12. Glossary

Terms are explained as they are used in this project; where a term has broader uses in linguistics or philosophy, the entry notes the sense intended here.

• AANN construction — the English pattern article + adjective + numeral + noun, as in a beautiful three days. Grammatically odd on its face (a pairing with plural days), yet natural to native speakers under specific conditions — which makes it a sensitive probe of graded grammatical judgment.
• Acceptability judgment — a speaker's (or model's) rating of how natural a sentence sounds, e.g. on a 1–7 scale. The standard data of experimental syntax.
• Alternation — a choice between two grammatical ways of expressing nearly the same content, e.g. give Mary the book vs give the book to Mary. Human choices between alternants follow statistical "soft constraints" rather than absolute rules.
• Anchor (human anchor) — an existing dataset of human judgments or behavior that an experiment's model results are compared against. The project's rule: every claim about human-likeness must name its anchor; results without one are labeled internal-contrast-only.
• Behavioral probe — an experiment that only looks at what a model says in response to prompts, never at its internal computations.
• BLiMP — the Benchmark of Linguistic Minimal Pairs: 67,000 published sentence pairs, each contrasting a grammatical sentence with a minimally different ungrammatical one, with recorded human agreement rates per grammatical phenomenon.
• Claim — in this project, a conclusion promoted from experimental results only after replication, surviving controls, and an adversarial review in a later session; deliberately scoped and fenced with caveats. Stronger than a result.
• Comparative correlative — the construction the X-er …, the Y-er …, as in The more you practice, the better you get. Its parts do not add up word-by-word to its meaning (that one quantity covaries with another) — the pattern itself carries it.
• Confidence interval (CI) — the range within which a measured value would be expected to fall on repetition of the experiment; "the interval excludes zero" means the effect is statistically distinguishable from no effect at the standard used.
• Confound — an unintended factor that could produce an experiment's outcome without the claimed cause; controls exist to remove confounds.
• Constructional meaning — meaning contributed by a grammatical pattern itself, over and above its words. The theoretical home of this idea is construction grammar, the school holding that grammar is a repertoire of form–meaning pairings at every size, from words to idioms to abstract sentence patterns.
• Contamination — the possibility that a model performs well on a published test because that test (or its answer key) appeared in the model's training data. A standing caveat for all public-dataset results.
• Control condition — a comparison condition built to capture everything about the test except the thing under study (e.g. same words without the construction), so that the difference isolates the thing under study.
• Corpus — a large recorded collection of naturally produced language (books, news, conversation transcripts) used for counting how speakers actually talk and write.
• Correlation (Spearman's ρ) — a statistic between −1 and +1 measuring how well one ranking tracks another; 0 means no relationship, 1 means identical ordering. Used here to compare model gradients with human gradients.
• Dative alternation — the choice between give Mary the book (double-object) and give the book to Mary (prepositional). A classic case of a probabilistic grammatical choice, with well-measured human preferences.
• Distributional shadow — this project's term for everything about a linguistic behavior that is already predicted by word-company statistics (which words occur near which). A behavior "beats the shadow" when it survives a control that captures those statistics; see §3.
• Distributional semantics — the research tradition treating a word's meaning as (or as reflected in) its pattern of co-occurrence with other words; the implicit theory behind LLM training.
• DWUG / DURel scale — DWUG (Diachronic Word Usage Graphs) is a published dataset in which human annotators rated pairs of real corpus sentences for how related the meaning of a shared target word is, on the 4-point DURel scale (4 = Identical, 3 = Closely Related, 2 = Distantly Related, 1 = Unrelated). The project's word-sense experiments use its English portion.
• Factive verb — a verb that presupposes the truth of its complement clause: realize in Sam didn't realize the door was locked still implies the door was locked; contrast non-factive suspect.
• Firewall (control) — the project's strictest control type: the compared strings are byte-identical (or use invented words), so no statistic of the strings themselves can explain a difference; only the surrounding context differs.
• Frozen design / pre-registration — fixing an experiment's materials, prompts, analysis, and success criteria before running it, so results cannot quietly reshape the test.
• Function word — a small closed-class word doing grammatical/logical work (the, of, some, every, because, although, not), as opposed to content words (cat, run, beautiful).
• Genitive alternation — the choice between the teacher's voice (s-genitive) and the voice of the teacher (of-genitive). Human speakers prefer the s-genitive more when the possessor is animate (a person or animal).
• Givenness / information structure — whether a referent is already "given" (mentioned, established) in the discourse or "new." Languages tend to place given material earlier; "information structure" is the general term for how sentences package given vs new.
• Grounding — connection between language and non-linguistic experience (perception, action). The symbol grounding problem asks how symbols could get meaning for a system that only ever manipulates other symbols.
• Internal-contrast-only — the project's label for a result that reports a difference within a model's behavior while making no claim of human likeness, because no license-verified human dataset anchors it.
• Island (syntactic) — a construction out of which question-formation cannot move: Who did a lot of doctors embarrass Winston Churchill without referencing? is ungrammatical because the without… clause is an "island." Island judgments require tracking deep, non-local structure, which makes them a hard stratum in grammar benchmarks.
• Large language model (LLM) — a neural network trained on very large text collections to predict the next word, from which conversational systems like the panel models are built.
• Minimal pair — two sentences differing in exactly one respect (one word, one inflection), so that any judgment difference isolates that respect.
• NLI (natural language inference) — the task of judging whether a premise sentence entails, is neutral toward, or contradicts a hypothesis sentence; a standard instrument in several of the project's probes.
• Nonce word — an invented word (e.g. wug, blicket) with no history in any text, used to test rules and preferences where no memorized statistics can help.
• Null result — a well-run experiment that finds no effect. In this project, nulls are published with the same care as positives, and several are load-bearing.
• OpenRouter — a commercial service giving unified access to many AI models; the project's route to its panel, and the meter for its budget.
• Panel — the project's fixed set of three test-subject models from three companies: claude-sonnet-4.6 (Anthropic), gpt-5.4-mini (OpenAI), gemini-3.5-flash (Google).
• Particle verb / particle placement — verbs like pick up, whose particle can precede or follow the object: pick up the book vs pick the book up. Human speakers' placement choices follow soft constraints involving givenness and length.
• Percentage point (pp) — the unit of difference between two percentages (75% vs 30% is a gap of 45 percentage points).
• Power (statistical) — an experiment's ability to detect an effect if it exists, mainly governed by sample size. Underpowered studies both miss real effects and produce unstable apparent ones; the project's standard for claim-carrying runs is roughly 100–150 test items.
• Presupposition — content a sentence takes for granted rather than asserts: John stopped smoking presupposes that John used to smoke. Projection is the survival of that presupposition when the sentence is embedded (John didn't stop smoking still implies he used to smoke); accommodation is a hearer's quiet acceptance of an unannounced presupposition.
• Prompt — the exact text sent to a model, including instructions and the question; the project treats prompt wording as part of the measuring instrument.
• Replication — re-running an experiment on fresh test items (here, usually weeks later, with the instrument frozen) to see whether the finding recurs.
• Residual — what remains of an effect after a control's contribution is subtracted; the project's measure of "beyond the shadow."
• Shadow-beater / shadow-saturated / shadow-depth — see §3: a behavior that survives its distributional control / one fully explained by it / the degree to which a phenomenon is already written into surface statistics at all.
• Temperature — a model setting controlling randomness of output; the project probes at temperature 0, the most deterministic setting, and treats each answer as a draw from a distribution even so.
• Treebank — a corpus whose sentences carry linguist-annotated grammatical structure; the Universal Dependencies (UD) treebanks used here exist for many languages under open licenses.
• Verifier — in this project, an independent session that recomputes an experiment's reported numbers from the raw outputs with separately written code before the result may be cited.
• Working surface — the project's term for an answer format that permits visible reasoning before the final answer (e.g. "think step by step, then end with FINAL: …"), as opposed to a forced single-token reply. Several capabilities are masked under the forced format and appear under a working surface.
• Word sense — one of the related meanings of a word (paper: the material vs the newspaper). Polysemy is relatedness among senses; homonymy is accidental sharing of a form by unrelated meanings (bank: riverbank vs financial institution). Lexicographers treat the boundaries as graded, not sharp.