Reference

Policies

These are the standing rules the product enforces — not aspirations. Each maps to a concrete enforcement point: an engine check, a blocking gate, or a governed review path. The narrative version, with the reasoning behind each position, is at Design Principles.

Sources & ingest

Policy	What it enforces
Feeders feed, they never author	Application and software inventories carry attribution evidence but can never create device records. Only trusted discovery sources mint new CIs.
Authoritative source owns each attribute	Per-attribute precedence, authored from evidence at onboarding, recorded field-by-field with the values it beat.

Identity

Policy	What it enforces
No hopeful imports	Records below the identity bar are blocked with the specific reason stated — a record you cannot identify is a scheduled duplicate.
Conflicts surface, never absorb	One-serial-many-hosts and one-host-many-serials are flagged for review; silent merging is prohibited.

Classification

Policy	What it enforces
Deployable leaf classes	Automatic classification targets only deployable leaf classes; abstract parents are rejected as rule targets at authoring, deployment, and runtime.
Collision-guarded rules	Rule authoring checks substring predicates against observed values across all sources; consumer-platform rules require a corroborating condition.
Doubt routes to review	Records the rules cannot place with confidence go to a review queue with evidence — never into a catch-all default.

CSDM attribution

Policy	What it enforces
Content evidence anchors confirmation	A placement is confirmed only on evidence found on the device itself; context can corroborate, never anchor; structural defaults never confirm.
Zero false confirms on the evaluation corpus	No engine release ships unless validation demonstrates zero false confirmations; operator confirms and rejects continuously feed the gate’s labels.
Endpoints park honestly	End-user devices attach to the management service that operates them — not force-attributed to business services on weak signals.
Coverage is three numbers, not one	Business-attributed / management-service / unassigned, computed by one shared predicate everywhere it appears.
Emission is staged; capture never is	What ships grows with adoption stage (crawl/walk/run/fly); what the system learns is never gated; promotion is an operator decision on the record.
The workbook is a view	The CSDM 5.0 workbook is generated from the governed model; hand-edited spreadsheets are never a source of truth.

Export

Policy	What it enforces
IRE-ready payloads, never table loads	Everything arrives at the instance as conditioned input to Identification and Reconciliation. Export quality gates block bundles that fail their checks.

Governance

Policy	What it enforces
The system authors, the operator reviews	Configuration is AI-authored with rationale onto the decision record; operator overrides are first-class, attributable, and take precedence.
Recommend, don’t assert	Every recommendation carries a confidence band, evidence, and review status; low-confidence inferences become questions, never silent facts.
Every engine change is regression-proven	Changes ship only after verification against pinned full-run baselines — byte-identical, or the difference is explained.
Data stays home	Single-tenant deployment in your environment; discovery data never leaves it.

Honesty note

Policies differ in how they are held: most are engine-enforced or gate- blocking; a small number are review-doctrine applied by the governed authoring path. We track the difference internally and promote doctrine to engine enforcement as the machinery lands.