The AI Canon is a free, method-backed reference library for AI knowledge. It ranks texts, not people. It invites correction. It sells nothing.
Check the math, not the curator.
Curation of AI knowledge has collapsed into affiliate listicles and opinion threads while the field itself compounds. The Canon's claim is narrow and testable: knowledge curation can be made auditable, reproducible, and challengeable. Which works belong to the canon of AI is decided by published method and verifiable evidence (citations, library holdings, syllabus adoption, sustained readership), never by taste alone, and never by anything money can buy.
The list is not the product. The method is the product. The list is its first proof.
The Canon is
- A curated, multilingual library of the books, papers, reports, and standards that define the AI field
- Evidence-ranked within each domain, under a published method and weights
- Versioned: every release tagged, every change logged, every rank movement traceable
- Open to challenge from anyone, with public resolutions
- Free, permanently
The Canon is not
- A ranking of people or companies. Voices and organizations are described, never scored
- A recommendation engine, a review site, or a bookstore
- A leaderboard across domains. A standard is never ranked against a novel
- Sponsored, affiliated, or advertised. No entry can be bought or featured
- Finished. It is maintained, corrected, and re-released
The first ranking is live, and you can check the math.
Pilot release pilot-v0.1. It ranks the papers domain under three published weighting scenarios, and it survived a two-iteration adversarial review (GATE A: pass). It is deliberately narrow and honest about it: books carry no harvested metrics yet, two evidence signals are live, coverage is partial, and every gap is declared rather than zero-filled. A rank is not a verdict on worth. It is a transparent output of declared evidence, weights, and missing-data rules at this release.
| Rank | Paper (academic view) | Year | Score | Evidence |
|---|---|---|---|---|
| 001 | Deep Residual Learning for Image Recognition | 2016 | 0.3250 | citations, recency |
| 002 | Random Forests | 2001 | 0.1093 | citations, recency |
| 003 | Long Short-Term Memory | 1997 | 0.0459 | citations, recency |
Coverage, stated plainly. The corpus is strong in English. The multilingual layer is in development, and the Chinese-language spine is a known gap. We will not describe the Canon as worldwide until that gap is closed. Chinese-literate readers are invited to nominate works and contest rankings through the challenge protocol, with evidence and with credit.
Rules the ranking cannot break
The full method is published with each release and is itself versioned. In brief: scoring is deterministic, every number carries its provenance, missing evidence is recorded and penalized rather than invented, domains never cross-rank, people are context and never contestants, and a rank is an output of declared evidence and weights, not a verdict on worth. Read the full method and weighting scenarios.
Disagreement is a feature. File it.
Anyone may challenge any entry, rank, metric, category, or method rule, including ours. A challenge is contested against the cited evidence, not against opinion. Every challenge gets a public identifier and a published resolution. See the challenge protocol and log, or download the audit package and rebuild the ranking yourself.
Honest about cadence. Absolute about conduct.
Every update is logged. Every correction is traceable. Every ranking can be challenged. The library is maintained as capacity allows, without deadlines we would resent, and without commercial influence of any kind. What is promised without qualification: no advertising, no affiliate links, no sponsored placement, no paid inclusion, ever. Nothing in this library is for sale, which is precisely why it can be trusted.