SYNTOLOGY The Modern Ontology for AI HomeExplorerDemoMethodologyEthosAboutDevelopers
Methodology

How a paper becomes a checkable part of the graph.

This page is the short version, written to be read by someone deciding whether to trust the output. Indexing now covers NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, ACL, EMNLP, NAACL and arXiv. The figures beside this text are from the first pass over the NeurIPS, ICML and ICLR core, which is the slice that has been measured end to end; the vision and language venues are being indexed against the same pipeline and will be reported the same way. Four of the five stages that write edges are running against the corpus today — identity resolution, review ingestion and code-link ingestion have all written real edges into the live graph. The one exception is model-derived relation labeling (Method/Dataset/Concept), which has been piloted at small scale but not run across the full corpus yet. Each stage below says which it is.

Verified, not projected
25,179
papers targeted in the first pass — NeurIPS, ICML and ICLR, 2020–2025
134
structural failures, each with a named error type on specific documents
25,045
papers parsed into the typed schema — 99.47% success, run end to end

The pipeline

five stages · running today or piloted at small scale, marked as such — the standing commitments are on the ethos page
{{ s.num }} {{ s.name }} {{ s.state }} {{ s.kind }} {{ s.caret }}
What happens

{{ s.what }}

Source
{{ s.source }}
Known limit

{{ s.limit }}

Rule one

Provenance on every edge, from the first edge written

Each relation carries where it came from — a structured API, an explicit sentence in the paper, or a model extraction with a confidence score. No edge is written without it. This is the difference between having a graph and having a traceable one.

Rule two

Reviews are recorded, never scored

One node per reviewer per paper, rating and confidence kept raw. Averaging across reviewer confidence is a known, biased statistic, so it is never done. Reviewer identity is a permanent omission — not a field to fill in later.

Rule three

Thin structure is surfaced, not hidden

Each node carries a readiness rollup — dense, sparse, or unresolved — computed from its own provenance mix. A deep trace over a sparse region says so before it runs, instead of returning a confidently wrong answer. The rollup is specified and gated into the retrieval design; it is computed once the edges it reads exist.

Coverage, stated as a bounded scope

The extraction mechanism was proven on a deliberately narrow core — NeurIPS, ICML and ICLR, 2020–2025 — before widening. Vision and language venues followed, and arXiv is indexed on a fixed cadence with its unreviewed status carried on the record rather than dropped.

{{ c.label }} {{ c.venues }}

This list is not comprehensive today and won't claim to be. Publication venues evolve; coverage will be documented as a known, bounded scope at every stage.

Improving from usage without knowing who's asking

Researchers exploring unfamiliar territory should be able to do so without the exploration itself being visible to anyone. That is a constraint on system design, not a values statement.

{{ p.title }}
{{ p.body }}
Ranking and extraction logic are designed to be auditable. Ask us anything about them.
media@syntology.ai Open the demo Ethos and intent
The Modern Ontology for AIsyntax + ontologymedia@syntology.aisyntology.ai