Discovery proposes. People decide.
Discovery reads what your existing tools already see about your estate, compares it with the CIs in your CMDB, and turns every difference into something a named person reviews and decides. It never quietly rewrites the CMDB, and every decision lands in the tamper-evident audit log.
Preview, not yet field-tested. Discovery is built and has been tested with simulated data only; no real scanner, Intune or cloud tenant has been connected in a customer environment yet. Run a pilot in a test tenant with two independent sources before you rely on it for production decisions or audit evidence.
The answer to “how do you know your CMDB reflects reality?”, with evidence behind it.
Sources on a safety ladder
A new source starts disabled and moves through connection test and observe-only before it can become active. Observe-only stages evidence and changes nothing. A run that arrives with a different scope, manifest, field or data class than you approved is refused and the source goes to quarantine. A run that would exceed its approved change cap stops before the first write.
Matching on strong identifiers, never on names
Each observation resolves to a CI by strong identifiers such as a device ID, serial, MAC or FQDN, never by name similarity. Where two CIs claim the same identity, nothing is applied and the pair goes to the CMDB reconciliation queue. Merging is always a human decision.
Per-attribute precedence
For each attribute, a policy decides which source may set it: Intune for the OS build, for example. An authoritative source that has gone stale produces a review instead of an overwrite, and an unranked source never wins. Owner, technical owner, criticality and critical-function support are manual-only for every source.
Soll/Ist deviations with an owner
When an observation contradicts the CMDB, the refusal becomes a deviation: the CI, the attribute, what the CMDB says and what was observed, with a severity, a due date and the CI's owner. Due dates run on the wall clock, because a wrong CMDB is wrong at night too. Deviations never close themselves; closing one needs a person, a decision and a written reason.
Independent verification you can seal
A CI counts as independently verified only when a different origin system confirms it with authenticated data. A scanner cannot confirm what the scanner itself discovered. The attestation reports verified, stale and single-sourced CIs per class, states its denominator in words, excludes any CI with an open deviation, and is sealed as an immutable, hash-chained artifact.
Ghost assets proposed, never deleted
When successful, complete runs stop seeing a CI, Discovery counts them. Failed, partial or observe-only runs never count. At the grace threshold it opens a retirement candidate, and only if no other origin system still sees the CI. Retiring keeps the record, its history and its provenance.
How this connects to regulation
Where this connects
What CMDB owners and auditors ask.
See TENSOR running on your own estate.
30 minutes, screen-shared, no slides. Bring the process you are least sure we handle and we will start there.