Indexing covers the major ML, vision and language venues plus arXiv. What has been measured end to end is narrower: a first pass over the general-ML core, where structural extraction succeeded on 25,045 of 25,179 papers. The graph is now real and live: identity resolution, review ingestion and code-link ingestion have all run and written real edges — over a million citation edges, tens of thousands of reviews and code links. The one stage still limited to a small pilot is model-derived relation labeling (Method/Dataset/Concept). Nothing on this page should be read as claiming more than that; the methodology page marks each stage as running, piloted, or not.
This page describes what is being built toward and the constraints being held while building it. It is written to be read back later, including by people checking whether we did what we said.
The methodology page draws a distinction between having a graph and having a traceable one. The difference is a constraint that was accepted before the first edge was written: no relation enters the graph without a record of where it came from. That constraint costs throughput. It is the reason coverage is three venues rather than thirty.
These principles are being held now, while the product is small and cutting a corner would go unnoticed, because the pressure to cut them arrives later — with a funder, a deadline, or a coverage number that would look better if provenance were optional. A principle that has only ever been held when it was cheap has not been tested.
So the commitments below are written with the specific thing that would falsify each one, and with its actual state — in force now, or locked in the schema and waiting on ingestion. If one of them stops being true, it will be visible in the output before it is visible in the copy.
{{ c.body }}
Decision-making authority over the mission and the operating principles is intended to sit with independent parties: researchers, open-science advocates, and people who use the graph — not people who fund it. The point of the separation is that the party who pays for the compute should not be the party who decides what counts as a verified claim.
Running a graph of this kind requires infrastructure that a project this size does not own. Organisations who provide it are named as partners or advisors, in this category, and hold no authority over extraction, ranking, retrieval or governance. Support is acknowledged; it is not converted into influence.
Syntology is being built with the intent to become a nonprofit or public-benefit structure — infrastructure that belongs to the field rather than a company with shareholders whose interests could pull against the mission. That is an aim and a direction, not a settled legal fact. It has not been filed, granted or approved.
The purpose of the structure is narrow: shorten the distance between published research and production use, by making verified, traceable knowledge easier to find and act on. Not making claims faster to generate — there is already an abundance of that.
Every node already carries a readiness rollup — dense, sparse or unresolved — computed from its own provenance mix. The roadmap item is to use that signal to decide where extraction and human review happen next, so effort concentrates on the weakest-evidenced regions of the graph instead of being spread uniformly at random.
This is a claim about allocation, not about autonomy. It does not mean less verification, less testing, or a system that supervises itself. A graph that knows which of its own regions are thin is a graph that can spend a fixed review budget where it buys the most, and that is the whole of the claim.
Trust in a system like this is earned in one direction and lost in the other, and it is earned slowly. Syntology has to prove the extraction holds at scale before it deserves any of the language above. If it ends up serving a few hundred researchers rather than becoming public infrastructure, the provenance record, the separation of governance from funding, and the refusal to sell rank stay exactly as they are.