On this page
Define the synthetic artifact and its source chain
The source dates shown in this guide identify documentation checks on 22 October 2026, not original publication dates. Follow each citation for the risk and ethics vocabulary being discussed. Begin with a controlled release id, format, schema or ontology version, source artifact references, generation activity, configuration reference, filter stages, and evaluator. Do not place private source rows, customer payloads, secrets, or sensitive examples in a public synthetic fixture.
| Lineage item | Synthetic record | Unknown if absent |
|---|---|---|
| Source set | source-release-03 reference | Which material influenced generation? |
| Generation activity | activity-synth-11 | Which process created the release? |
| Configuration | config-ref-11, content withheld | Which settings and model version? |
| Transform stages | filter, dedupe, map | What changed after generation? |
| Output release | synthetic-release-04 | Which artifact was evaluated? |
| Agent or role | synthetic owner role | Who supplied or approved the record? |
| Intended use | evaluation fixture only | Which use boundary applies? |
A lineage record should distinguish source use from source copying. Record the relationship and evidence location without claiming that the output cannot reveal source information. A synthetic label is not proof that the source was removed, and a new identifier is not proof that a person or organization cannot be inferred.
Measure coverage as a set of declared checks
Do not summarize a synthetic release as representative without defining the population, slices, comparison variables, and evaluation method. Record which checks were run and which were unavailable. A release can resemble an aggregate while missing rare, boundary, multilingual, temporal, or minority patterns. The unknown state belongs in the result.
Synthetic coverage worksheet
{
"release_id": "synthetic-release-04",
"source_reference": "source-release-03",
"slices": ["common-case", "edge-case", "time-window-2"],
"checks": [
{"name": "schema_shape", "status": "observed"},
{"name": "rare_slice", "status": "unavailable"},
{"name": "source_reconstruction_probe", "status": "not_run"}
],
"limitations": ["coverage for rare slices is unknown", "privacy-related evidence is method-specific"],
"disposition": "conditional"
}| Dimension | Evidence to retain | Safe wording |
|---|---|---|
| Schema | Fields, types, constraints, version | Structure was checked for the named release |
| Distribution | Declared variables and comparison window | Observed comparison is scoped |
| Subgroup | Named slice and evidence status | This slice is observed or unknown |
| Temporal | Source and generated time windows | Time coverage was or was not checked |
| Utility | Task-specific evaluation record | Result applies only to this evaluation |
| Disclosure risk | Method, probe, and unresolved gaps | No automatic privacy conclusion |
Write limitations that survive a favorable fixture
NIST guidance is a reminder to connect claims to a method and threat model. A synthetic release can carry generation uncertainty, inherited source bias, missing subpopulations, memorized patterns, or new artifacts introduced by filtering. Record the evidence that supports each observation and avoid turning one pass into a universal claim.
| Observed fact | Tempting overclaim | Qualified disposition |
|---|---|---|
| No exact duplicate in a sample | The output contains no source material | Report sample scope and unknown remainder |
| Schema matches source | The output is representative | Keep schema and coverage checks separate |
| Common slice resembles source | All groups are preserved | Name unchecked slices |
| No probe was run | The data is private | Privacy-related evidence unavailable |
| A utility test passed | The data is fit for every use | Name task, slice, and release |
| A filter removed records | Risk was eliminated | Record filter scope and residual unknowns |
The UK Government, (opens in a new tab) guidance supports considering intended use and mitigations. Use that as a prompt for an owner-specific review, not as a replacement for a privacy, security, or legal process. If the source population or generation settings are unavailable, say so plainly.
Use a lineage and limitation decision table
A useful review ends with a disposition and a next evidence request. Keep the generated artifact, lineage chain, evaluation observations, and owner decision as separate records. The table uses invented identifiers and avoids raw source examples.
| Check | Evidence state | Next action |
|---|---|---|
| Source reference | linked / partial / unavailable | Request controlled source reference |
| Generation activity | versioned / descriptive only / unknown | Record workflow and configuration |
| Schema and ontology | matched / changed / unknown | Version the mapping and hold ambiguity |
| Coverage slices | observed / unobserved / unavailable | Name required evaluation slices |
| Utility evidence | task-specific / not run | Limit wording to the checked task |
| Privacy-related evidence | method-specific / unavailable | Escalate to the appropriate owner process |
| Residual limitations | listed / incomplete | Keep release conditional until reviewed |
Keep synthetic-data findings scoped and qualified
This review can document lineage, declared generation stages, checked slices, unavailable evidence, and residual limitations. It cannot establish that data is private, unbiased, representative, accurate, lawful, authorized, or suitable for every downstream use. Unknown evidence remains unknown until the owner supplies a bounded check.
Pair this guide with the data provenance log for artifact history and annotation label ontology versioning for generated labels and definitions. After supported non-health scope, authority to describe the collection, and a controlled lineage packet are established, the qualified appraisal request path may be a next preparation step. It does not generate data, certify privacy, create a quote, match a buyer, issue a legal opinion, or grant a license.
Takeaways
- Record source, generation activity, configuration, transformations, output identity, agent role, and intended use.
- Report coverage and utility by named slices and tasks, with unavailable checks visible.
- Do not treat synthetic labels, new identifiers, or one favorable fixture as a privacy guarantee.
- Keep inherited bias, generation uncertainty, and residual limitations in the disposition record.
- Separate lineage evidence from privacy, legal, ownership, and downstream-use decisions.