ARCHITECTURE
The Nervous System
Four processing layers, from edge to core. Signals enter raw and gain context, reliability metadata, and health annotations as they move inward. Each layer adds something the previous one couldn't.
The metaphor isn't decoration. A nervous system filters stimuli at the periphery, integrates signals in the middle, and makes judgments at the core. billing-brain does the same thing with knowledge.
Each layer is a separate module. They share the in-memory graph but have distinct responsibilities. A search query touches Filter. A "show me how these connect" touches Integrate. A "how much should I trust this?" touches Harden. A "what's degrading?" touches Maintain.
The first thing that happens to a question is filtering. Which nodes in the graph are relevant? The answer comes from three tiers, tried in order.
Direct ID lookup or title substring. If you already know the node's name, this is instant. Highest confidence — you asked for it by name.
The graph includes a classification taxonomy with hundreds of signals drawn from real ticket analysis. If your question contains a signal phrase — "payment failed," "double charged," "stuck in processing" — the taxonomy routes you to the right domain nodes.
Tokenized text search across node IDs, titles, and bodies. The broadest net. Catches what the other tiers miss, with lower confidence.
Every result carries its match tier. You always know how the node was found — exact, taxonomy-routed, or keyword-matched — so you can weight it appropriately. No black-box relevance scores.
Filter finds nodes. Integrate shows how they connect. This layer walks edges, assembles subgraphs, and finds paths between distant concepts.
Expand any node to see what it connects to — outgoing edges (what it grounds, enables, triggers), incoming edges (what grounds it, contains it, complicates it), and the adjacent nodes at each end.
Give it a set of node IDs, get back every edge between them plus cross-edges reaching outside the set. The shape of the connections reveals relationships that aren't obvious from individual nodes.
How does this fact connect to that mechanism? BFS traversal finds the shortest path, ignoring edge direction. The path steps show which edge types link the concepts.
Explore an entire domain — all its nodes grouped by kind, internal edge density, and cross-domain connections. See which domains are tightly coupled and which are islands.
Integration is what turns a bag of relevant nodes into a connected understanding. The graph has structure — this layer makes that structure visible.
The most important layer. Harden doesn't find information or connect it — it tells you how much to trust it.
Pick a fact or thesis. Harden walks incoming grounds edges recursively until it reaches data nodes — observations, signals, metrics. The chain is the provenance. If the chain is empty, the fact is ungrounded.
Strong — backed by fresh data nodes. Degraded — backed by data, but some sources are stale. Ungrounded — no data backing at all. An asserted claim. The verdict is computed, not guessed.
The graph includes explicit contradicts edges where two nodes disagree. Harden surfaces these so you see conflicts rather than getting silently inconsistent answers.
Every answer from billing-brain is either grounded or explicitly flagged as ungrounded. There's no middle state where a claim pretends to be verified when it isn't.
A knowledge graph that can't detect its own decay is a liability. Maintain watches the graph's health and surfaces what's degrading.
Structural warnings grouped by type — verbose bodies, ungrounded facts, unobserved metrics, untested theses, broken references. Each warning type has a count, a description, and examples.
When a metric goes stale, Maintain traces which facts depend on it. A stale metric isn't just a missing number — it's a chain of claims that just lost their foundation.
Asserted facts ranked by connectivity. The most-connected ungrounded facts are the highest-priority debt — if they're wrong, the most other nodes are affected.
Weekly digest data from classified support tickets. Category-level trends show which problem areas are growing, stabilizing, or shrinking. The graph connects trends to the mechanisms that explain them.
Maintain is the difference between a knowledge base and a knowledge system. The base stores information. The system knows when that information is going bad.
A typical interaction touches multiple layers in sequence.
Filter catches "activation failures" via taxonomy routing — it's a known signal in the subscription-lifecycle category. Returns relevant facts, mechanisms, and metrics.
Integrate finds paths between the activation failure nodes and the dunning domain. Reveals shared mechanisms, common entities, and edge types that link the two territories.
Harden traces evidence chains for the key facts. Some are grounded in recent BigQuery observations. Others are asserted claims from code reading — flagged as lower confidence.
Maintain shows that two metrics in the chain haven't been refreshed this week. The facts they support are still structurally sound but epistemically degraded. Time to re-run the queries.
By the time an answer reaches you, it carries its own trust score.