Evidence policy
iilib separates sourced facts, graph relationships, editorial classification and comparison notes so generated research views can be traced back to structured records.
1 · Sources
Product pages, official documentation, pricing pages, provider documentation, official repositories and first-party announcements are preferred for factual records. A published factual entity must reference at least one source, and current indexing checks require primary-source coverage for the factual entities in the public graph.
2 · Entities
Tools, agents, models, APIs, companies and model families have structured records with an official URL, source references and a verified date. Generated research tables read those records directly rather than copying values into separate hand-maintained pages.
3 · Relationships
Factual relationships require evidence. Company-to-entity edges use relations such as develops, provides, operates and owns. Model availability is kept distinct from product model use, and category membership is kept distinct from factual provider relationships.
4 · Model semantics
A product-to-model uses edge is reserved for documentation that assigns a model to a current role or mode. integrates_with means the model or provider can be selected or accessed. available_via connects a model to an API. Research pages preserve these distinctions.
5 · Comparisons
Comparison pages combine entity facts and graph context with a smaller curated difference layer. iilib does not convert unlike dimensions into a single ranking score, and a comparison is published only when the pair has a plausible research reason and enough concrete differences to be useful.
6 · Categories
Category membership is editorial taxonomy, not a claim of ownership or technical integration. Category research pages expand from direct members into verified model, API and provider relationships, while featured entries remain a separate editorial layer.
7 · Freshness
A verified date records when the underlying facts were checked against their cited sources. It is not a promise that volatile facts such as pricing or product availability cannot change later. Automated operations reports track aging records so high-volatility data can be reviewed again.
8 · Indexing
A record can exist in the knowledge graph without receiving a public indexable page. Model families and categories are kept unindexed when they are too thin to justify a useful standalone page. The goal is graph completeness without manufacturing search pages.
9 · Open models
Open Models research records weights access, license class, commercial-use terms, redistribution, derivative rights, source-code availability, training-data transparency and checkpoint availability separately. Downloadable weights alone are not treated as proof of OSI-style open-source status. Unknown or conditional rights remain explicit rather than inferred.
10 · Representative coverage
iilib selects representative models that add a meaningful research dimension instead of indexing every parameter size, quantization, community fine-tune or packaging variant. Use-case labels support navigation and filtering but are editorial taxonomy, not performance rankings.
Generated research
AI Coding Landscape, LLM API Pricing, Model Context Windows, Open Models Research and Company → Product Graph are generated from the same entity and relationship records used by detail pages. Corrections should update the source, entity or relationship record; the research views then change with the graph.
Back to Research Engine →