Open Source Stats

verified statistics · every figure sourced · comparison tables, not single numbers

Reconstructed history — schema and dictionary

2,508 points across 3 metrics · CC-BY-4.0

Read this before you use the data

These are reconstructed points, not measured readings. Each one comes from a dated Wayback Machine capture of a GitHub repository page. They are sparse - captures happen when someone archived the page, not on a schedule - and nothing is interpolated. If you need a value for a date with no capture, the honest answer is that this dataset does not have one.

Do not merge these into a series of measured readings without labelling them. A backfilled point and a measured point are different evidence, even when they describe the same counter.

Files

File	Format	What it is
reconstructed-history.csv	`text/csv`	One row per point. Same fields as below, flat.
reconstructed-history.json	`application/json`	The points plus name, license, method, limits and citation — self-describing on its own.
reconstructed-history.schema.json	`application/schema+json`	JSON Schema (2020-12) for the JSON export. Validate against it rather than parsing hopefully.

Points per metric

Metric code	Points
`github_forks`	847
`github_stars`	848
`open_issues`	813

Field dictionary

Field	Type	Unit	Example	Notes
`repo`	string · owner/name	—	`anomalyco/opencode`	GitHub repository in owner/name form, using the CURRENT name. Renamed repositories still store points captured under their previous name; the name that was captured is visible in `source`.
`metric`	enum · `github_stars`, `github_forks`, `open_issues`	metric code	`github_stars`	Which counter the value belongs to. Never mix metrics in one series: they are different units.
`as_of`	string · date	UTC day	`2026-04-01`	The calendar day the capture represents. Not the reading time - the capture time is in `archived_at`.
`value`	integer	count	`96883`	The number GitHub printed on that page. Null is never stored as zero; an unreadable capture is omitted entirely.
`confidence`	enum · `exact`, `beyond_cap`	—	`exact`	`exact` means the counter was read directly. `beyond_cap` is reserved for the API path that stops at 40,000 stars and does not occur in archive captures.
`method`	enum · `wayback_repo_page`, `stargazers_binary_search`	—	`wayback_repo_page`	How the point was obtained. Archive captures need no token and have no cap; the API method is blocked on this account.
`source`	string · uri	—	`https://web.archive.org/web/20260401060630/https://github.com/openai/codex`	The exact archive capture the value came from. Any row can be re-checked by opening this URL - that is the point of publishing it.
`archived_at`	string · date-time	ISO-8601 UTC	`2026-04-01T06:06:30+00:00`	When the archive made the capture. Differs from `as_of` (the day it represents) when a capture happened late at night in another zone.
`reconstructed`	boolean	—	`true`	Always true in this file. Present so a consumer cannot mistake a reconstructed point for a measured reading when merging with other data.

This table and the JSON Schema are generated from one field definition in the build, so they cannot disagree.

Known limits

Irregular and sparse. Gaps are real and are not filled.
Values can fall: stars drift down slightly when people unstar, and fork counts can drop sharply when GitHub cleans up forks — one tracked repository lost 2,333 forks in three weeks while its stars rose. Both are published rather than filtered.
Repository names are current names; a point captured under a previous name keeps that name in source, so the lineage stays checkable.
Coverage is not uniform: repositories with a long archiving history have up to 34 captures, others far fewer.

How to cite

Open Source Stats, “Reconstructed repository history”, https://opensourcestats.com/datasets/reconstructed-history/. Reconstructed from Wayback Machine captures; CC-BY-4.0. State that the figures are reconstructed rather than measured.