The source is a single database table named human_validations, keyed on the author recorded against each row. The table held 1,145 independent AI dual-rater re-grades recorded by a second model rather than by a person (n=1,145, dated 2026-08-19). The name asserts a property about authorship at the level of the container, while the author field records that property for some rows and records a model for others.
The same table held 9 validations recorded by a human operator (n=9, dated 2026-08-19). No measurement of the consequences of the mismatch between the container name and the recorded author is reported, and no cost figure is available for either path.
Method
The evidence consists in two row counts drawn from the same table on the same date. Both row counts were taken on 2026-08-19. The first is a tally for independent AI dual-rater re-grades recorded by a second model, 1,145 (n=1,145). The second is a tally for validations in that same table recorded by a human operator, 9 (n=9). Each figure carries its own n and is reported separately. No total row count for the table is reported, so no ratio, share, or proportion is computed here, and none should be inferred by juxtaposing the two tallies.
Several properties that would bear on interpretation are not reported. The mechanism that populates the author field is not reported, so whether the value is written by the process that creates the row or attached afterwards cannot be established. Whether the two categories are mutually exclusive, and whether together they exhaust the table, is not reported. No agreement statistic relating the model-recorded re-grades to the human-operator rows is reported, and no interval is reported for either count. When the table was named, and whether the name was fixed before or after the dual-rater path began writing to it, is not reported. What the underlying grades assess is not specified in the evidence and is not asserted here.
Results
The table records 1,145 independent AI dual-rater re-grades recorded by a second model. The same table records 9 validations recorded by a human operator. Both figures are dated 2026-08-19.
The table name therefore functions as a provenance assertion that the author field does not uniformly support. A consumer reading only the schema receives one claim about who produced the rows. A consumer reading the author field receives two different claims depending on the row. Nothing in the evidence indicates which artefact any downstream process actually reads.
Interpretation
Data lineage is conventionally defined as the tracking of how data is generated, transformed, transmitted, and used across systems over time, and provenance records are described as supporting forensic activities including error and compromise detection, auditing, and compliance analysis (Wikipedia, Data lineage). Under that definition a table name is not a provenance record. It is a container label that happens to encode an agent claim, and it carries no per-row attribution, no timestamp, and no revision history. The row-level author field is the only artefact in this evidence that discharges the function.
The W3C PROV family places agent attribution in the data model itself: PROV-DM as a relational data model, PROV-O as an OWL ontology for conceptual modelling, and PROV-N as a human-readable syntax, with PROV-ML applying the family to the scientific machine learning lifecycle (Provenance Tracking in Large-Scale Machine Learning Systems). The same paper argues that provenance tracking is often overlooked because machine learning systems prioritise performance optimisation and scalability over interpretability, and because recording and storing such metadata reduces efficiency. A table name encoding an agent claim is the low-cost substitute that this argument predicts: it requires no per-row write and no schema change, and it becomes wrong silently once a second writer is added.
The failure mode has a documented analogue in dataset licensing. An audit tracing the lineage of text finetuning datasets across widely used text data collections found that most licenses for popular datasets on GitHub and Hugging Face are listed as "Unspecified", and re-annotation by the auditors found that analysed Hugging Face licenses often fell into a different use category than the original author's license, frequently labelled more permissive (The Data Provenance Initiative). In both cases a label attached at the artefact level diverges from what the underlying records support, and the divergence runs towards the more favourable reading. A container named for human authorship reads as stronger evidence of human confirmation than a mixed table warrants.
Frameworks designed for human-AI collaborative data creation locate the guarantee at the ingestion event rather than at the name. TableVault is presented as a metadata governance framework built around a guaranteed record of data ingestion, robust operation status tracing, lineage capture of execution parameters, and a standardised metadata layer queryable from within model executions (Metadata Management for AI-Augmented Data Workflows). The same paper invokes the FAIR principles, holding that metadata should be linked to the original data source, accessible to subsequent practitioners, contain detailed lineage information, and follow a standardised format. Measured against that standard, the two counts reported here are the beginning of a lineage record rather than an instance of one.
For a measurement of competency mastery, the distinction matters at the point where a grade is treated as confirmed. A validation row recorded by a second model and a validation row recorded by a human operator support different claims about independence, and the difference is invisible to any consumer that resolves provenance from the table name. Whether any consumer in this system does so is not established by the evidence.
Limitations
The two counts are separate measurements with separate sample sizes and are not combined. Nothing here supports a statement about the proportion attributable to either author type, because the total row count is not reported.
The 9 human-operator rows are too few to support any inference about behaviour, reliability, or throughput in human validation, and none is offered.
Only one table on one date is observed, so no trend, no rate of accumulation, and no change following any intervention can be reported. The recorded author is taken at face value; the evidence does not establish that the field reflects the actual writer of the row. No agreement, concordance, or error-rate statistic relates the model-recorded re-grades to the human-operator rows. No cost figure is available for either path, so the efficiency argument in the provenance literature is invoked as a mechanism and not as a measured quantity for this system. No claim is made about the intent behind the table name, since that is not reported.
Related work
The consequences of losing origin information as datasets are shared, edited, and reproduced, and the framing of dataset compliance as a rateable property, are set out in the Compliance Rating Scheme paper, which also reports a study finding that nearly half of popular AI training datasets include data whose creators were neither asked nor informed about the inclusion while their licenses make the use appear permissible (Compliance Rating Scheme). Broader treatments of consent, credit, and the limited reliability of machine unlearning assign responsibility for provenance jointly to data creators, AI developers, regulators, and researchers (Data Authenticity, Consent, & Provenance for AI). Commentary in Nature Machine Intelligence notes that many questions remain about the provenance of training datasets, license enforcement, and creator consent, and emphasises the importance of developers engaging with dataset metadata (A step forward in tracing and documenting dataset provenance).
Further work
Backward and forward lineage can be combined into an end-to-end audit trail running from a reference point to final destinations (Wikipedia, Data lineage); applying that construction to the validation path would establish which downstream consumers resolve authorship from the row and which from the container. Three further measurements follow from the present counts: agreement between model-recorded re-grades and human-operator records on rows where both exist, the cost per validation on each path, and the accumulation rate for each author type over time. None is reported in the present evidence.
Sources
- Provenance Tracking in Large-Scale Machine Learning Systems
- Data Authenticity, Consent, & Provenance for AI
- Metadata Management for AI-Augmented Data Workflows
- Compliance Rating Scheme: A Data Provenance Framework
- A step forward in tracing and documenting dataset provenance
- The Data Provenance Initiative
- Data lineage - Wikipedia