Answer Assurance
Source-grounded answers, claim checks, refusal tests and an exportable evidence packet for every relied-on response.
Inspect the evidence chain →Solutions
Useful AI, bounded by evidence.
ANULUM does not begin with a preferred model or appliance. We begin with the confidential work, the custody boundary and the proof required before a professional may rely on the result.
01Custody02Acceptance03Change04Exit
Five bounded solution lines
Each can stand alone. Components are selected only after the workload and evidence gates are known.
Governed Agents
Optional local agents with named identities, tool allow-lists, human gates, budgets, receipts and immediate stop controls.
Inspect the control plane →Evidence Memory
Versioned local retrieval and durable context whose sources, changes, retention and deletion can be reconstructed.
Inspect the memory boundary →Secure Document Intake
OCR, classification, source hashes, duplicates, authority, retention classes and deletion evidence before retrieval begins.
Inspect the intake path →Model Stewardship & Exit
No silent upgrades. Model and licence review, regression gates, signed promotion, rollback and client-owned exit artefacts.
Inspect controlled change →Private AI Fit Check
Generate a local planning record before disclosing documents or committing to a paid review.
Run the fit checkEngineering services
Build the system—and test the boundary.
ANULUM also undertakes bounded software and security work when the problem extends beyond a private-AI product line. Both services use the same written scope, evidence, handover and ownership discipline.
Custom Software Engineering
Internal tools, AI integration, scientific computing, APIs, performance work and maintainable operations for confidential or technically difficult workflows.
Inspect the engineering contract →Security Engineering Services
Authorised reviews and testing, code analysis, hardening, incident readiness, AI security, remediation and controlled retest evidence.
Inspect the security boundary →Availability update · 29 July 2026
Kimi K3 Sovereign Evaluation is now scopeable.
Official weights and serving guidance unblock a bounded rack-scale engagement: checkpoint custody and licence review, workload and tool-schema evaluation, eight-accelerator or multi-node architecture, site envelope, and a documented stop/go decision. It does not include an unmeasured production-performance promise.
One system
Infrastructure is only one layer.
A credible private-AI system joins archive quality, identity, model serving, source-grounded retrieval, evaluation, recovery and operating ownership. A fast model on a large GPU does not prove that the answer is supported, authorized or recoverable.