Download
CC BY-SA 4.0 lets you extract, reuse and share all of this, in bulk. Nothing here requires an account. The archive is served from its public source; Syntology's layer is a stamped file with the provenance record beside it.
The archive archive 2025-07-28
The five tables this site is built from, byte-identical to what it serves (the build reads them directly):
- pwc-archive/papers-with-abstracts (576,261 papers)
- pwc-archive/links-between-paper-and-code (300,161 links)
- pwc-archive/evaluation-tables (leaderboards)
- pwc-archive/datasets (15,008)
- pwc-archive/methods (8,725 rows, of which most are spam; see what is not shown)
How the tables were reshaped is stated on the attribution page; the exact exclusion counts for this build are on what is not shown.
Syntology's layer Syntology
- graph_layer.jsonl.gz: one line per paper, holding its harvested samples (entry name, repository, path, language, verification level, contract verdict, metamorphic tier, fingerprint flag, licence, status, and, where the graph records them, the fixture's input signature, the type its run returned and when the level was recorded) and its repository links with their provenance. Each sample entry in a line is one paper–sample link, so identical code held for several papers appears once under each. Provenance record, with the licence and the status counts: graph_layer.jsonl.prov.json. 63,466 papers, 294,260 distinct samples in 361,576 paper–sample links. Three fields are left out of a sample entry, never written empty, when the file carries no value for them:
signature, present on 239,587 links and absent on 121,989, the arguments the sample's fixture called it with, each as name:type/rank/element kind;fixture_out_type, present on 267,275 links and absent on 94,301, the type name the fixture's run returned, with no value; andlevel_recorded_at, present on 252,010 links and absent on 109,566, when Syntology's graph wrote the verification level, not when the code executed, and never filled in with a build date. The signature and the returned type were recorded by two separate runs of the same fixture, and a returned type says nothing about whether the output was right. For 8,412 links the graph records a signature the file leaves out, because it is not a faithful record of the arguments the fixture passed (7,616 are the harness's own probe of a class; 796 are a printout of an object in Syntology's harness where an argument's type belongs); each such line names its reason insignature_withheld, and the provenance record counts each one underfield_exclusions. - /index/manifest.json: the paper-to-code index, keyed by arXiv id. It covers every arXiv id with a paper page here (231,180 ids in 236 files: one per arXiv month at
/index/code/<yymm>.json, andold.jsonfor ids from before 2007). Each id gives its page, its JSON twin and Syntology's counts for that paper: harvested samples, how many ran, split into runs with no instrument failure and runs where Syntology's instrument failed (those are failures of Syntology's instrument, not of the code), how many are unverified, and pointer-only samples. It also givesn_repos_linked, the distinct repositories Syntology's layer links to the paper from any source; that is not the paper page's "harvested from N repositories", which counts only the repositories its samples came from. 61,025 of the ids have a line in Syntology's layer. The other 170,155 are listed with"syntology": null: Syntology holds no run record for them, which does not mean they have no code. Entries are ordered by arXiv id and nothing is ranked. The manifest lists every file with its sha256, size and entry count, and the coverage figures the home page prints. Per-sample detail stays in each paper's twin; papers with no arXiv id, and papers in Syntology's layer with no page here, are counted in the manifest, not listed. - repo_status.json.gz: repository reachability from one probe window (2026-09-16 to 2026-09-18), 86,714 repositories, with the census's own vocabulary (ok / unanswered / gone / empty_repo / exists_but_no_tree). Provenance: the file's own
_provenanceblock (it carries its stamp inline; there is no separate sidecar). - Per-page JSON twins: append
.jsonto any paper, benchmark, task, dataset or method URL, or to a/code/<name>or/census/<name>page. The code-name list is also at /search/code.json.
Each sample's status uses the six values the pages use (ran_honours, ran_violates, ran_draft_wrong, ran_fixture, ran, unverified), written into the file by the same function that labels every page, status_of in the site's build, from the recorded level, report and contract verdict: a level below 1, or a level with no report behind it, is unverified whatever its verdict. The values are explained on about. Code bodies are not in this file; the graph service at syntology.ai serves a sample's text only when two licence records both allow it. license_inline_ok is this paper's link: its repository's licence is on Syntology's permissive list (MIT, Apache-2.0, BSD and similar) and the copy is cleared. node_license_inline_ok is the stored copy of the same code that Syntology serves, with node_upstream_license and node_repo, the repository it was first harvested from, which may be another paper's. served_inline is true only when both allow the text, and withheld_clause says in Syntology's own words which record keeps it a pointer and why (null when it is served inline). Every page on this site counts pointer only by served_inline. On 17,699 links (5,897 distinct samples, 6,853 papers) this paper's link allows the text and the stored copy does not. Syntology's layer is published under CC BY-SA 4.0, like the archive material.