Methodology & trust

What this is, and is not

The Intelligence Papers is a map of how modern AI was actually built — the dependency network of architectures, training methods, data practices, infrastructure, alignment techniques, evaluation frameworks, and open implementations. It is deliberately not a “top papers” list, a citation leaderboard, or a collection of AI-generated summaries with no sources. Fame and operational importance are kept visibly distinct.

Selection criteria

A work is included when it introduced a mechanism, removed a concrete bottleneck, or became load-bearing infrastructure that later systems depend on. The first release seeds the Transformer lineage and its antecedents, infrastructure, alignment, reasoning, multimodality, retrieval, code, evaluation, and open-weight families — 211 records in all.

Verification tiers

Each record carries a verification state, shown on every card and record page with a non-colour marker:

• Reviewed (211 of 211). The record’s primary source is linked and its identifier (arXiv ID / DOI / canonical URL) confirmed, and its thesis and summary were written from the primary literature rather than copied from the source corpus. All of the initial corpus has now reached this state.

“Reviewed” is an editorial bar, not a claim that every field was independently re-fetched field-by-field. Summaries are concise syntheses of well-documented work; where a claim matters, follow the primary-source link on the record and check it directly.

A candid limitation. The source corpus was compiled in an environment where live citation-database APIs were network-blocked. Every citation count is therefore an order-of-magnitude estimate from training knowledge, not a live-verified integer. Counts are shown only with their source and retrieval date, flagged as estimated, and used as one weak signal — never as a ranking mechanism. We disclose this rather than paper over it.

Evidence tiers (for dependency edges)

• Direct — explicitly cited or acknowledged in the primary source.
• Strongly supported — clear architectural or methodological inheritance, even if not name-cited.
• Probable — substantial technical similarity plus circumstantial evidence.
• Speculative — plausible, not publicly verifiable.
• Unknown — insufficient evidence.

Influence — six dimensions, no single ranking

Ranking research by citations alone rewards fame over dependency. Instead every record carries six independent dimensions (0–10): academic, frontier-system, engineering, industrial, open-source, and underappreciated. The directory lets you sort by any of them.

Where a single composite is shown, its formula is public and unweighted: composite = mean(the six dimensions), to one decimal place. It is a transparent convenience for sorting, not a claim of objective universal rank. The seed values are derived from the source corpus’s importance/influence scores and refined per record during verification; they are inference, and labelled as such.

Openness vocabulary (kept strict)

Weights being downloadable does not make a model open source. We keep these distinct: fully open source (weights and training code/recipe under an OSI-style licence), open weight (weights downloadable, recipe partly or wholly undisclosed), source available, research-only, usage-restricted, API-accessible, and proprietary. Where a work carries no openness signal it is marked unknown rather than guessed.

Closed-model claims

Internals of proprietary models are never stated as fact without disclosure. Inferred methods are labelled as inference, and model families carry an explicit architecture-disclosure and training-data-disclosure level.

AI assistance

This archive’s software and its editorial synthesis were produced with AI assistance and human direction. Metadata for the verified spine is checked against primary sources; editorial claims are the author’s synthesis and are open to correction. The full dataset is downloadable from the dataset page so any claim can be independently checked.