System Architecture

Deterministic Infrastructure.
Zero Hallucinations.

Causel abandons probabilistic text generation in favor of a mathematically bound, single-tenant retrieval architecture. Every probabilistic operation is bounded by a deterministic verification gate. The gates are the moat. The model is the utility.

“We do not summarize the law; we litigate the data. Every claim is bound to a coordinate. Every contradiction is gated by temporal scope and custodian access. Every draft is attacked before it leaves the building.”

Five Technical Pillars

01

Geometric Provenance

Every fact in the Causel knowledge graph is bound to a verifiable bounding box on a specific page of a specific source document. The constrained LLM extractor produces only exact-quote excerpts; each is reverse-mapped to the source PDF’s character coordinate index. Extractions that cannot be reverse-mapped are rejected and never enter the system.

{ "claim": "Deponent denies presence", "bounds": [142.5, 804.1, 305.2, 816.0], "bates": "PLAT-0042" }
02

Bipartite Resolver

Email, Slack, and Teams are unified into a single weighted communication graph. A 1/√N broadcast discount prevents ambient channel noise from corrupting the baseline. An isolation forest then surfaces pairwise communication gaps that exceed a configurable severity threshold.

weight(edge) = 1.0 / sqrt(channel_members) // Direct messages: weight = 1.0
03

Epistemic-State NLI

For every sworn claim, the system extracts the temporal anchor date, generates the inverse hypothesis, and runs a deterministic graph query bounded by custodian access and temporal scope. Only candidates that pass both deterministic gates are sent to the cross-encoder NLI classifier.

Filters: temporal_scope(< anchor_date) AND custodian_access(claimant) → then NLI
04

Adversarial Drafting

As an associate drafts a motion, an actor-critic agent runs in the background. The critic queries the knowledge graph for contradictions, the deterministic judge profile lookup retrieves the vulnerability vector, and the hedge generator proposes a strategic revision. Three-iteration cycle cap.

Cycle: Critic → Judge Lookup → Severity Gate → Hedge Generator → repeat (max 3)
05

Confounder-Gated Bayesian Profiler

Every published ruling is classified as procedural or merits by a deterministic LLM gate. Procedural rulings are blocked from updating the doctrinal vector. Merits rulings are weighted by counsel disparity before Bayesian shrinkage against the prior.

if ruling_basis == "procedural": BLOCK // Standing, jurisdiction, timeliness → no update

Infrastructure Stack

Single-Tenant AWS Cells

Every AmLaw 50 deployment provisions a dedicated, isolated AWS VPC. Your data never touches a multi-tenant database or a shared inference endpoint. End-to-end data locality inside the firm’s cell.

  • Dedicated RDS PostgreSQL (pgvector)
  • Dedicated SageMaker Inferentia Instances
  • Customer-Managed KMS Keys (BYOK)
  • Isolated VPC with no cross-tenant routing

Inference Bifurcation

Reasoning models (Claude Sonnet, Opus) routed through Bedrock to keep inference traffic on-network. Task models split between SageMaker Async with Spot Instances for batch ingestion (scale-to-zero, cost-efficient) and SageMaker Inferentia for real-time UI inference (sub-second latency, always-warm).

Batch Path

SageMaker Async + Spot

Scale-to-zero on idle

Real-Time Path

SageMaker Inferentia

Always-warm endpoints

Full Stack

Compute

  • SageMaker Inferentia (real-time NLI)
  • SageMaker Async + Spot (batch ingestion)
  • AWS Bedrock (Claude Sonnet / Opus reasoning)

Storage

  • RDS PostgreSQL + pgvector (knowledge graph)
  • S3 (document store, encrypted at rest)
  • ElastiCache Redis (session, hot cache)

Orchestration

  • Temporal (workflow engine, checkpoint recovery)
  • LangGraph (adversarial agent framework)
  • EventBridge (async event routing)

Security

  • AWS KMS (AES-256, customer-managed keys)
  • CloudTrail (immutable audit logging)
  • IAM + SSO (RBAC, ethical wall enforcement)

Intellectual Property

The five technical pillars described above are the subject of a provisional patent application filed with the United States Patent and Trademark Office. The specification covers geometric provenance binding, multi-channel broadcast-discounted communication graphs, epistemic-bounded contradiction retrieval, adversarial drafting with cyclical critic-hedge agents, and confounder-gated Bayesian doctrinal profiling.

Patent Pending