The ontology held 424 competencies at the September 2026 reading date (n=424). The number of frameworks represented in it was 127 at the same reading (n=127). The framework count describes breadth of sourcing. It does not indicate, for any single competency in the ontology, whether that competency is attached to recorded observation, nor how many competencies are so attached. The criteria by which a competency was admitted to the ontology, and by which a framework was counted as represented, are not reported in the evidence available here.
Corpus and ontology
Two stores are involved, and they are distinct. PeerLibrary, the cross-tenant licensable corpus, held 259,205 behavioural primitives at the September 2026 reading (n=259,205). Of those 259,205 primitives, 256,286 were cleared as licensable (n=259,205).
The platform assessment corpus is a separate store. It held 17,403 atomic behavioural observations at the September 2026 reading (n=17,403). Commercial conversations assessed numbered 1,384 at that reading (n=1,384). These two figures rest on different measurements. No mapping between them is reported in the evidence available.
The ontology itself is a label set over competencies. Whether its entries index into either store, and by what key, is not reported.
Method
The counts above were read as of a single date in September 2026, and each carries the sample size recorded against it. No interval is reported for any of them. The procedure that produced each count is not described in the evidence available here, which means the following cannot be stated: whether counts were taken at the same moment or reconciled after the fact, whether duplicate entries were removed before counting, and whether any entry was counted under more than one framework.
No agreement statistic is reported for the assignment of competencies to frameworks. No reliability figure is reported for the atomic observations. Where a published taxonomy paper would report the third of the conventional steps in taxonomy construction, evaluation, following term selection and relationship construction as set out in a bottom-up taxonomy study of the software application domain, the evidence available here supplies only counts of terms. The evaluation step is unreported rather than reported as negative.
Results
The ontology contained 424 competencies. Framework representation stood at 127. Both figures are counts of ontology contents and neither is a measurement of evidential support for the contents.
The licensable clearance figure is the one count in the set that reports a property of individual records rather than a total: 256,286 of the 259,205 PeerLibrary primitives carried licensable clearance. The basis on which clearance was granted or withheld is not reported.
For the platform assessment corpus, 17,403 atomic behavioural observations were held. The distribution of those observations across competency labels is not reported. The number of competencies in the ontology that have at least one attached observation is therefore unknown from this evidence, and so is the number with none. A count of 424 is compatible with a highly concentrated distribution and with a uniform one; nothing here distinguishes the two.
Unreported admission criteria
The following are not stated in the evidence available here, and are therefore treated as unreported rather than as absent or present:
- The threshold a candidate competency had to meet for admission, including whether any minimum of recorded observation was required.
- Whether frameworks were weighted, and whether appearance in more than one framework affected admission.
- The rule for deduplicating constructs that appear in several frameworks under different names, and whether any such rule was applied before the count of 424 was taken.
- Who adjudicated admission, how many adjudicators were involved, and what agreement they reached.
- Whether any of these criteria were fixed before the ontology was populated.
- What counted as a framework being represented: one competency traced to it, a majority of its constructs present, or some other condition.
The last of these bears directly on the interpretation of 127. A framework represented by a single borrowed construct and a framework mapped in full contribute equally to a count of frameworks.
Interpretation
Counts of taxonomy size are common in published work and are generally accompanied by a second quantity that carries the evidential load. A multilayer taxonomy for large language models partitions capability into domains and subskills across primitive, constructed and integrative layers, and reports alongside that partition how the underlying literature distributes over it: the authors screened papers from four conference venues and mapped a language-model-focused subset through multi-model annotation, consensus and arbitration, finding research attention concentrated in language-semantic competence and reasoning while six of the domains appeared in fewer than 2% of papers, as they describe. The domain count tells the reader how the space was partitioned. The distribution tells the reader which partitions are populated. The evidence available here supplies the first kind of quantity for the ontology and not the second.
The same asymmetry appears in occupational taxonomies, where published totals differ substantially between systems. O*NET is described as covering 1,016 occupations on one comparison of the two systems. ESCO is described in that same comparison as covering 3,039 occupations, with the larger figure reflecting finer stratification of the same underlying work rather than more evidence about it. A larger count can record a finer partition, a broader intake of sources, or duplication that survived deduplication.
For an ontology intended to support inferences about mastery, the relevant standard is not the size of the label set. Validity in the framing adopted by the Standards for Educational and Psychological Testing attaches to the interpretation of scores for particular uses, a reframing introduced in the 1999 revision, and the Standards do not prescribe particular development procedures. On that framing, a competency count is a description of an instrument's structure. A framework count is likewise a structural description. Both bear on validity only through evidence not present here.
Limitations
All six counts are single-date readings with no interval, and no time series is available, so no statement about growth or stability can be made. The 17,403 observations are reported under their own sample size (n=17,403). The 1,384 assessed conversations are reported under a different one (n=1,384), and no link between the two measurements appears in this evidence, so no per-conversation rate can be computed. The 259,205 primitives sit in PeerLibrary and the 17,403 observations sit in the platform assessment corpus, and no correspondence between the two stores is reported. Cohort composition behind any of the counts is not reported, and no cohort is described here for that reason. Whether the 424 competencies are mutually exclusive is not reported, so the count may include overlapping constructs.
Further work
A distribution of the atomic behavioural observations over competency labels would establish how much of the ontology is evidenced and how much is currently label only; the observation store held 17,403 records at the reading date. A second quantity worth having is an agreement statistic for the assignment of competencies to frameworks, which would establish whether the figure of 127 is reproducible by an independent annotator. A stated admission rule, with the date it was adopted, would make the count of 424 interpretable as something other than the outcome of an unspecified process. Prompting and fine-tuning approaches to taxonomy construction have been compared directly in published work, with the reported finding that prompting outperformed fine-tuning even without training on the dataset while producing structural constraint violations that are harder to post-process, in a comparative study; whether any automated step contributed to this ontology is not reported, and that question would need answering before such comparisons apply.
Sources
- Automatic Bottom-Up Taxonomy Construction: A Software Application Domain Study
- From Isolated Tasks to Structured Capabilities: A Multilayer Taxonomy for Large Language Models
- Dual-Taxonomy Career Guidance: ESCO-O*NET Mapping Explained
- Standards for Educational and Psychological Testing (SAGE Encyclopedia of Educational Research, Measurement, and Evaluation)
- Prompting or Fine-tuning? A Comparative Study of Large Language Models for Taxonomy Construction