Architecture / Importance Engine

The layer that decides which signals deserve weight.

Importance Engine sits above governed context and evidence feeds. It compares signals, weighs recurrence and contradiction, and returns either a low-latency score or a richer forensic explanation for human decision support.

Core layers

How Importance Engine is structured

01

Signal intake

Receives governed context, product behaviour, support posture, commercial indicators, local session state, or evidence summaries through explicit boundaries.

02

Weighting and scoring

Uses result types, Success Points, negative signals, confidence, recurrence, and mode-specific scoring to identify what should be acted on first.

03

Decision support output

Returns scores, explanations, dossiers, or response-pressure guidance that can be reviewed, audited, and challenged by the user.

Operating flow

The request path through the product

Context

UCL facts provide business grounding

The engine can use governed context snapshots when enterprise facts need to inform scoring.

Evidence

Forensic mode keeps provenance visible

Evidence, contradiction, confidence, and negative-signal handling are exposed without publishing private scoring tables.

Action

The next move is weighted

The output helps choose whether to prioritise, reassure, challenge, escalate, or investigate further.

Governance

Boundaries the architecture must respect

Privacy

Local-first emotional boundary

Klopp-style conversation and weight handling is documented as local-first. Cloud analytics are aggregate-only and must not include prompts, responses, raw weights, or context facts.

Audit

Forensic explanation path

Forensic mode is positioned around provenance, factor explanations, confidence, and contradiction handling.

Control

No silent sensitive automation

The public architecture frames the engine as decision support, not as an unaccountable automated decision maker.

Integration points

Where it connects to the wider stack

UCL

Governed context

Pull facts from Scout or Fortress when business context is needed.

MCP/API

Agent-facing tools

Expose scoring and dossier paths to controlled investigation and assistant workflows.

Local LLM

Loopback enrichment

Where local model use is enabled, the design keeps sensitive runtime boundaries explicit.

Evidence boundary

What this page proves, and what it does not claim yet

Built

Local engine surfaces

The repo contains Rust crates for core scoring, Klopp, Forensic, MCP/API, diagnostics, and privacy documentation.

Deferred

Native and model proofs

Local model, native vector-store, and full benchmark proofs are not claimed as production-ready from this marketing page.