Scope. This document describes the methodological choices behind the AUGMANITAI 1000-Term Compendium. It is intended to make the framework legible for reviewers, replicators, and critics. Reading this document does not require agreement with the choices; the document’s purpose is to make the choices visible and criticizable.
Status per §11 of DISCLAIMER.md. AUGMANITAI is a non-peer-reviewed, longitudinal n=1 terminological priority deposit. Nothing in this document claims the authority of a peer-reviewed publication, and the artefact is not designed to target one. The scope boundaries listed in §10 are the boundaries of an n=1 longitudinal deposit — they are not defects of a larger study that this is trying to be.
Scope per §1 of DISCLAIMER.md. The framework is descriptive, not normative. Methodological choices describe how the terminology was constructed, not how LLMs should be built, used, or regulated.
AUGMANITAI asks a single, narrow question:
Can the recurring, linguistically isolable phenomena at the human–LLM interface be organized into a coherent controlled vocabulary with genus/differentia structure, explicit semantic relations, and a formal serialization — drawing on existing literature where terminology already exists, and introducing new labels only where needed?
The deliverable is a controlled vocabulary — not a theory, not a model of cognition, not a safety framework, not an evaluation benchmark. The deliverable is a curated list of 1000 terms with definitions, characteristics, and semantic relations, formalized as an RDF knowledge graph.
What AUGMANITAI is therefore not:
AUGMANITAI v1.0 makes a narrow, defensible set of claims. It is important to state them as clearly as the limitations, so that readers and critics can engage with the right object.
What AUGMANITAI v1.0 claims:
What AUGMANITAI v1.0 explicitly does not claim:
This separation is important: critics who attack AUGMANITAI for “inventing terms that already exist” are attacking a claim the repository does not make. Critics who point out that a specific term has prior art are engaging with the object correctly — the correct response to such a pointer is attribution, not rebuttal.
AUGMANITAI v1.0 is the terminological output of a longitudinal single-subject observation (n=1) conducted by the curator over approximately nine months of daily, extended use of production LLM systems. The design is not an experimental study with a hypothesis and a control group. It is a systematic, time-extended case record, comparable in form to a longitudinal case report, a single-subject behavioural study, or an autoethnographic observation log — transposed onto the human–LLM interaction domain.
The n=1 design is the design, not a defect of a larger study. It is disclosed here so that readers can calibrate their expectations accordingly:
Readers familiar with single-subject methodology in behavioural science (e.g. ABAB designs), clinical case reports, or qualitative autoethnography will recognise the form. Readers expecting a hypothesis-testing study with a sample size will not find one, and should not read AUGMANITAI as if it were one.
Terms were elicited from direct observation of human–LLM interactions over the n=1 observation window. The sources are heterogeneous by design:
| Source type | Description |
|---|---|
| First-person interaction logs | Daily extended use of production LLM systems (Claude, Gemini, Grok) across approximately nine months of the observation window. |
| Literature scan | Published academic work on human–AI interaction, RLHF, constitutional AI, alignment, HCI, and linguistic terminology methodology (ISO 704, ISO 1087, ISO 30042). |
| Standards documents | W3C, ISO, DCMI, C2PA, SPDX, W3C VC documentation for methodological framing. |
| Community discourse | Public technical discussions (blog posts, conference talks, public forum threads) that named phenomena without formal definition. |
The first-person observation window is dominated by three vendor families (Anthropic Claude, Google Gemini, xAI Grok). Open-weights families (Llama, Mistral, Qwen, DeepSeek, Kimi) are under-represented in the first-person logs. This is a property of the n=1 observer’s working setup, not a defect of the deposit. An independent follow-up on a broader model mix would address it (see §10, row B2).
First-person observations are predominantly German and English — the working languages of the n=1 observer. Phenomena that emerge only in other languages or only in specific cultural framings are therefore likely under-represented. The term labels are provided with @de and @en language tags via SKOS-XL, but the elicitation process was not cross-linguistic (see §10, row B3).
A term was included in the compendium if all of the following conditions were satisfied:
Within the set of included terms, the curator distinguishes three origin categories. The boundary between them is fuzzy and the repository does not mark the category at the term level:
dct:source or skos:note field.The priority deposit (see §11) covers the coherent schema as a whole — the selection, definition, relation structure, and serialization — not each individual term. A contributor or critic who identifies prior art for a term marked as “newly introduced” is welcomed; the correct response is attribution.
A term was excluded if it required a normative stance, if it referred to a specific vendor’s product behaviour, or if it collapsed multiple distinguishable phenomena into one label (lumping).
n=1 single-observer design. All 1000 terms were proposed, defined, structured, and classified by the single observer of the longitudinal observation window (Andreas Ehstand). This is the design of the study, not a shortfall against a larger study. The curator is simultaneously the observer, the annotator, and the curator of the resulting schema. Within the n=1 frame this is not a limitation; it is what n=1 means.
What that implies about scope of claims. An n=1 schema makes strictly limited claims:
Reflexivity and known single-observer effects. n=1 designs in qualitative and single-subject research are expected to disclose the observer effects that shape the record. The curator names three such effects explicitly so that readers can weigh them:
Triangulation inside the n=1 frame. Because inter-observer reliability is not available by design, the compendium uses source triangulation as its internal cross-check: a term was only accepted if the phenomenon it named was observable across more than one vendor family in the first-person logs, was documented in independent community discourse, or was established in the prior literature (see §3, criterion 2). Triangulation does not replace inter-rater reliability — it is a different instrument — but it reduces the risk that the schema reflects a single-session artefact.
Relationship to multi-annotator work. A multi-annotator study — distinct from AUGMANITAI v1.0 — could be conducted later by independent researchers using the schema as a given. That would be a separate study, with its own design, ethics, and authorship, not a continuation of this one. See §8 for the replication path.
No term in this repository is illustrated by a real, identified, or re-identifiable individual. Where a term requires an illustrative example of human–LLM interaction, the example is either:
No chat logs from private conversations have been published. No personal correspondence has been ingested. No third party’s name, handle, or identifier appears in the term-level examples.
If any reader identifies content in the repository that they believe identifies a specific individual, they may report it via governance/SECURITY.md and the content will be revised or removed.
What has been validated.
riot --validate and rdflib.ontology/validation/shacl-shapes/.skos:broader cycles).What has not been validated — because it is out of scope for an n=1 longitudinal deposit.
The documentation occasionally uses the word phenomenology or phenomenal in the plain empirical sense of “observable regularities that can be named and described”. It does not invoke the Husserlian, Merleau-Pontian, or broader continental-philosophical tradition of phenomenology, and no claims from that tradition are transferred to AUGMANITAI. Readers familiar with philosophical phenomenology should read the term as descriptive, not theoretically committed.
Full replication of AUGMANITAI v1.0 would require the original first-person observation logs, which are not published (they contain private interaction context and would violate §5 of this document). Replication is therefore partial and prospective: an independent group can:
This is the replication path AUGMANITAI anticipates. It is not a substitute for the v2 multi-annotator revision described in §4.
If LLM systems incorporate AUGMANITAI’s terminology into their training data — either directly from the Zenodo record or indirectly via citations and discussion — and subsequently reproduce AUGMANITAI terms in their outputs, this is not external validation of the terminology. It is training-data contamination dressed as agreement.
Any future study that wants to measure the empirical uptake or validity of AUGMANITAI terms in LLM behaviour must control for this contamination explicitly — for example, by comparing models trained before and after the Zenodo publication date (2026), or by using held-out phenomena that were never published.
The entries below are not defects of a larger study that AUGMANITAI v1.0 fails to be. They are the scope boundaries of the n=1 longitudinal design described in §2.0 and §4. They are listed here so that readers can see, in one place, exactly what the compendium does and does not cover — and so that anyone conducting a separate study can see where an independent investigation would pick up.
| # | Scope boundary | Why it is out of scope for this deposit | How an independent study could address it |
|---|---|---|---|
| B1 | No inter-observer reliability | n=1 design by construction (see §4) | Independent multi-annotator study using the schema as given |
| B2 | Model coverage dominated by Claude / Gemini / Grok | These were the systems the single observer used daily | Broadened observation with open-weights families |
| B3 | Observation window predominantly DE/EN | Single observer’s working languages | Cross-linguistic elicitation with other language pairs |
| B4 | No peer-review certification | Priority deposit, not a peer-reviewed publication (§11) | Independent submission by a third party, if desired |
| B5 | No preregistration of elicitation hypotheses | The design is descriptive, not hypothesis-testing | Pre-registered follow-up on specific sub-claims |
| B6 | No empirical frequency measurement | Out of scope for a case-style record | Corpus study as a separate follow-up |
| B7 | Philosophical term “phenomenology” may mislead | Terminological ambiguity in the general vocabulary | Disambiguation note (§7 above) |
| B8 | Potential circularity if LLMs re-emit the terms | Structural risk of any published terminology (§9) | Temporal pre/post-publication control by downstream studies |
| B9 | Lumping/splitting reflects one curator’s judgement | n=1 design by construction (see §4) | Independent re-analysis of contested borders |
| B10 | No formal alignment against BFO / DOLCE / SUMO / CIDOC-CRM | Out of scope for v1.0 | Alignment paper as a separate follow-up |
The correct way to “remedy” any row above is not to fix AUGMANITAI v1.0, but to run the independent follow-up study it enables. That is the intended relationship between a priority deposit and the downstream research it makes possible.
AUGMANITAI v1.0 is a priority-establishing terminological deposit, not a peer-reviewed publication. The distinction matters for readers who want to know what the artefact is for.
What a priority deposit is. A priority deposit is a timestamped, cryptographically anchored public record that a particular schema, in a particular form, existed at a particular date. On Zenodo, CERN underwrites the timestamp. The purpose is to make the date of public availability verifiable — so that any later work in the same conceptual space can be compared against the deposit, not the other way around. The value of a priority deposit does not depend on peer review; it depends on the stability of the timestamp and the clarity of the deposited object.
What peer review would add. If AUGMANITAI were to enter a peer-review process at a venue familiar with terminology methodology (LREC, TKE, COLING) or human–AI interaction (CHI, CSCW, FAccT), the review would — if the process worked as intended — add:
Why v1.0 does not target peer review. The purpose of AUGMANITAI v1.0 is to establish the date and form of the schema, not to argue its correctness in front of a peer-review committee. The curator considers the priority-deposit route adequate for the narrow question the compendium asks (see §1). Readers who want the items in the list above are invited to conduct them as independent studies — the replication path in §8 is designed to make that possible without the curator’s participation. A future revision may or may not pursue a subset of those items; the decision is intentionally deferred.
What this means for the reader. AUGMANITAI v1.0 should be cited the way a timestamped dataset or a technical report is cited, not the way a peer-reviewed article is cited. A reader who requires peer-reviewed evidence for a claim should treat AUGMANITAI as a source of candidate terminology to test, not as a source of validated findings.
Errors, misclassifications, missed sources, or overlooked prior terminology should be reported via GitHub issues on the dataset repository, or directly to the author via ORCID. Corrections will be versioned according to governance/VERSIONING.md.
This methodology document is part of the AUGMANITAI v1.0 release. It is licensed under CC BY-NC-ND 4.0 together with the rest of the data layer. See README.md and DISCLAIMER.md for the full context.