Evidence ledger

Type: Evidence ledger
Status: Active
Canonical for: Claim-level naming evidence, evidence strength, direct URLs and explicit non-claims
Depends on: Verified primary sources and documented design-system artifacts

ID	Claim or observation	Evidence label	System/domain	Exact artifact	Direct URL	What it proves	What it does not prove
DS-01	One kind contains hierarchy, consequence and appearance values	Demonstrated	Carbon	Button usage/API	https://carbondesignsystem.com/components/button/usage/	Public semantic dimensions are mixed	That agents select incorrectly
DS-02	One appearance contains prominence, consequence and experience identity	Demonstrated	Atlassian	Button docs	https://atlassian.design/components/button/	Property labels do not guarantee one dimension	Agent effect size
DS-03	appearance mixes hierarchy and visual treatment, with hierarchy supplied in prose	Documented	Fluent 2	React Button usage	https://fluent2.microsoft.design/components/web/react/core/button/usage	Names and prose work together	Machine performance
DS-04	variant, treatment and static-color are separate	Demonstrated	Spectrum	Button API	https://opensource.adobe.com/spectrum-web-components/components/button/	Public APIs can factor dimensions	That factoring improves selection
DS-05	category, variant and states are separated; destructive prominence varies by context	Demonstrated + Documented	GitLab Pajamas	Button docs/API	https://design.gitlab.com/components/button/	Consequence and hierarchy are orthogonal	Optimal universal schema
DS-06	tone, color intensity and variant are separate	Demonstrated	Shopify Polaris	Web-component guide	https://shopify.dev/api/pos-ui-extensions/2025-10/using-polaris-components	Factored current public model	Agent accuracy
DS-07	secondary explicitly denotes a secondary color	Documented	USWDS	Button docs	https://designsystem.digital.gov/components/button/	Cross-system collision with hierarchy meaning	That the term is unusable
DS-08	Warning styling is reserved for final serious action	Documented	GOV.UK	Button docs	https://design-system.service.gov.uk/components/button/	Workflow stage affects presentation	General agent behavior
DS-09	Appearance-named Button types carry hierarchy guidance in prose	Demonstrated + Documented	Material Web	Button docs	https://material-web.dev/components/button/	Appearance names need usage rules	Relative machine performance
DS-10	variant is explicitly visual while color, contrast and loading are separate	Demonstrated	Radix Themes	Button API	https://www.radix-ui.com/themes/docs/components/button	Clean appearance model is possible	Product-intent selection quality
AI-01	Tool metadata alignment, descriptions and pretraining exposure affect selection	Empirical adjacent evidence	LLM tool calling	BiasBusters	https://www.microsoft.com/en-us/research/publication/biasbusters-uncovering-and-mitigating-tool-selection-bias-in-large-language-models/	Models use semantic metadata and learned priors	Design-system naming effect
AI-02	Overlapping tool names/descriptions reduce selection performance	Empirical adjacent evidence	LLM tool calling	ToolScope	https://arxiv.org/abs/2510.20036	Confusable schemas are harder	UI design causality
AI-03	Similar tools and relationships create selection challenges; relational experience helps	Empirical adjacent evidence	LLM tool calling	ToolExpNet	https://aclanthology.org/2025.findings-acl.811/	Alternatives/relations can aid selection	Specific component architecture
AI-04	Renaming tool and parameter identifiers while holding descriptions/examples/tasks fixed changes tool selection and parameter identification	Empirical adjacent evidence	LLM tool calling	PA-Tool	https://arxiv.org/abs/2510.07248	Tool/parameter names are causal signals in the tested settings	Enum-value naming or design-system selection effects
AI-05	Noise and opacity in tool/parameter names reduce robustness across tool-use stages	Empirical adjacent evidence	LLM tool calling	RoTBench	https://arxiv.org/abs/2401.08326	Name robustness matters in tool schemas	Isolated enum-value or design-system effects
AI-06	Intent-aligned endpoint names and name–description conflict materially change selection in a controlled technical report	Empirical adjacent evidence; non-peer-reviewed report	API method selection	Hood semantic method naming	https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6460766	Method names and conflicts can affect frontier-model selection and calibration	Generalization to enum values; peer-reviewed confirmation
AI-07	Hard decoding constraints can improve structural validity while lowering semantic/executable correctness in tested small models	Empirical adjacent evidence	Structured generation	Constraint Tax	https://arxiv.org/abs/2605.26128	Validity and correctness must be measured separately	Optimal design-system constraint architecture
AI-08	Joint optimisation of a tool name and description can strongly bias tool selection	Empirical adjacent evidence; preprint/under review	LLM tool selection	ToolTweak	https://arxiv.org/abs/2510.02554	Tool metadata is a causal attack surface when name and description are changed together	Name-only effects, enum-value effects or design-system selection effects
CV-01	Concepts are separable from preferred/alternate labels, definitions and relations	Documented	Controlled vocabularies	SKOS Reference	https://www.w3.org/TR/skos-reference/	Labels need not be canonical identity	Executable design rules
CV-02	Roles, states and properties are distinct semantic kinds with constraints	Documented	Accessibility	WAI-ARIA 1.2	https://www.w3.org/TR/wai-aria-1.2/	Typed semantics and conformance matter	Product intent semantics
NEG-01	No credible public direct design-system naming experiment was located	Not established	Cross-domain review	Search boundary	n/a	A direct evidence gap exists	That no private or future study exists
NEG-02	No verified controlled study in the reviewed bundle isolates enum-value names while holding property name, descriptions, behavior, tasks and choices fixed	Not established	Schema enum selection	Bundle audit	n/a	The closest naming question remains an experiment, not a conclusion	That enum names have no effect