Time-series encoding that captures event cadence, periodicity, and sequence.

GEOSPATIAL

Location-aware embeddings for proximity analysis and regional clustering.

Adversarial Validation

The Continuous Critic

Epistemological honesty is not optional. The Continuous Critic scores every emergent structure for coherence, flags contradictions, and assigns a legibility tier. The system knows what it doesn't know.

Solid
Defined
Emerging
Haze
Whisper

Document Lifecycle

The state machine

Every document progresses through a deterministic state machine. Each transition is observable, auditable, and reversible.

PENDING
ENRICHING
ENRICHED
EMBEDDING
EMBEDDED
CRYSTALLIZED

See the architecture in action.

Request Demo