open record

How arXiv papers use one another

A citation says that one paper names another. It doesn't say what the paper does with it. For every arXiv paper Pith holds, we read each sentence in the paper's LaTeX source that carries a \cite and recorded what the citing paper does with the work it cites: uses its method or its data, builds on it, compares with it, cites it as support or as background, or says it is wrong. We also linked papers whose main results are the same or conflict. Every quoted sentence is the authors' own; every label carries its probability.

citing sentences typed	77,374,763
links between arXiv papers	19,809,958
paper pairs compared by main result	470,729 stored, 66,814 shown

How each label was checked

Pith's classifier assigned every label. To measure it, other model families read samples of the same text without seeing its answer. Two such families don't always agree with each other, so agreement is counted on the items where they agree: this measures how often the classifier gives the answer two independent readers converge on. A label appears on paper pages only at the confidence shown here and only after passing its check: citation labels needed agreement of 0.85 or more on a fresh sample, and relations between papers needed a 95% lower bound of 0.85 or more.

label	shown when	agreement	95% lower bound	check it passed	checked by
Citation intent (method, data, extension, comparison, support, background)	Classifier intent probability 0.8 or more	69 of 76 (91%)	0.82	agreement of 0.85 or more on a fresh sample	GPT-5.6 Luna and Grok 4.7, blind, from the same citing sentences. 220 citing sentences drawn at random from the reviewed-corpus run (100 uniform plus rare intents); counted where the two families agree with each other.
Contests the cited work	Classifier contest probability 0.7 or more	106 of 114 (93%)	0.87	agreement of 0.85 or more on a fresh sample	GPT-5.6 Luna and Grok 4.7, blind, from the same citing sentences. 699 sentences whose intent answer put 0.2 or more on a dispute.
Papers state conflicting results	Classifier probability 0.8 or more and GPT-5.6 Luna giving the same relation	118 of 120 (98%)	0.94	95% lower bound of 0.85 or more, at least 100 items	Grok 4.7 and DeepSeek V4 Flash, blind, from both papers' title, main claim, contribution and abstract. 150 shown relations sampled across similarity bands.
Papers state the same result	Classifier probability 0.8 or more and GPT-5.6 Luna giving the same relation	218 of 239 (91%)	0.87	95% lower bound of 0.85 or more, at least 100 items	Grok 4.7 and DeepSeek V4 Flash, blind, from both papers' title, main claim, contribution and abstract. 400 shown relations in two random samples.
check_claim: papers reported as supporting or contradicting a statement	Reported by the MCP tool with the classifier's probability	110 of 112 (98%)	0.94	95% lower bound of 0.85 or more, at least 100 items	GPT-5.6 Luna and Grok 4.7, blind, from the statement and each paper's title and main claim. 120 reported pairs from 70 checked statements; planted checks: a paper's own main claim found it among the supporters 40 of 40 times, and a recorded conflicting paper appeared among the contradictions 25 of 30 times.
verify_citation: whether a cited paper's sentences support a citing sentence	Reported by the MCP tool with the classifier's probabilities	139 of 160 (87%)	0.81	agreement of 0.85 or more on fresh samples	GPT-5.6 Luna and Grok 4.7, blind, from the sentence and the cited paper's own sentences. 208 real citing sentences in two random samples; planted checks: 60 of 60 true attributions supported and 60 of 60 misattributed citations rejected.
One paper's result extends the other's (stored, not shown)	Classifier probability 0.8 or more and GPT-5.6 Luna giving the same relation	200 of 231 (87%)	0.82	not passed	Grok 4.7 and DeepSeek V4 Flash, blind, from both papers' title, main claim, contribution and abstract. 400 relations in two random samples; not shown on pages because the lower bound is below 0.85.

Candidate pairs for main-result comparison are each paper's ten nearest neighbours by title and main claim (BGE-base embeddings, cosine similarity 0.85 or more). Similarity only proposes a pair; the classifier and the confirming family decide the relation. A conflict states what each paper says; it does not say which is right.

The dataset

Release v1-20260925, built 2026-09-25. CC BY 4.0 for the labels and structure; quoted text keeps its source paper's license.

Sentence, preceding sentence and bibliography text are included only when the citing paper is licensed CC BY, CC BY-SA or CC0. For other papers the text is withheld; text_sha256 is the SHA-256 of the UTF-8 sentence as extracted, and the row id resolves on pith.science.

table	files	rows	compressed size
citations/: one row per citing sentence	30,964	77,374,763	10.7 GB
claims/: one row per compared pair	10	470,729	0.01 GB

The manifest lists every file with its row count, size and SHA-256. Download a file at /connected-record/data/<path> using the path the manifest gives. The files reproduce the live database: 19,809,958 arXiv links (match) and 470,729 compared pairs (match).

Fields in a citation row

id (citing arXiv id, version and sentence id), section, text_sha256, text and before when the license allows, refs (bibliography key, arXiv id, DOI), intent with intent_p and all intent_probabilities, load_bearing_p, contests_p, and the question bank and model that produced them.