Practical guide

What evidence shows data classification is operating?

A written classification policy shows intended design. Operating evidence shows whether information was actually discovered, assigned an authorised class, labelled or handled accordingly, reviewed, and corrected when the process failed. The distinction matters because a current policy and a configured tool can coexist with unclassified repositories, inconsistent labels, and unresolved.

By Two Prune

A written classification policy shows intended design. Operating evidence shows whether information was actually discovered, assigned an authorised class, labelled or handled accordingly, reviewed, and corrected when the process failed. The distinction matters because a current policy and a configured tool can coexist with unclassified repositories, inconsistent labels, and unresolved exceptions. This guide provides an evidence test, not a legal or compliance determination. Requirements vary by jurisdiction, contract, information type, and system context. Use current specialist review for those decisions.

State a testable control objective

Begin with one sentence that links population, action, rule, and period:

> During the quarter, in-scope documents in the approved collaboration repositories are assigned a permitted classification by an authorised rule or person, retain the classification attribute through defined transfers, and have exceptions reviewed by the data owner within the stated service target.

This is narrower than “sensitive data is protected.” It identifies what should leave evidence. NIST SP 800-60 provides a method for mapping information types and information systems to security categories based on potential impact. NIST SP 800-53 control AC-16 addresses the association of security and privacy attributes with information, permitted attribute values, interpretation, and retention. These sources support testing both the classification decision and the attribute that carries it.

Build an evidence chain

Collect evidence in seven linked layers:

Keep the layers separate. A screenshot of configured labels does not prove that the full population was scanned. A scan total does not prove that the assigned value was correct. A closed ticket does not prove that affected objects were remediated.

  • Authoritative rule: approved taxonomy, permitted values, handling rules, effective date, and owner.
  • Population: system inventory and a reproducible list of in-scope objects for the period.
  • Procedure: how classifications are assigned, changed, inherited, challenged, and removed.
  • Configuration: enabled rules, label definitions, repositories covered, identities authorised, and deployment history.
  • Transactions: object-level label events, rule matches, user overrides, transfers, and timestamps.
  • Exceptions: failed scans, unsupported file types, override reasons, overdue reviews, and remediation.
  • Oversight: owner review, trend analysis, decisions, and next review date.

Reconcile the population before sampling

Sampling only helps after the reviewer knows the population. Reconcile at least three totals:

| Population total | Source | Test | | --- | --- | --- | | In-scope repositories | Approved system inventory | All expected repositories appear in scan coverage | | Eligible objects | Repository export | Scan output plus documented exclusions equals eligible objects | | Classification events | Audit/event store | Events reconcile to labelled, unchanged, failed, and excluded objects |

Investigate rather than netting away differences. A repository export of 48,200 eligible files and a scan report of 47,900 is not “99.4% complete” until the missing 300 files are identified. They may be unsupported, newly created, inaccessible, duplicated, or omitted by configuration. Each explanation has a different risk and owner.

Then select a purposeful sample that covers high-impact classes, each assignment route, recent changes, transfers, overrides, and exceptions. Add a random component if the objective is to estimate an error rate, and document the sampling frame and limitations. Do not extrapolate from hand-picked successful examples.

Test each sampled object

For every object, retain an evidence row:

| Test field | Evidence question | | --- | --- | | Information type | What content or business process makes the object in scope? | | Assigned class | Is the value permitted by the current taxonomy? | | Basis | Which rule, owner judgment, or inherited attribute assigned it? | | Attribute trace | Can the label be linked to the object and relevant transfer event? | | Handling | Did access, sharing, retention, or other defined treatment match the rule? | | Change control | Was any reclassification authorised and logged? | | Exception | Was a failure or override recorded, reviewed, and remediated? | | Review date | Is the evidence current for the stated period? |

Record pass, fail, unresolved, or not applicable for each field. An unresolved test is not a pass. Preserve the source path, timestamp, reviewer, and exact reason so another person can repeat the test.

Worked example: quarterly evidence for a collaboration repository

Consider a synthetic quarterly review of a product team's collaboration repository. The permitted values are Public, Internal, Confidential, and Restricted. Automated rules can propose Confidential or Restricted, but a named data owner approves Restricted overrides.

The inventory lists 12 repositories; the scanner covers 11. The missing repository was created during a migration and contains 420 files. Across the covered repositories, the source export lists 18,640 eligible files. Scan results contain 18,500 objects, with 90 unsupported files and 50 unexplained omissions. The review therefore opens two exceptions: missing repository coverage and 50 unreconciled objects.

The reviewer selects 30 objects: ten rule-assigned high-impact items, five user overrides, five transferred files, five recently reclassified items, and five randomly selected Internal items. Results are:

The evidence record does not say “the control is effective.” It says 26 of 30 sampled objects passed the defined tests, three failed, and one was unresolved; 420 files were outside scanner coverage and 50 objects were not reconciled. The accountable owner blocks the unlabelled transfer route, adds the missing repository, reconstructs the 50-object difference, and schedules retesting. The review date and each exception owner are retained.

  • 26 have an authorised value and a traceable assignment event;
  • two Restricted files were shared through a route that did not preserve the attribute;
  • one override lacks the required owner approval;
  • one Internal file is unresolved because the source content was deleted before review.

Distinguish design, operation, and outcome

Report three conclusions separately:

NIST's current Cybersecurity Practice Guide draft on data classification describes discovering, identifying, and labelling sensitive unstructured data with classification technologies. Because it is an initial public draft, use it as implementation context rather than a final requirement. Current system records and approved organisational rules remain the evidence for a local operating conclusion.

Two Prune may assess how a participant identifies relevant sources, tests a population, handles exceptions, verifies AI-assisted analysis, and communicates uncertainty in a simulated task. It does not establish legal compliance, certify a control, or replace a data owner or qualified reviewer.

  • Design evidence: the taxonomy, procedure, roles, and configuration define how classification should work.
  • Operating evidence: reconciled populations, event records, samples, and exceptions show what happened during the period.
  • Outcome evidence: downstream access or handling results require their own tests and cannot be inferred from label presence alone.

Sources

NIST SP 800-60 Volume 1 Revision 1, Guide for Mapping Types of Information and Information Systems to Security Categories

  • NIST SP 800-53 Revision 5, Security and Privacy Controls for Information Systems and Organizations
  • NIST SP 1800-39 Initial Public Draft, Data Classification Practices

Sources: [1] [2] [3]