Keep the boundary
of every claim visible.
This is a descriptive study of the public integer-mult-bounds repository. It covers a frozen census of 633 PRs opened October 7–12, 2026, from 77 author accounts. The baseline is ce8f38b4dcc6baf3df9150b4ba00778ed38d3d85. The conservative common cutoff for event analysis is 2026-10-12 23:33:31 UTC. Collection occurred over an interval; it is not an atomic snapshot.
Evidence layers that answer different questions
| Layer | What an edge means | What it does not establish |
|---|---|---|
| Recorded communication | Deduplicated non-author comment or formal review, sender → PR author; 142 events in the common-cutoff analysis. | Every intended addressee, coauthorship, private discussion or evaluation quality. |
| PR reference | Current-body URL or explicitly prefixed PR number. The atlas orients cited antecedent → citing PR; the statistical reference plane uses source → target. | Dependency, positive endorsement, novelty, or when the link first appeared. |
| Audited technical relation | A purposively selected, exact-excerpt-backed source statement such as combines, modifies, audits or critiques. | Independent validation of the mathematical claim or complete coverage of intellectual contributions. |
| Participation | Public account authorship, comment event authorship, or attributed participation in a curated idea record. | A human/AI identity, sole inventor, verified professional background or scientific credit. |
Numbers depend on the extraction definition
The original URL-only body graph has 438 pairs, including 11 pairs to later-opened PRs. Adding explicit PR shorthand gives 2,241 body pairs, including 78 later-opened pairs. Body plus title extraction yields 2,257 pairs; including the other collected fields such as comments and reviews yields 2,283 pairs. No extraction label turns a mention into a verified inheritance relationship. Removing later-opened targets still cannot certify historical visibility for the remaining current-text links.
The idea atlas contains 744 objects: 633 PR containers, 23 curated idea labels, 84 account nodes and four recorded comment-event nodes. The 84 accounts include participants beyond the 77 PR authors. Semantic auditing covers 29 PRs, 65 membership/role records and 36 typed relations. This is a selected sample, not an independent-idea census.
Time, censoring and adoption
The first-response curve is the Kaplan–Meier complement. Of 633 PRs, 124 have a recorded non-author response by the common cutoff; 509 are right-censored. The full-census response median is not reached. Shaded intervals show pointwise 95% author-cluster bootstrap sensitivity from 800 resamples, not causal or population certainty.
First merge and first closure are competing outcomes. PR #166 closed and later reopened, so its first closure appears in the event curve while its snapshot state remains open. Snapshot states are 131 merged, 48 closed without merge and 454 open. Cohort retention uses completed calendar days, preserves explicit active/cohort denominators, and does not imply long-term departure.
How to read the 2D and 3D atlas
2D lanes distinguish accounts, PRs, ideas and comment events in small neighborhoods. Wider maps use a fixed layout from the frozen graph. Spatial distance is not a measured intellectual distance. The optional 3D view separates opening time, family lane and object type; the same nodes and evidence are available in 2D and the relation table. Cohort playback is navigation by PR opening time, not a reconstruction of what participants could see then.
The static claim figure shows ten source-reported conditional κ records in three selected series, with a separate scale in each panel. The dashboard can show an individual series or all series on one shared axis. They do not define an assumption-matched, verified performance frontier. PR #633 reproduces #627’s reported κ while adding an audit claim. Practical speed, guarantee strength, study replay and repository adoption are kept separate.
Download and reproduce
Every chart has SVG/PDF exports and CSV companions. Dashboards can export filtered records directly. These files are the numerical and evidence companions to this website; the larger sanitized source corpus is reserved for the planned dataset release.
- Selected figure gallery with downloads
- 633-PR follow-up table · Account activity · Recorded communication events
- Graph extraction sensitivity · Audited relations · Idea memberships · Conditional claim records
- Recursive graph features · Feature hierarchy
- Published-file checksums and version manifest
A Hugging Face dataset card and reproducible packaging scaffold are being prepared. No Hugging Face release exists yet. This website makes no claim that the full dataset is ready for unrestricted redistribution under one license.
Attribution and limits
Source software retains its Apache-2.0 attribution. Discussion text belongs to its original authors; the software license is not asserted to relicense their conversations. Public source handles, URLs, revisions and excerpt provenance are retained. Private email addresses, credentials, session transcripts, local machine details and private correspondence are excluded from the website.
Deleted comments, earlier mutable-body revisions, inaccessible commits and unrecorded conversations remain missing. No personal identities or professional ties are inferred from account names. All observed associations and proposed hypotheses remain exploratory. The website uses bundled Plotly.js, no analytics, no remote fonts and no third-party data requests.