Systems and records
Synthetic data provenance disclosure: label fields observed, inferred or generated
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Label origin at the field level, and at the value level wherever origin varies. Observed fields were captured from a real source and only reformatted or redacted; inferred fields were computed or predicted from other fields by a named rule or model; generated fields were produced by a generative process. Each label carries its source, method, inputs, generator and validation status, and a validator rejects records whose labels contradict each other.
Provenance trail
Synthetic data provenance
Key takeaways
- A dataset called real can hold imputed and model-labeled fields; a dataset called synthetic can hold observed values copied from seeds.
- Three origin classes cover most cases: observed, inferred and generated, with a flag for values reported by a person rather than captured.
- Redaction and format changes keep a field observed, but the transformation must be listed.
- Set origin per value whenever a column mixes classes, such as a code that is observed for most rows and imputed for the rest.
- A validator should reject contradictions, such as an observed field with a generator reference or an inferred field with no inputs.
- Evaluate generated fields against untouched observed records, and report which rare categories the generated share misses.
How do you label observed, inferred and generated dataset fields?
Give every field an origin class in the schema, and switch to a per-value label wherever a single column mixes classes. Then attach to each label the evidence that makes it checkable: where an observed value came from, which rule or model produced an inferred one and from which inputs, and which generator, version and seed produced a generated one.
The definitions need to be strict enough that two people label the same field the same way. Observed means captured from a real source system, sensor or document at the time of the event, changed only by formatting, normalization or redaction. Inferred means computed or predicted from other fields, whether by a deterministic rule, an imputation method or a classifier. Generated means produced by a generative process, such as a language model, a simulator or a statistical synthesizer, without a one-to-one real value behind it.
Buyers need this because the same dataset can be described honestly as real and still contain many values no one observed. The disclosure turns a marketing label into something a training or evaluation pipeline can filter on.
Start with the fields a model will be scored on, not the whole schema. If a triage assistant will be evaluated on priority and resolution code, those two columns need the strictest labels and per-value scope first; descriptive fields that only feed training can follow. That order keeps the first version of the disclosure small enough to finish and useful on day one.
Why is a dataset-level synthetic label not enough?
Because most datasets offered for training are mixtures, and the mixture sits inside records, not between them. A support ticket export can hold observed messages, a priority an agent typed, a sentiment score from a model and a summary written by a language model, all in one row. Calling that dataset real or synthetic is wrong either way.
Dataset documentation already asks the right question at a coarse grain. Datasheets for Datasets asks whether the data for each instance was directly observable, reported by subjects, or indirectly inferred or derived from other data, and whether reported or inferred data was validated. MLCommons' Croissant RAI vocabulary gives dataset-level places for the collection type, the imputation protocol and, where relevant, the method used to generate synthetic data.
Those answers are written once for the whole dataset, usually as free text. They tell a buyer that imputation happened, not which values were imputed. A field-level disclosure turns the same answers into structured labels a filter can act on, so an evaluation set can be restricted to observed values and a training set can weight or exclude generated ones.
Three origin classes and where each is usually mislabeled
The classes are deliberately few. More classes produce more disagreement between labelers, and the extra detail belongs in the method and input attributes rather than in the class name. The common mislabels are worth knowing, because they are what a reviewer should look for first.
| Origin class | Definition | Examples | Common mislabel |
|---|---|---|---|
| Observed | Captured from a real source at the time of the event; changed only by format, normalization or redaction | Timestamps from a ticketing system, sensor readings, a customer's message with names masked | Values back-filled later from another system are labeled observed when they were inferred |
| Observed, reported | Entered by a person describing something, not captured by a system | Agent-set priority, self-reported job title, survey answers | Treated as ground truth when it is one person's judgment |
| Inferred | Computed or predicted from other fields by a named rule, imputation method or model | Category from a keyword rule, imputed missing codes, model sentiment labels | Model labels shipped as observed because they were in the source system |
| Generated | Produced by a generative process without a one-to-one real value | LLM-written summaries, simulator trajectories, synthesized rows | Lightly edited generated text relabeled as observed after human review |
The field-level origin disclosure schema
The schema below has one entry per field, plus per-value overrides where a column mixes classes. It is intended to travel with the dataset as a machine-readable file next to the data card. Required-when rules keep it short: an observed field does not need a generator reference, and a generated field does not need a source column.
At a glance
The field-level origin disclosure schema
Two attributes do most of the work in practice. origin_scope forces a supplier to admit when a column is mixed, which is the most common way origin gets lost: a code that was typed by agents for years and filled in by a classifier after a migration looks uniform in the export. intended_use turns the disclosure into an instruction a pipeline can follow, so nobody has to reread a data card to know which fields belong in an evaluation set.
Keep the file versioned with the data. When a supplier retrains the sentiment classifier or changes the summary prompt, the inferred and generated values change even though the source records did not, and the as_of and derivation_method entries are the only way a buyer will notice.
| Attribute | Purpose | Allowed values or format | Required when |
|---|---|---|---|
| origin_class | States how the value came to exist | observed, observed_reported, inferred, generated | Always |
| origin_scope | Says whether the class holds for the whole column or varies by value | field or per_value | Always; per_value requires a value-level origin column |
| source_ref | Points to the system, table and column an observed value came from | System name, object, field path | Observed and observed_reported |
| transformations | Lists changes that keep a value observed | Ordered list: format, normalize, redact, tokenize | Any observed field that was changed |
| derivation_method | Names the rule or model that produced an inferred value | Rule ID with version, or model name with version | Inferred |
| inputs | Lists the fields the inference or generation used | Field names from this schema | Inferred and generated |
| generator_ref | Identifies the generative process | Model or simulator, version, prompt or config ID, seed set ID | Generated |
| as_of | Dates the value or the rule or model used | ISO date or datetime | Always for inferred and generated; event time for observed |
| validation_status | Records how the label or value was checked | unchecked, sample_checked with sample size, fully_checked | Inferred, generated and observed_reported |
| intended_use | Signals where the field is safe to use | train, evaluate, both, neither | Always |
How does the schema map onto Croissant, PROV and media metadata?
It sits on top of what teams already publish. In the Croissant 1.1 format, each field in a record set has a source property naming where its data comes from, and a DataSource describes how values are extracted and transformed to populate the field. That is a natural home for source_ref and transformations, with origin_class added as an extension property.
For lineage across systems, W3C's PROV-O vocabulary already has the distinction you need. Its definition of derivation covers the transformation of one entity into another, an update that produces a new one, or construction of a new entity based on a pre-existing one, and wasGeneratedBy links an entity to the activity that produced it. Inferred and generated values map to derivations and generation activities, which lets provenance tools trace them.
Images and video have a vocabulary of their own. IPTC recommends that software creating images with trained AI models records the digital source type trainedAlgorithmicMedia, while digitalCapture covers camera originals and compositeSynthetic covers composites that include a synthetic element. When media files sit inside a record, carry their IPTC value into origin_class rather than inventing a second label.
Regulation points the same way at a coarser grain. For high-risk AI systems, Article 10(2) of the EU AI Act requires data governance practices covering data collection processes and the origin of data, and preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation. Field-level origin labels are one practical way to keep that record, though the Act does not prescribe this format.
What does a valid worked record look like?
The record below is illustrative. It is invented to show the schema in use and is not drawn from any customer's data. It represents one closed support ticket from a help desk export, prepared for training a triage assistant.
At a glance
What does a valid worked record look like
Three details make this record pass. The redacted message stays observed because masking changed its content but not its origin, and the transformation is listed. The two model outputs name their model and version, so they can be regenerated or dropped when the model changes. And the resolution code admits it is mixed instead of claiming a single class for the whole column.
- ticket_token: observed; source_ref help desk, tickets, id; transformations tokenize; intended_use both.
- opened_at: observed; source_ref tickets, created; transformations format to UTC; intended_use both.
- customer_message: observed; source_ref first public comment; transformations redact names, emails and order numbers; intended_use both.
- priority: observed_reported; set by the agent; validation_status unchecked; intended_use train.
- product_area: inferred; derivation_method keyword rule v3; inputs customer_message; validation_status sample_checked, 200 tickets; intended_use train.
- sentiment: inferred; derivation_method classifier v1.4; inputs customer_message; as_of model date; intended_use train.
- resolution_summary: generated; generator_ref language model and version, prompt ID, seed none; inputs the full ticket thread; validation_status sample_checked, 100 tickets; intended_use train only.
- resolution_code: observed for most rows and inferred for some; origin_scope per_value; value-level column marks which.
What should the validator reject, and what does a failing record look like?
A validator turns the disclosure from a promise into a check. Run it on every record before delivery and on a sample after receipt. The rules below catch most contradictions without needing to know anything about the domain.
Some checks need data the record does not carry. Spot-check observed fields against the source system's audit history where the supplier can provide it, because a value written by an integration user long after the event was probably not observed. Compare inferred values with a fresh run of the named rule or model on a sample, since a mismatch means the derivation_method entry is stale.
- Every field has an origin_class and an intended_use.
- Observed fields have a source_ref and no generator_ref.
- Inferred fields have a derivation_method, at least one input and an as_of date.
- No inferred or generated field lists itself among its inputs.
- Generated fields have a generator_ref and are never marked intended_use evaluate.
- Columns with origin_scope per_value have a value-level origin for every row.
- A field derived from a generated input cannot be labeled observed.
- Fields with validation_status sample_checked state the sample size.
Reading the validator output on a failing record
The following failure is illustrative. A second ticket record arrives with resolution_code labeled observed at field scope, but the supplier's notes mention that codes were back-filled with a classifier for older tickets. The validator does not know about the notes; it fails the record for other reasons, and the reasons lead to the same place.
First error: resolution_summary is labeled observed but carries a generator_ref, which observed fields may not have. Second error: product_area lists resolution_summary among its inputs, so a field derived from generated text is in the chain, yet product_area is labeled observed. Third warning: resolution_code is field-scoped observed, but a spot check against the help desk audit log finds the code was written by an integration user months after closure.
The fix is to relabel resolution_summary as generated, product_area as inferred with its rule and inputs, and resolution_code as per_value, with the back-filled values marked inferred. In the illustrative export, that moves about a third of resolution codes out of the evaluation set, which is exactly the decision the disclosure exists to support.
Notice what the validator did not need. It did not judge whether the summaries were good or the classifier accurate; it only checked that the labels agree with each other and with the evidence attached. Quality evaluation is a separate step, and it is much easier once origin is settled, because each class can be measured against the right reference.
Which real exceptions does the generated share fail to represent?
Usually the rare ones. Generators reproduce what is common in their inputs, and repeated training on generated outputs makes this worse: a 2024 Nature paper on model collapse reports that models trained on recursively generated data lose the tails of the original distribution first. In a field-level disclosure, that shows up as categories present in observed values but thin or absent in generated ones.
Check it directly. Hold out a set of observed records the generator never saw, then compare category counts per field between the observed holdout and the generated share. Report every category that appears in the holdout and not in the generated values, with its holdout count. Score any model trained on the mixture on the untouched observed records only, because evaluating on generated fields measures agreement with the generator, not with reality.
Look at combinations as well as single fields. A generator can reproduce every product area and every priority on its own and still never produce an urgent ticket about a rarely used product, because that pairing was thin in its inputs. Cross-tabulate the two or three fields your model will rely on most, and list the combinations that appear only in observed records.
Asking for field-level origin labels in a SourceX request
Because a SourceX dataset is sourced from U.S. businesses only once a buyer asks, the origin schema can sit in the request next to scope, volume, format, timing and licensing needs. Name the fields you need observed for evaluation and the fields where inferred or generated values are acceptable for training. De-identification is done by SourceX through approved processing routes, which is a transformation of observed fields that the disclosure should list rather than a change of class.
Frequently asked questions
What is the difference between inferred and generated data?
Inferred values are computed or predicted from other fields in the same record by a named rule or model, so they can be traced to specific inputs. Generated values are produced by a generative process without a one-to-one real value behind them. Both should be disclosed, with different evidence.
Does redacting or de-identifying a field make it synthetic?
No. Masking names or tokenizing IDs changes the value's format but not where it came from, so the field stays observed with the transformation listed. Replacing a redacted name with an invented one does introduce generated content, and that replacement should be labeled.
Are labels from human annotators observed or inferred?
They are judgments about the data rather than captures of an event, so treat them as observed_reported when a person entered them in the source system, or as inferred when added later by an annotation process, with the guideline version as the method. Either way, record how they were checked.
Why should generated fields be excluded from evaluation sets?
Scoring a model against generated values measures how closely it matches the generator, not how it performs on real cases. Keep evaluation on observed values from records the generator never saw, and report generated-field results separately if at all.
Is field-level origin disclosure required by law?
Not as a specific format. The EU AI Act asks high-risk AI providers to govern data origin and preparation operations, and field-level labels are one practical way to keep that record. Counsel should review what a particular system and jurisdiction require.
Sources
- Datasheets for Datasets asks whether each instance's data was directly observable, reported by subjects, or inferred or derived from other data, and whether it was validated. Source
- The Croissant RAI 1.0 vocabulary defines dataset-level properties for collection type and imputation protocol and lists the synthetic data generation method among compliance information. Source
- Croissant 1.1 gives each field a source property and uses DataSource to describe how data is extracted and transformed to populate a field. Source
- PROV-O defines derivation as transforming, updating or constructing a new entity from a pre-existing one, and links entities to the activities that generated them. Source
- IPTC recommends the digital source type trainedAlgorithmicMedia for AI-generated images, alongside values such as digitalCapture and compositeSynthetic. Source
- Article 10(2) of the EU AI Act requires data governance for high-risk AI data covering data origin and preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation. Source
- A 2024 Nature paper reports that models trained on recursively generated data lose the tails of the original distribution. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.