Keeping the model current

How STATE3 stays true.

Most CMDBs drift out of date within a year, because nothing in them decides what to do when the automated answer and the human answer disagree. That decision is what we built. Automated where it's safe, human where it matters, and a pipeline that knows which is which.

The upkeep pipeline

Two routes into the graph.

Which route a fact takes depends on where it came from. Stages 02 and 03 apply only to data that arrives through staging.

Get it in.

Connectors run on a schedule and land in staging, as do CSV imports run through the ingestion hub. Entity-specific imports, UI edits and MCP writes go straight to the graph instead.

Make sense of it.

The rules engine reads what arrived and proposes Add, Update, Delete or Merge actions. Each is applied automatically if confidence is high, or queued for review if it isn't.

Trust it.

Uncertain actions queue for review with field-level diffs, and you approve or reject each one. Fields a person has edited are protected from automation.

01 · The problem

Why most tools fall behind.

Auto-import everything from all your systems, and the data goes stale, because nobody trusts it enough to fix it. Require manual entry for everything, and the platform doesn't get adopted. Most CMDB and EAM tools made one of those two choices and lived with the consequences. So we didn't pick. We built a pipeline that decides which changes are safe to apply automatically, and a UI that lets a human resolve the cases the machine isn't sure about.

02 · For the technical evaluator

How it works.

Three subsystems, each relying on the one before it to get its part right. Read to whatever depth you need.

Get it in.

Several paths bring data into STATE3, and they do not all work the same way. Scheduled connectors always write to staging, and so does a CSV run through the ingestion hub. Those are bulk, automated sources that have to be reconciled against what is already there. An entity-specific CSV import, a direct edit in the UI, and an MCP write skip staging entirely: each one is a deliberate act against a known record, so it is validated and written to the graph.

STATE3 ingestion: two routes into the graph Two routes into the STATE3 graph. Route 1 carries the scheduled connectors — Entra ID, Azure DevOps, Azure Resources, Digicert and the Data Feed — together with CSV imports run through the ingestion hub. It passes through a staging layer and then the rules engine, where proposed actions are either applied automatically or queued for human review, before reaching the graph. Route 2 carries entity-specific CSV imports, direct edits made in the user interface, and MCP write tools; these are validated as they are written and go straight into the graph, with no staging and no review queue. The two routes stay separate until they reach the graph. ROUTE 1 · VIA STAGING Entra ID Azure DevOps Azure Resources Digicert Data Feed CSV — ingestion hub ROUTE 2 · DIRECT CSV — entity-specific import Direct UI edits MCP write tools ROUTE 1 Staging layer dump tables Rules engine auto applied · or queued for review ROUTE 2 Validated on write no staging · no review queue STATE3 graph the living model On-ramps Both routes end in the graph
Route 1 passes through staging and the rules engine, where the review queue applies. Route 2 goes straight to the graph.

Connectors (Entra ID, Azure DevOps, Azure Resources, Digicert and the extensible Data Feed) run on a schedule and pull directly from the source system. The CSV import wizard covers everything without a connector: map CSV columns to STATE3 fields, mark anything you don't need as Skip, and upload, either through the ingestion hub or as an entity-specific import.

Every import is tenant-scoped, sanitised against CSV-injection prefixes, row- and column-capped, and written in batches.

Column mapping: CSV import wizard

CSV header Maps to STATE3 field
Asset Tag asset_tag
Make manufacturer
Model model_name
Serial serial_number
Location Skip

Five columns from a source CSV mapped to their STATE3 equivalents, with a Skip option for anything the organisation doesn't track.

Make sense of it.

This stage applies to the staged route. Once data is in staging, the rules engine reads it and produces a list of proposed actions. Each action is one of four types: Add a new entity, Update an existing one, Delete something the source no longer reports, or Merge two records the engine recognises as the same thing. Each proposed action is tagged either Auto (high confidence, so apply it) or Manual (queue it for human review). Confidence is shaped by alias lists, ignore lists, and transformation lists maintained per tenant.

Decision card: rules engine output

Auto applied

Update

Asset 1234 · location:
Hamilton → Tauranga
(alias match)

Manual · queued

Merge

‘SQL Server’
‘Microsoft SQL Server’
(queued for review)

Two proposed actions: one auto-applied on an alias match, one queued for a human to resolve.

Trust it.

Once actions are proposed, three mechanisms make sure what ends up in STATE3 is what your team believes is true.

Review queue

Every uncertain action shows you the proposed change as a field-level diff. You approve or reject it, with a comment if you need one.

Diff card: action review

Update Database Server #DB-014
Current Proposed
Edition Standard Enterprise
Version 2016 2019

Confirmation

When you've looked at a record and you know it's correct, you confirm it. STATE3 records who confirmed it and when. You can confirm a single record or bulk-confirm an entire model in one operation.

Confirmation chip

Confirmed by service.owner@example.org · 2026-04-29

Manual edit protection

Once a person has edited a field, the rules engine never overwrites it.

03 · What relies on it

Lenses, AI answers and board figures all read this data.

The five lenses on the home page rely on the upkeep. So do the MCP queries an AI assistant runs on your behalf, and any number a CIO defends in a board meeting. If the graph is wrong, so is every answer read from it.