Reliable knowledge for agentic AI
Three orthogonal layers — calibrated trust, change awareness, targeted revalidation — composed into one reference architecture. Open-source Rust crates. Production-validated research foundations.
of companies abandoned AI initiatives (S&P 2025)
calibrated accuracy via multi-source fusion (KV)
of tests catch 99.9% of failures (Meta TAP)
revalidation for 90%+ queries (Salsa)
The Problem
Models work. The knowledge they operate on does not.
Enterprise knowledge degrades silently: facts go stale at different rates, contradictions accumulate across sources, and when something changes, nothing indicates which downstream decisions are now unreliable.
Calibrated Trust
Every fact must carry a confidence score grounded in evidence — not just vector similarity.
Change Awareness
When a fact changes, the system must know what else is affected across the dependency graph.
Targeted Revalidation
Maintenance must focus on what actually changed — not reprocess everything on a fixed schedule.
No current approach delivers all three simultaneously.
Why Current Approaches Fail
Three patterns, three failure modes
RAG
Ranks by vector similarity — topical relevance, not factual reliability
A three-year-old code comment and yesterday's API spec score identically. No mechanism for confidence calibration, staleness detection, or selective revalidation.
Scheduled Rebuilds
Uniform cadence applied to non-uniform change
Google TAP (2B LOC, ~1 commit/sec): 50% of targets changed <14 times/month; a volatile minority changed hundreds of times. Nightly rebuilds reprocess the stable majority and miss the volatile subset.
Long Context Windows
More tokens = more sources of confidently wrong answers
Models ignore mid-context information ("lost in the middle"). TAP showed that beyond 10 dependency hops, testing produces a 99:1 noise-to-signal ratio.
The Architecture
Three orthogonal layers, one composition
Each layer solves a single dimension using the appropriate mathematical tool: Bayesian fusion, graph propagation, and durability classification. UCM composes them into a unified framework that addresses all three failure modes.
Calibrated Trust
Research Foundation
Google Knowledge Vault (KDD 2014)
Noisy-OR Bayesian fusion
Extracted 1.6B facts from the web using four independent methods. Multi-source agreement compounds confidence. When the system outputs 0.96, empirical accuracy is 0.96.
Enterprise Mapping
A claim confirmed by CRM data, an API response, a support ticket, and a Confluence page — quantify agreement weight, flag single-source claims.
Change Awareness
Research Foundation
Meta Predictive Test Selection (ICSE-SEIP 2019)
Graph-distance impact propagation
Thousands of commits/day, millions of tests. Impact attenuates predictably with dependency-graph distance: shortest-path length is the strongest failure predictor. 33% of tests caught 99.9% of real failures.
Enterprise Mapping
API contract changes → directly dependent docs take a large confidence hit; three hops away, smaller; ten hops, negligible.
Targeted Revalidation
Research Foundation
Salsa Engine — Rust IDE (2019+)
Durability classification + incremental re-compute
Every input classified by expected volatility: stdlib changes every 6 weeks (high durability), source files every few seconds (low durability). On change, only dependent computations revalidate. Result: ~300ms → ~0ms for 90%+ of queries.
Enterprise Mapping
Business rules (quarterly), API contracts (monthly), feature flags (hourly) — each with revalidation frequency matched to actual volatility.
Implementation
7 Rust crates, published on crates.io
Each crate maps to a specific layer of the architecture. All are open source, tested, and documented with architecture notes and research references.
Entity model, SCIP identity scheme, Noisy-OR confidence math, in-memory graph
Append-only event sourcing — Datomic-inspired mutation log with temporal queries
Multi-source ingestion pipeline — Markdown, JSON, YAML with configurable extractors
Bayesian confidence fusion + graph-distance impact propagation engine
Volatility classification, staleness scoring, revalidation scheduler
REST/gRPC API surface — query, ingest, and observe endpoints
CLI for local development, graph inspection, and pipeline debugging
Case Study
Banking: unauthorized refund prevention
When a new compliance rule supersedes an existing refund policy, agents must know immediately — not after the nightly rebuild. UCM propagates the change through the dependency graph and revalidates only the affected downstream decisions.
Policy change ingested
New compliance rule enters via ucm-ingest. The event log in ucm-events records the mutation with full provenance.
Impact propagated
ucm-reason computes graph-distance impact: refund-approval rules take a large confidence hit; indirectly related KYC docs receive a smaller adjustment.
Targeted revalidation
ucm-observe marks only the affected subgraph for revalidation. Stable subgraphs (account structures, historical transactions) are untouched.
Privacy-sovereign context for enterprise agents
UCM handles knowledge reliability. Krava handles privacy and security — identity-less passkey auth, encrypted-at-rest memory, and agentic routing with session isolation. Together, they enable portable, private, reliable context across models and apps.
Get Started
Book a technical consult
We help engineering teams adopt UCM + Krava for their specific workloads — from architecture review to production deployment.
Architecture Review
Map UCM's three layers to your knowledge topology
Proof of Concept
Validate confidence calibration on your data
Production Deployment
Integrate Rust crates + Krava SDK into your stack
Frequently Asked Questions
Common questions
How is UCM different from RAG?
RAG ranks by vector similarity — topical relevance, not factual reliability. UCM adds three layers on top: confidence scoring (Bayesian fusion), change tracking (graph-distance propagation), and incremental revalidation (volatility classification). It can augment an existing RAG pipeline or replace the retrieval layer entirely.
Is UCM production-ready?
UCM is a reference architecture backed by published Rust crates (v0.1.3 on crates.io). Each layer is grounded in production-validated research (Google Knowledge Vault, Meta TAP, Salsa), but the integrated pipeline has not been tested on enterprise data. We offer consulting engagements to validate and deploy UCM for your specific workloads.
What does Krava add?
Krava provides the privacy and security layer: identity-less passkey authentication (FIDO2), encrypted-at-rest memory, and agentic routing with session isolation. UCM handles knowledge reliability; Krava handles data sovereignty. Together they enable portable, private, reliable context across models and applications.
How do I calibrate confidence scores for my data?
UCM's calibration layer uses Noisy-OR Bayesian fusion — multi-source agreement compounds confidence. You configure which sources (APIs, databases, documents, tickets) feed each entity, and the system learns calibrated scores. When the output says 0.96, empirical accuracy should be 0.96.
Can I use UCM with my existing LLM stack?
Yes. UCM is model-agnostic — the Rust crates handle confidence scoring, graph traversal, and revalidation independently of the LLM. The ucm-api crate provides REST endpoints that integrate with any retrieval or generation pipeline. Krava's adapter layer supports multiple backends (Anthropic, Tinfoil, etc.).
What industries is this best suited for?
Any domain where knowledge changes at different rates and errors carry real cost: banking/compliance (policy changes affecting agent decisions), document management (contract and regulatory updates), healthcare (protocol and formulary changes), and enterprise SaaS (API contracts and feature flags).